They should be looking over the other shoulder. Because what’s coming for CRM doesn’t even have a sales team yet.

Redpoint surveyed CIOs across the Fortune 500 and mid-market on which software categories they’d be most open to replacing with AI-centric vendors. The results, ranked:

Salesforce Automation: 83%.

Not 83% who are mildly curious. Eighty-three percent who ranked CRM in their top two categories most open to AI-native replacement. Customer service management came in second at 56%. ITSM third at 55%. ERP and procurement tied at 50%.

CRM isn’t just the most vulnerable category. It’s 27 points more vulnerable than the second-place category. That’s not a gap -- it’s a canyon.

The Two Numbers That Tell the Whole Story

Two numbers. The tension between them is the entire strategic picture.

Number one: 19% of CIOs have seriously considered replacing their CRM in the last twelve months. That’s the action metric. One in five enterprise buyers has already done the work -- evaluated alternatives, run cost analyses, maybe even started a POC.

Number two: 83% are open to AI-centric CRM alternatives. That’s the intent metric. Five in six would switch if the right product showed up.

The gap between 19% and 83% is 64 percentage points. That gap is the most important number in enterprise software right now.

Think of it like housing. Nineteen percent of homeowners are actively listing their house. Eighty-three percent would sell at the right price. The market hasn’t moved yet. But here’s the twist the housing metaphor misses: in real estate, sellers wait for the right price. In software, sellers wait for the right replacement. And unlike houses, software replacements get better every quarter.

The 83% aren’t waiting for motivation. They’re waiting for a product good enough to justify the migration pain -- something that makes a rep’s morning pipeline review feel like talking to a colleague instead of clicking through 14 Salesforce tabs. That product is getting built right now.

What’s the catalyst? I think it’s already happened: it’s the public failure of Agentforce.

Agentforce and the Credibility Collapse

Salesforce bet big on Agentforce as the answer to the AI-native threat. Agents inside the CRM. Autonomous workflows. The pitch was: you don’t need to replace Salesforce, because Salesforce is becoming AI-native.

The market’s response has been more damaging than any competitor could have engineered.

That auto OEM executive again: “Agentforce has been oversold. It’s not a game changer. It’s a smart chatbot. If I just look at Agentforce as a chatbot, then there are much better chatbots out there, and the price at which Agentforce is coming in, I expect a miracle.”

This isn’t a random critic. This is a buyer who wanted it to work. Who probably championed the Salesforce relationship internally. Who went to Dreamforce and got excited about the demo. And then deployed it and found a chatbot.

The pattern across all three mega-vendors is identical. ServiceNow: “I wouldn’t even think twice about switching” if a startup offered better at a better price. Microsoft Copilot: pricing “is not going to stick” because it “literally doubles your E3.”

Every major platform tried the same play -- bolt AI onto the existing product and charge more. Every major platform is hearing the same thing back: not good enough, too expensive, feels like a tax on our existing contract.

This is the credibility collapse. The incumbents had one shot to prove that AI-native meant “we evolved” rather than “we added a feature.” They missed. And now that 83% open-to-replacement number has a tailwind of disappointment behind it.

The Salesforce Moat -- What It Actually Is

The bearish case is easy. The nuance matters more.

Salesforce’s moat was never the product. The CRM itself -- contacts, opportunities, pipeline views -- is a database with a UI. It’s been replicable for a decade. HubSpot proved that. What made Salesforce untouchable wasn’t the software. It was three things:

Data gravity. Twenty years of customer interaction data sitting in Salesforce orgs. Every call logged, every email tracked, every deal stage recorded. Migrating that data isn’t just a technical project -- it’s a political one. Sales leaders built their reporting on Salesforce. CFOs built their forecasting models on Salesforce data. Ripping it out means rebuilding institutional knowledge.

The admin ecosystem. There are over 200,000 certified Salesforce administrators. These people built careers on Salesforce expertise. They influence buying decisions. They resist platform changes because platform changes threaten their livelihoods. It’s not a conspiracy -- it’s rational self-interest at scale.

The AppExchange. Thousands of ISV integrations. CRM isn’t a standalone product -- it’s a platform wired into every other system. Replacing Salesforce means replacing or rebuilding every integration point.

That’s a real moat. It’s held for twenty years. The question is whether AI dissolves it. I think the answer is yes.

Read more

That’s not a market correction. That’s the market telling you that “works everywhere” was never a moat. It was a vulnerability with good branding.

Redpoint’s 2026 market update puts the number on what a lot of operators have been feeling for months. Horizontal SaaS didn’t just underperform. It got repriced as a category.

The last twelve months, indexed: vertical SaaS up 2-3%. Infrastructure up 2%. Horizontal SaaS down 35%. That’s not a correction. That’s a reclassification. The market looked at three categories of software and decided one of them has a fundamentally different risk profile than the other two.

The question isn’t whether the market is right. It’s why.

The Three Layers, Three Fates

Redpoint breaks the software universe into three layers and explains why AI hit each one differently. This is the slide I keep sending to founders who ask me “is my category safe?”

Vertical SaaS (+2-3%). Owns the irreplaceable moat: data plus compliance. When you’re the system of record for a dental practice or a property management company, AI is a feature you add, not a threat you face. The switching cost is existential, not cosmetic. Nobody is ripping out their EHR because a startup has a better chatbot. They might add the chatbot on top, but the vertical platform stays.

Infrastructure (+2%). AI is a tailwind, not a displacement risk. More AI means more compute, more data pipelines, more observability. Every agent someone deploys needs to run somewhere, log somewhere, fail somewhere. Infrastructure companies sell picks and shovels to both sides of the AI war. Their multiples compressed a bit, but their thesis got stronger.

Horizontal SaaS (-35%). And here’s the line that hit me: “Accidentally optimized for replaceability.”

That phrase is doing more work than any three-word description has a right to. Let me unpack it.

Horizontal SaaS was designed -- intentionally, deliberately, as a strategy -- to serve every industry equally. That was the pitch to investors. “We don’t need to build vertical-specific features. Our platform is flexible enough to work everywhere.” And it was true. Horizontal tools do work everywhere. The problem is that “works everywhere” and “integrated deeply with none” are the same statement.

I’ll be honest -- I’ve pitched “works everywhere” as a strength in past roles. I’ve sat in the room and said “we don’t need vertical-specific features, our platform is flexible enough.” It sounded like a moat. It was a vulnerability with good branding.

When your product is a coordination layer -- moving information between people, triggering workflows, displaying dashboards -- you’re doing exactly what AI does natively. You’re a middleman between humans and data. And AI just cut out the middleman.

The coordination problem that horizontal SaaS was built to solve? AI solves it natively. Not by being a better tool for coordination, but by eliminating the need for coordination in the first place. When an agent can pull data from six systems, synthesize it, and take action, the dashboard that displayed data from those six systems becomes a loading screen nobody needs.

Two Playbooks for Eating the Incumbents

So the horizontal giants are vulnerable. Who’s coming for them? Redpoint maps two distinct startup playbooks, and the distinction matters enormously for which incumbents should be scared and which have time.

Playbook 1: Augment the Enterprise. Sell alongside the incumbent. Own the AI layer on top. Don’t replace Salesforce Service Cloud -- sit on top of it. Sierra and Decagon are doing this in customer service. Moveworks is doing it on ServiceNow. Legora on LexisNexis. Lovable on Figma.

This playbook is smart because it doesn’t require the customer to rip and replace. It’s additive. The incumbent barely notices until the AI layer is handling 60% of the workload and the customer starts asking why they’re paying full price for a platform that’s mostly a data store now.

Think of it like ivy on a building. Looks decorative at first. Even charming. The building owner likes it -- adds character, doesn’t cost anything, and guests compliment it. Then one year you realize the ivy is load-bearing and the building is just a frame. That’s the moment the incumbent realizes the “AI integration partner” owns 60% of the customer’s workflow and they’re just the database underneath.

Playbook 2: Attack the SMB/Mid-Market. Rebuild from scratch for the segment the incumbent under-serves. Attio vs. Salesforce. Serval vs. ServiceNow. Linear vs. Atlassian. Rillet and Everest vs. NetSuite.

This one is the classic disruption pattern -- go downmarket where the incumbent’s cost structure can’t follow. But AI accelerates it because the rebuild is 10x faster. What used to take a startup three years and $20M in funding to build (a credible CRM alternative) now takes 12 months and $3M. The attack surface expanded because the cost of building enterprise-grade software collapsed.

Both playbooks are dangerous, but in different timeframes. Playbook 1 is a slow squeeze -- the incumbent keeps the account but loses the value and eventually the pricing power. Playbook 2 is a fast grab -- the startup takes the customers the incumbent never prioritized and works upmarket.

If I’m a horizontal SaaS CEO, I’m worried about Playbook 1 in my enterprise accounts and Playbook 2 in my mid-market. Which means I’m worried about everything.

Read more

Redpoint’s 2026 deck says every software company is standing at that same fork right now. I agree. I’ve spent the last year inside a 20-year-old healthcare software company with thousands of customers trying to force exactly this kind of reinvention. And I can tell you the hardest part isn’t the strategy. It’s not even the technology. It’s the moment someone in the room says “but our margins are fine” and everyone exhales, because that sentence is the most comfortable way to avoid the uncomfortable conversation.

“Have a re-founding moment” is investor language. Operators need a playbook.

The Incumbents Are Telling On Themselves

Before we get to the how, let’s sit with the why. Redpoint surveyed executives at Fortune 500 companies and the quotes are devastating -- not because they’re surprising, but because they’re specific.

An auto OEM exec on Salesforce’s Agentforce: “I think Agentforce has been oversold. It’s not a game changer. It’s a smart chatbot. If I just look at Agentforce as a chatbot, then there are much better chatbots out there, and the price at which Agentforce is coming in, I expect a miracle and that’s not happening.”

A Fortune 500 exec on ServiceNow: “If there’s a startup that came along and said, ‘We can do this better at a better price,’ certainly they’ll beat the price of ServiceNow because ServiceNow is always going to be the highest. I wouldn’t even think twice about switching.”

Another Fortune 500 exec on Microsoft Copilot: “Microsoft recognizes that the pricing that they released for Copilot is not going to stick because it literally doubles your E3.”

These aren’t analysts speculating. These are buyers -- the people writing checks -- saying out loud that the emperor’s AI clothes don’t fit. When your biggest customers describe your flagship AI product as “a smart chatbot” and your pricing as “expecting a miracle,” you don’t have a product problem. You have a credibility problem.

And credibility, unlike software, can’t be patched in a sprint.

The Rosetta Stone

The most useful slide in Redpoint’s entire 60-page deck is page 43. It’s a simple two-column comparison: Traditional SaaS vs. AI-Native. I’ve stared at it for weeks now, and every time I look, I see a different failure mode in my own organization.

Let me walk through what this actually means when you’re sitting in the chair.

Executive Team. Traditional: “been there done that.” AI-Native: “First-principles thinkers. No playbook exists. Speed matters.” I’ve been in rooms where the most senior person’s primary qualification is having scaled a SaaS company in 2014. That experience is now a liability dressed up as a resume line. The playbook they ran -- PLG motion, land-and-expand, seat-based pricing -- was built for a world where software was deterministic and customers needed training. That world is gone.

Product Development. Traditional: customer-led (listen, build to spec). AI-Native: possibility-led (understand models, build ahead). This one cuts deep. I spent years building the muscle of “talk to customers, build what they ask for.” It’s good practice. It’s also a guaranteed way to build yesterday’s product. When your customer asks for a better dashboard, they’re not asking for the agent that could eliminate the need for the dashboard entirely. Customer-led development in an AI world means your roadmap is always one paradigm behind.

Engineering. Traditional: deterministic. AI-Native: probabilistic. This isn’t just a technical distinction -- it’s a hiring filter, a QA philosophy, and a product liability question all wrapped into one word. I watched a team spend three months trying to make an LLM-powered feature pass the same regression test suite they use for their REST APIs. The tests kept “failing” because the outputs varied. The outputs were supposed to vary. That’s the whole point.

Sales. Traditional: packaged product sold as-is. AI-Native: Forward-Deployed Engineer model. You know what FDE means in practice? It means your first 50 customers each get a slightly different product. It means your sales team needs to understand what the technology can do, not just what the current SKU does. I watched a sales rep demo our traditional product for 45 minutes and then spend 5 minutes on the AI features because -- his words -- “I don’t want to promise something that might work differently next week.” He’s not wrong. But that instinct will kill you.

Pricing. Traditional: seat-based, predictable ARR. AI-Native: consumption or outcome-based, experimental. This is the one that makes CFOs break out in hives. Seat-based pricing is beautiful. It’s predictable. It makes forecasting easy. It also has nothing to do with value delivery in an AI world. When one agent can do the work of ten seats, you’re either charging for outcomes or you’re watching your TAM collapse. We tried floating outcome-based pricing internally. Finance asked how to model it. I said, “You can’t, not yet.” That meeting ended early.

The Re-Founding Moment

Redpoint frames this as an architectural shift comparable to the cloud transition. They point to historical parallels: Microsoft and Adobe embraced cloud and outperformed. IBM and Oracle resisted and lost share.

That framing is correct and completely unhelpful.

Here’s why: when Microsoft went to cloud, Satya Nadella had a $400 billion balance sheet and could afford to cannibalize Office for a decade. When Adobe moved to subscriptions, they had a near-monopoly in creative tools and could weather the transition dip. The lesson isn’t “be brave like Satya.” The lesson is “you probably don’t have Satya’s margin of error, so you need a different plan.”

Read more

Let me walk through what the deck actually shows, and why it matters more than the panic.

The Infrastructure Is Real This Time

The first thing Redpoint establishes is that this isn’t the dotcom bubble. AI capex as a percentage of GDP is now a distinct investment wave, comparable to railroads, telephony, and the internet buildout (Slide 2). Hyperscaler CapEx is going from $309B in 2024 to $779B by 2027, a 36% CAGR (Slide 3). That’s real capital, not speculative fiber.

The part that kills the bubble narrative: OpenAI and Anthropic are each generating $20B+ in ARR. Over 90% of buildout capacity is pre-committed. Near-zero data center vacancy rates. 1B+ monthly active users four years post-ChatGPT (Slide 4). As Logan Bartlett, who co-wrote the deck, put it on LinkedIn: “The infrastructure isn’t running ahead of demand. Demand is pulling the infrastructure forward.”

This is the opposite of 2000. In 2000, you had Cisco trading at $0.37 EPS while fiber sat at 3% utilization. Today, Nvidia is at $4.06 EPS and capacity is oversubscribed. That distinction matters because it tells you the correction in software stocks isn’t about the AI thesis being wrong. It’s about the AI thesis being right, and software being on the wrong side of it.

The Three-Layer Divergence

Software is down 20% YTD, the worst-performing sector in the S&P 500 (Slide 12). Public SaaS median NTM multiple is at 4.1x, the lowest since Redpoint started tracking in 2007 (Slide 13). But the selloff didn’t hit evenly. And the unevenness is where the real signal lives.

Redpoint breaks it into three layers (Slide 16-17):

Vertical SaaS: +3% LTM. These companies own the irreplaceable moat: proprietary data plus compliance infrastructure. AI can’t displace Veeva in life sciences or Toast in restaurants without re-acquiring years of industry-specific data, regulatory relationships, and transaction embedding. For vertical SaaS, AI is a feature, not a threat.

Infrastructure: +2% LTM. More AI deployment means more compute, more data, more observability. Snowflake, Datadog, MongoDB, Cloudflare all get AI tailwinds. The multiple compressed, but the thesis didn’t.

Horizontal SaaS: -35% LTM. This is the kill zone. These companies, as Redpoint puts it, “accidentally optimized for replaceability.” They built to serve every industry equally, which meant integrating deeply with none. Their core value prop was answering “who is doing what and when,” which is a coordination problem. AI solves coordination natively.

That -35% isn’t panic. It’s the market correctly pricing the fact that horizontal SaaS moats were always an illusion. The switching cost was familiarity, not dependency. And familiarity is worthless when a new tool is 10x better on day one.

The Agent Maturity Curve Changes the Math

Here’s where the Redpoint deck gets genuinely important. Slide 7 maps the agent maturity curve:

• Copilots (where we’ve been): TAM is ~$0.5T, the existing US software spend

• Task Agents (where we are now): TAM grows to ~$1.2T, adding services automation

• Workflow Agents (next 18-24 months): TAM hits ~$2.8T, adding operational payroll

• Autonomous (the endgame): TAM reaches $6.1T+, unlocking knowledge worker payroll

We’re at Task Agents. Minutes of independent runtime. Human oversight still required. But even at this stage, the TAM has already doubled beyond software. The companies winning here aren’t selling seats. They’re selling outcomes against labor budgets.

This is why 58% of CIOs say AI features are the #1 driver of software spend increases (Slide 19), and simultaneously, 54% are pursuing vendor consolidation and 45% of AI budgets are replacing existing software (Slide 20). Only 3% of CIOs expect AI to lead to more vendors. The market is converging, not expanding.

Read those numbers together. CIOs are spending more on software, but only when that software has AI capabilities baked in. And they want fewer vendors, not more. The pricing model is shifting underneath everything: 46% of CIOs expect usage or outcome-based pricing to become more common, while 35% expect flat subscriptions to decline. The per-seat model that built the SaaS industry is unwinding in real time.

AI spending is going up. Software vendor count is going down. Pricing is shifting from seats to outcomes. If you’re a horizontal SaaS company selling per-seat licenses for coordination software, those three facts are a death sentence unless you pivot fast.

The Newspaper Analogy (Why Quarterly Beats Don’t Matter)

Redpoint draws a chilling parallel on Slide 22: US newspaper stock returns started collapsing around 2002, but consensus forward earnings held up until 2007 before falling off a cliff. Stock prices led earnings by roughly five years.

Apply that to SaaS. Companies are still posting decent quarters. Revenue is still growing. But 85-95% of SaaS enterprise value is terminal value, not explicit forecast period (Slide 23). A strong quarter moves less than 5% of the value. You can’t disprove a 10-year existential question with a 90-day result.

The market sees this. In three months, the implied long-term growth rate for a representative SaaS company collapsed from 4.7% to 1.1% (Slide 24). The near-term estimates didn’t change. The terminal value did. Public investors are telling you they believe SaaS growth asymptotes to near-zero over the long term.

Even if a company beat estimates, the stock can drop the next day. Now they know why.

Where the Vulnerability Is Highest

The CIO survey data on Slide 30 tells you exactly where to worry. When asked which categories they’re most open to replacing with AI-centric vendors:

• Salesforce Automation: 83% (this should terrify Salesforce)

• Customer Service: 56%

• ITSM: 55%

• ERP: 50%

And the incumbents are, in Redpoint’s assessment, “missing the boat” (Slide 42). One auto OEM executive on Agentforce: “It’s been oversold. It’s not a game changer. It’s a smart chatbot.” A Fortune 500 executive on Microsoft Copilot: the pricing “literally doubles your E3” and they’ve decided against enterprise-wide rollout. A global industrial executive on ServiceNow: “If there’s a startup that came along and said, ‘We can do this better at a better price,’ I wouldn’t even think twice about switching.”

The AI-native companies are operating at a fundamentally different efficiency level. Cursor generates $6.1M in ARR per employee. Lovable: $3.4M. Compare that to the incumbents they’re disrupting: Salesforce sits at $0.54M. Atlassian: $0.46M (Slide 55). Cursor is producing 11x more revenue per human than Salesforce. That’s not a marginal improvement. That’s a structural cost advantage that compounds every quarter. It means AI-native companies can undercut incumbents on price, outspend them on R&D as a percentage of revenue, and still generate better margins.

As one commentator on the deck noted: the incumbents can’t decouple revenue from headcount. They sell seats. They staff with humans. Their entire economic model assumes that revenue scales linearly with people. AI-native companies broke that assumption on day one.

Three Questions to Ask Monday Morning

1. Are you vertical or horizontal? If you can’t name the specific regulatory framework, industry data set, or transaction type that makes your product irreplaceable in your market, you’re horizontal. Start the re-founding conversation (Redpoint’s term) now. Not next quarter.

2. Can your revenue decouple from headcount? The AI-powered P&L model Redpoint outlines (Revenue +25-35%, headcount costs -15-20% across every function) only works if you can deliver more value with fewer people. If your revenue model is fundamentally tied to the number of humans on your team or your customer’s team, you’re on the wrong side of the curve.

3. What happens at Workflow Agents? We’re at Task Agents today. The jump to Workflow Agents (hours of independent runtime, multi-step business processes across systems) is coming in the next 18-24 months. That jump takes the addressable market from $1.2T to $2.8T. But it also means AI goes from executing discrete tasks to orchestrating entire workflows. Every product that exists primarily as workflow orchestration software becomes a feature of the agent, not a platform the agent runs on.

The Window

Slide 67 is the one I keep coming back to. Across three platform shifts (internet, cloud, mobile), the durable winners were founded in years 4-5 after the shift began. ChatGPT launched in November 2022. We’re in year 4. 2026-2027 is the window.

That’s not a VC pitch. That’s a pattern backed by three decades of data. The companies that will own the next era of business software are being built right now. Some of them are incumbents who execute a genuine architectural reset. Most of them are AI-native startups running at $3-6M ARR per employee while the incumbents are stuck at $500K.

The prescription starts with one honest question: when AI agents can orchestrate entire workflows, what part of your product is still irreplaceable?

If you can’t answer that clearly, the market already has.

The Stories

Anthropic Accidentally Leaked Its Most Dangerous Model

The noise: breathless coverage of “Mythos,” the secret model tier above Opus that Anthropic says is too dangerous to release. Cybersecurity exploitation capabilities. Pentagon contracts. $380B valuation. The mystery box is doing its job.

The signal: A safety-focused company exposed 3,000 internal files through a CMS misconfiguration. Forget the model for a second. The company whose entire brand is “we’re the responsible ones” just demonstrated that operational security is harder than alignment research. The model itself might be genuinely dangerous. The leak proving Anthropic can’t secure its own CMS? That’s a different kind of dangerous. Consumer subs doubling and an $18B revenue target tell you the business is working. The question is whether “too dangerous to release” becomes a moat or a liability.

OpenAI Killed Sora. Nobody Should Be Surprised.

The noise: AI’s first major product death. The end of an era. Video generation dreams dashed.

The signal: $15M per day in inference costs against $2.1M in total lifetime revenue. That’s not a product death. That’s a math problem that finally got solved. Sora was always a demo dressed up as a product. The real story: OpenAI at $850B valuation and $20B+ ARR can afford to kill things that don’t work. That’s a sign of discipline, not failure. Video generation stays in ChatGPT for subscribers, which is exactly where it belongs: a feature, not a platform.

The White House Wants to Preempt Every State AI Law

The noise: the National AI Policy Framework is either “pro-innovation” or “regulatory capture,” depending on which side of the aisle you read.

The signal: this framework has already failed legislatively twice. It’s sector-specific, creates no new regulatory body, and preempts state laws. Translation: the federal government wants to prevent a patchwork of 50 state regulations without actually regulating anything itself. Industry loves it because the alternative is California writing the rules for everyone. Whether you think that’s good or bad depends on whether you trust sectors to regulate themselves. (History says no.)

Claude Code’s Auto Mode Changes the Developer UX Game

The noise: incremental product update. Classifier-based permissions. Less clicking.

The signal: the approval loop was the single biggest friction point in AI-assisted coding. Every time a developer had to click “yes, proceed” for an obvious action, the tool lost momentum. Auto mode is Anthropic betting that the classifier can distinguish between “rename this variable” and “delete the production database” well enough to remove the human from routine decisions. If it works, this is the kind of UX shift that changes daily adoption numbers, not quarterly strategy decks.

OpenAI Bought Six Companies in 2026. That’s Not a Model Company.

The noise: OpenAI acquired Astral (uv, Ruff) and Promptfoo. Smart picks. Developer tooling. Evaluation infrastructure.

The signal: six acquisitions in a single year. Add the $50B Amazon partnership for stateful agent runtime on Bedrock. This is a platform company building distribution, not a research lab building models. Sam Altman is running the Microsoft playbook: own the developer toolchain, own the cloud runtime, own the ecosystem. The model becomes the kernel nobody thinks about. Whether that’s good for the industry depends on how tightly they lock the stack.

MCP Hit 97 Million Monthly SDK Downloads

The noise: protocol adoption numbers go up. Good for Anthropic.

The signal: 4,750% growth in 16 months. That’s not adoption. That’s infrastructure. When a protocol layer hits this kind of trajectory, it stops being a feature and starts being a standard. MCP is becoming the USB-C of AI agent connectivity. The companies not building MCP integrations today will be retrofitting them in six months.

Atlassian Cut 1,600 People and Blamed AI

The noise: another round of tech layoffs dressed up as strategic transformation.

The signal: “AI pivot” is becoming the “blockchain strategy” of 2026. When a company lays off 10% of its workforce and leads the press release with AI, the community calls it correctly: cost-cutting with better PR. The real AI pivots are happening at companies that are hiring into new roles, not cutting existing ones. Watch what companies build, not what they say during layoff announcements.

Google Went From 5.4% to 18.2% Chatbot Market Share in a Year

The noise: Gemini is catching up. Google is back in the AI race.

The signal: the fastest growth trajectory in the market, driven by the most aggressive pricing anyone has seen. Gemini 3 at $2/$12 per million tokens is a price that forces everyone else to respond. Google is doing what Google always does: subsidize adoption with infrastructure advantages nobody else can match. DeepMind’s research pipeline feeding directly into products that run on Google’s own chips and cloud. The 114 model price changes across the industry in March alone tell you this pricing war is unsustainable for everyone except the company that owns the data centers.

Meta Shipped Open-Weight MoE Multimodal Models

The noise: Llama 4 Scout and Maverick are here. Open source wins again.

The signal: Meta keeps commoditizing the model layer, and every release makes the gap between open-weight and proprietary models smaller. Mixture-of-experts architecture in open weights means smaller companies can run competitive models on reasonable hardware. LlamaCon on April 29 will likely accelerate this. The strategic play is clear: if models are commoditized, the value moves to applications, data, and distribution. Meta has all three.

Robotics Got $1.2 Billion in March Alone

The noise: LeCun’s AMI Labs raised $1.03B. Physical AI is the next frontier.

The signal: JEPA architecture (Joint Embedding Predictive Architecture) is LeCun’s bet that the path to general intelligence runs through world models, not language models. A billion dollars says he’s not alone in that belief. The robotics funding surge is the market saying that software-only AI is approaching diminishing returns for certain problem classes. Moving atoms, not just bits, is where the next decade of value creation lives. And unlike software agents, physical robots need real-world training data that can’t be synthesized from the internet. That data moat is why the money is moving now, before the window closes.

Perplexity Launched a “Personal Computer.” Yes, Hardware.

The noise: a search company is selling an always-on Mac mini AI agent. Quirky pivot.

The signal: Perplexity is testing a thesis that matters. The browser is a bad form factor for AI agents that need persistent context, local file access, and always-on availability. An always-on device that runs your AI agent locally solves the context window problem by never closing the session. It’s early, probably too early. But the insight underneath is correct: cloud-based chat interfaces are a transitional form factor. The companies thinking about what comes after the chat window will define the next generation of AI UX.

OpenAI Foundation Pledged $1B for 2026

The noise: OpenAI is serious about safety. Zaremba is leading. Big number. Good optics.

The signal: context matters. This is the same organization that made six acquisitions, partnered with Amazon for $50B, and killed Sora when the economics didn’t work, all in the same quarter. A billion dollars in safety funding against that backdrop is either genuine commitment or the cost of maintaining a narrative while you scale as fast as possible. The proof will be in what the foundation actually funds, who has independence to publish findings that embarrass the parent company, and whether any of it slows down the commercial roadmap. I’ll believe it when I see a safety finding delay a product launch.

The Pattern

Two forces defined this week. The first: the split between “too dangerous” and “too expensive.” Anthropic won’t ship Mythos because of safety concerns. OpenAI killed Sora because the economics didn’t work. These are fundamentally different reasons for products not reaching users, and the industry needs to stop conflating them. Safety decisions and business decisions require different frameworks, different oversight, and different public accountability.

The second: the platform war is over before most people realized it started. OpenAI’s six acquisitions, Amazon’s $50B partnership, MCP’s infrastructure trajectory, Google’s pricing aggression. These aren’t companies competing on model quality anymore. They’re competing on ecosystem lock-in, developer toolchains, and distribution. The model layer is being commoditized from above (proprietary platforms) and below (open-weight releases from Meta). The winners will be decided by who owns the integration points, not who has the best benchmark score.

If you’re building on AI right now, the question isn’t which model to use. It’s which platform you’re willing to be locked into for the next five years.

“7.5 million AI agents. 75,000 humans.”

That’s 100 agents per person. At NVIDIA. Not in a pitch deck. Not in a research paper. At one of the most operationally sophisticated companies on the planet.

Analysts heard a headline. VCs heard a thesis. I heard someone finally put a number on the thing I’ve been arguing from the field for the last year: the AI-native org doesn’t add agents to the existing structure. It replaces the structure entirely.

I run a 103-agent system. Not as a proof of concept. As my actual operating architecture for GTM work — content, research, competitive intelligence, sales enablement, revenue operations, the whole stack. And what Huang described isn’t science fiction from where I’m sitting. It’s Tuesday.

The ratio itself isn’t the insight. It’s what the ratio demands of the humans who remain.

The End of “Managing” as a Job Description

The typical VP of Sales manages 8-12 people. Maybe 20 if the span is wide. The job is 1:1s, coaching, pipeline reviews, conflict resolution, career development. It’s relationship-intensive, time-intensive, and — let’s be honest — it doesn’t scale.

Now imagine that same VP orchestrating 100 agents. The skillset inverts completely.

You don’t coach agents. You configure them. You don’t do 1:1s. You review output quality and adjust routing logic. You don’t resolve conflicts between personalities. You resolve conflicts between competing optimization functions.

This isn’t an incremental change. It’s a different job. And most leaders I talk to haven’t internalized that yet.

I was on a call two weeks ago with a CRO at a $400M SaaS company. Smart guy. Aggressive growth targets. He told me he was “exploring AI agents for the team.” When I asked what that meant, he described giving his reps access to a chatbot that could draft emails.

That’s not a 100:1 org. That’s a 0.1:1 org with better autocomplete.

Which Human Roles Survive (And Which Don’t)

These are the human roles that get more valuable at 100:1, not less:

Architects — Someone has to design the system. Which agents exist, how they connect, what data flows between them, where human judgment gates sit. This is the new org design. When I built my routing table — literally mapping task patterns to specific agents across 9 departments — that was architecture work. No agent did that for me.

Governors — At 100:1, one misconfigured agent creates cascading failures. Someone has to set guardrails, monitor for drift, audit outputs, and decide when the system’s confidence threshold requires human override. I run drift detection on my agent specs. If the source material changes but the agent hasn’t been updated, I get flagged. That governance layer is a human job.

Relationship holders — Agents don’t build trust. They don’t read the room in a board meeting. They don’t know that your champion at the prospect company just went through a divorce and needs you to be a human being before you’re a seller. The judgment calls — when to override the system, when to ignore the data, when to just listen — those stay human.

Taste-makers — Agents can generate a hundred variants of a LinkedIn post. Someone has to know which one sounds like you. The editorial judgment, the brand instinct, the “that’s not quite right” feeling — that’s irreplaceable. I review every piece of content my writing agent produces. The agent gets me 80% there. The last 20% is taste.

What gets absorbed into agent workflows? Research. First-draft creation. Data analysis. Competitive monitoring. Lead scoring. Email sequencing. Meeting prep. Follow-up cadences. Report generation. Most of what a typical BDR, SDR, or marketing coordinator does today.

That’s not a prediction. That’s what’s already happening in my system.

What Managing Agents Actually Requires

Let me be specific about what my day looks like orchestrating 103 agents, because the “100:1 ratio” sounds clean and the reality is messier.

Routing is the new management. Every task that enters my system hits a routing layer. Is this a Tier 1 obvious match — a writing task that goes straight to the content writer? Or is it Tier 3 — a novel cross-domain problem that needs multiple specialists in sequence? Getting the routing right is 60% of the job. Bad routing wastes more time than bad agents.

Memory is the new moat. My agents share a knowledge graph. When the competitive intelligence agent discovers something, the content writer can reference it. When the product marketing agent updates positioning, the sales enablement agent adjusts its playbooks. This compound learning is the actual competitive advantage — not any individual agent’s capabilities.

Context windows are your bottleneck. Every agent has a finite amount of information it can hold at once. Managing what goes into that window — and what gets folded, summarized, or archived — is a genuine operational discipline. I built a context management system with token budget tiers. It sounds like infrastructure. It’s actually the difference between agents that hallucinate and agents that produce.

Observation beats instruction. The best agents in my system aren’t the ones with the most detailed prompts. They’re the ones where I’ve built feedback loops — tracking what works, what doesn’t, and adjusting the system. I have a proactive monitor that checks 9 conditions across the system: inbox age, content staleness, queue status, agent freshness, orphan notes. The system tells me where to look. I decide what to do about it.

The Token Salary Signal

One more thing Huang said that deserves separate attention: NVIDIA plans to offer engineers $100K-$150K in AI compute tokens on top of their cash compensation.

Read that again. They’re not offering a perk. They’re offering capacity.

An engineer with $150K in compute tokens can spin up more agents, run more experiments, train more models than an engineer without them. The gap between the two isn’t about money. It’s about leverage. One engineer ships the output of ten. The other ships the output of one.

This is where the 100:1 ratio meets compensation design. If your agents are your workforce, and tokens are what powers those agents, then token access IS workforce capacity. The companies that figure this out first will attract and retain the engineers who can actually operate at 100:1 scale.

What To Do This Week

If you’re leading a GTM team and the 100:1 number made you feel something — excitement, anxiety, skepticism — channel it into one action:

Audit your current ratio.

Not “how many AI tools does my team use.” That’s the wrong question. The right question is: how many autonomous workflows run without a human in the loop?

Count them. Be honest. If the answer is zero, you’re not behind on AI adoption. You’re behind on architecture. The tools are available. The system design is what’s missing.

Huang didn’t announce new technology at GTC. He announced a new operating model. The technology has been here. The question is whether your org structure has caught up.

The leaders who build their 100:1 architecture in 2026 will own their categories by 2028. The ones waiting for it to “mature” will be buying the playbook from the first group.

Start with one agent. Then five. Then the ratio will tell you where to go next.

Here’s what happened, and what it actually means.

Jack Dorsey Cut 10,000 Jobs — and Didn’t Blame the Economy

The noise: Block is shrinking. Another tech layoff story. File next to the others.

The signal: Dorsey said something no executive in any prior technology wave has said in public: these cuts are not driven by financial difficulty. They’re driven by the growing capability of AI tools.

Read that again. He didn’t say “we’re rightsizing for efficiency.” He didn’t say “market conditions.” He said: AI got better, so we need fewer people. That’s a fundamentally different statement than anything we heard during the cloud era or the mobile era.

Oracle said the same thing, different math: 20,000 to 30,000 cuts to free up $8 to $10 billion for AI infrastructure. Not because they’re struggling. Because they’re repositioning.

Both companies are restructuring in anticipation of AI capability — not because they’ve already proved the results. That’s a strategic bet of the highest order. And whether you agree with the human cost or not, as a GTM leader you need to understand what’s being signaled: the largest organizations in tech are now treating headcount and AI infrastructure as substitutes on the balance sheet.

The question for every revenue team this week isn’t whether AI replaces your reps. It’s whether your competitors are already modeling that math.

Google Built a Design Tool That Made Figma’s Stock Drop 8.8% in a Day

The noise: Another Google Labs experiment. Probably won’t ship. Figma is fine.

The signal: Google Stitch isn’t a Figma clone. It’s a fundamentally different workflow. You describe what you want in natural language. You get a high-fidelity, interactive prototype with voice control. The thing that stopped me was DESIGN.md — a structured file of agent-friendly design rules that other AI systems can read and follow automatically.

That’s not a design tool. That’s design infrastructure. The prototype isn’t the end product — it’s the prompt for the next agent in the chain.

Figma fell 8.8% because investors understand what the market sometimes doesn’t: when the interface layer becomes conversational, the category leaders who built moats around GUI complexity are exposed. Figma’s moat was the interface. Google just made the interface optional.

I’ve been saying for two years that the companies most at risk from AI aren’t the ones in obviously automatable categories — they’re the ones whose entire competitive advantage is UX sophistication in a world where UX becomes a text input.

Microsoft, Anthropic, and the Infrastructure Layer

The noise: More AI partnerships. More enterprise licensing tiers. More announcements with no immediate impact.

The signal: Microsoft launched Copilot Cowork alongside a new E7 licensing tier, with Claude underneath. At the same time, Anthropic shipped Claude Cowork and Dispatch — persistent agent threads that work across Claude Desktop, iOS, and Android. Start a task on your laptop. Check in from your phone. Claude is still working. Recurring scheduling, on-demand triggers, Claude as a PowerPoint add-in.

This is the moment AI stops being a feature inside your software and becomes the connective tissue between your work. The infrastructure framing is intentional. You don’t think about your email client as a productivity tool — it’s just how work happens. That’s where Microsoft and Anthropic are pointing.

Anthropic also opened Enterprise tier — including Claude Code and Cowork — to self-serve purchase. No sales call required. $100M committed to a partner ecosystem. They’re moving from enterprise sales motion to infrastructure adoption motion. That’s a meaningful signal about where they think growth comes from next.

If you’re running a GTM team and you haven’t put a Claude subscription in your budget this quarter, you’re already a cycle behind.

Claude Code Gets Voice and a Million-Token Context Window

The noise: Developer tool update. Irrelevant unless you write code.

The signal: Push-to-talk via spacebar. One million token context window for Max, Team, and Enterprise subscribers. 64,000 default output tokens.

I keep watching non-technical GTM leaders walk past Claude Code like it doesn’t apply to them. It applies to you. With a 1M token context window, you can feed Claude your entire CRM export, your full competitive intelligence file, your last 12 months of sales calls, and your pricing model — all at once — and ask it to build you a territory plan. That’s not coding. That’s thinking at scale.

(Thursday I’m publishing a paid guide: how to use Claude Code in 20 minutes if you’ve never written a line of code in your life. This is the one I wish existed when I started.)

2.5 Million People Are Trying to Quit OpenAI

The noise: Internet drama. Brand controversy. Will blow over.

The signal: The #QuitGPT movement hit 2.5 million pledges after OpenAI’s Pentagon AI deployment deal. Reddit and X filled with migration guides to Claude and Grok. This matters less as a politics story and more as a market structure story.

OpenAI built its consumer lead on being the default. Defaults are sticky until they’re not. The fact that 2.5 million people went looking for migration guides — and found them — tells you the switching cost is lower than OpenAI’s market share would imply. That’s a vulnerability. When defaults break, they break fast.

Meanwhile, OpenAI is approaching $25B ARR with IPO signals for late 2026. They’re scaling revenue and facing brand fragmentation simultaneously. That combination deserves watching.

Apollo Went Fully Agentic and Came for the Whole Stack

The noise: Apollo added AI features. Sales tool competition continues.

The signal: Apollo reframed itself as a replacement for legacy sales stack — not an add-on. End-to-end agentic workflows. 20,000 weekly users in week one.

This is the Revenue Nervous System pressure point I’ve been watching. The legacy stack — your Salesforce, your Outreach, your ZoomInfo, your Gong — was built for human-in-the-loop workflows. One system per step of the motion. Agentic infrastructure doesn’t need handoffs between systems because the agent crosses the system boundaries itself.

If Apollo can execute research, sequence, personalize, follow up, and log — without a rep touching it — the question isn’t whether it’s better than your current stack. It’s whether your current stack is a collection of parts that an agent can replace wholesale. For most teams I talk to, the answer is yes.

AI Advertising Hits $57 Billion and Buyers Are Searching in Chat

The noise: AI is coming for digital advertising. Google is threatened.

The signal: IAB Tech Lab launched the AAMP Agent Registry — an infrastructure layer for AI systems to buy ads autonomously. Google’s AI Mode hit 75 million users. AI advertising is projected at $57 billion in 2026, up 63% year over year.

The buyer journey is changing underneath your funnel. Your ICP isn’t starting their research on Google anymore — they’re starting in conversational interfaces. If your content isn’t structured for retrieval by AI systems, you’re invisible at the top of the funnel in a way you can’t see in your analytics.

Post-sales is pre-sales. And pre-sales now includes what the AI says about you before the buyer ever clicks.

The 65-Point Gap Is Where the Race Is Being Run

Here’s the number that ties the whole week together.

45,000 tech layoffs in March. 20% AI-attributed — up from less than 8% in 2025. 76% of organizations are deploying agentic AI. Only 11% have it in production.

76 minus 11 is 65. That’s the gap between running a pilot and building infrastructure. And that gap is where the competitive race is being run right now.

Every story from this week points at the same shift. Dorsey and Oracle aren’t cutting because AI is replacing jobs today — they’re cutting because they believe AI will replace jobs at scale, and they’re building toward that world. Google Stitch isn’t threatening Figma because it’s better software — it’s threatening it because it moves design from a discrete workflow step to an input in a larger agentic chain. Microsoft and Anthropic aren’t shipping enterprise features — they’re shipping the connective tissue that makes AI the operating layer, not the tool layer.

The organizations in the 11% aren’t smarter. They’re not better funded. In most cases they’re not even better at AI. They’re better at one thing: they stopped treating AI as a project and started treating it as infrastructure. Infrastructure doesn’t get a quarterly review. It gets engineered.

Revenue is engineered, not hoped for. And right now the engineering is happening inside that 65-point gap.

That’s the week. The executives who are cutting now are the ones who decided the bet is real. The question is whether you’re building the infrastructure or watching it get built around you.

The second kind starts from a blank page and asks a completely different question: if we were designing this process today, knowing what AI agents can actually do, how would we build it?

That gap—between retrofitting AI onto old processes versus building new ones with AI at the center—is going to define which companies win the next decade. And most companies are firmly, stubbornly camped in column one.

Why “AI on top” is a trap

There’s a seductive logic to layering AI onto your existing workflows. It’s lower risk. It’s easier to explain to your team. You don’t have to blow anything up.

But it ignores a fundamental problem: most of our workflows weren’t designed with AI’s actual capabilities and constraints in mind. They were designed for humans, augmented by software. That’s a completely different architecture.

AI agents have real limitations. Context windows aren’t infinite. Handoffs between systems are clunky if you don’t think carefully about data flow. If you just drop an agent into a process built for a human, you’re not getting leverage—you’re getting a more expensive workaround.

But AI agents also have capabilities that most processes are completely failing to use. They can run in parallel at a scale that no team can match. They can write and execute code on the fly. They can call APIs, transform unstructured data, and connect disparate systems without waiting on IT to build a custom integration. They don’t get tired at 4 PM on a Friday.

When you layer AI onto an old workflow, you get maybe 20% improvement. When you redesign the workflow from scratch around what AI can actually do, you get something unrecognizable—in a good way.

The “infinite engineers” question

Here’s the mental model that’s been reshaping how I think about this.

When an agent can write code and interact with any API, it’s not just an automation tool. It’s effectively an expert engineer you can deploy against any business problem, at will, without a headcount request or a six-month hiring process.

So the question I’ve started asking about any process is this: what would you do differently if you had an infinite number of capable engineers who could write software specifically for this workflow?

Not a generic software platform. Not an off-the-shelf tool. Engineers who would actually sit down, understand your process end-to-end, and write custom code to solve your specific version of the problem.

What if those engineers connected your disparate data sources so every part of the process had the full context it needed? What if they wrote automation for every repeated task, no matter how small? What if they built custom integrations between your systems that no vendor was ever going to prioritize? What if they could comb through thousands of rows of unstructured data and surface exactly what each step of your workflow actually needed?

When I ask that question, most processes look completely different.

This isn’t theoretical

I’ve seen it get real in places you wouldn’t expect.

In marketing, we’re used to thinking about content production as a human-driven workflow with AI writing assistance sprinkled in. But if you redesign from scratch—treating agents as the engineering backbone—you’re not producing content faster. You’re running continuous audience signal analysis, dynamically adjusting messaging, and triggering distribution workflows based on engagement patterns, all in parallel, all automatically. That’s a different function.

In finance, the traditional close process involves a lot of humans moving data between systems, reconciling exceptions, and waiting on approvals. An agent that can write code and hit APIs doesn’t need most of that scaffolding. It becomes something closer to continuous reconciliation—not a monthly sprint.

In sales operations, think about how much time goes into enriching data, researching accounts, and building the context a rep needs before a conversation. Most of that is repeated, structured work that a capable engineer would have automated three years ago if anyone had asked. Agents can do it now, at scale, for every account, every time.

Not every process has this upside—I want to be clear about that. Some workflows are genuinely human judgment-intensive in ways that don’t yield to this kind of redesign. But there are far more that aren’t than most leaders realize.

The real barrier isn’t technology

Here’s what I keep running into: the companies that are stuck in column one aren’t stuck because the technology isn’t ready. They’re stuck because they’re asking the wrong people to redesign the workflows.

Process owners—the people who know how workflows actually run—are often the last people to imagine what those workflows could look like if built from scratch. They’re too close to it. They optimize incrementally. That’s not a criticism; it’s how expertise works.

The redesign question requires a different kind of thinking. It requires someone who can look at a process and ask: “If I weren’t constrained by how we’ve always done this, and I had access to agents that could write and run code, what would I actually build?”

That’s a product mindset applied to operations. And most organizations don’t have many people operating that way.

The teams that figure it out—that start treating AI agents as their unlimited engineering capacity—will run circles around the ones still optimizing their old playbooks. Not because they have better AI access. Everyone has access.

Because they asked a better question first.

It’s a series of LEGO images. First, a pile of loose bricks in every color — total chaos, no structure. Then sorted by color. Then arranged in neat rows. Then stacked into clean columns as a bar chart. Then, finally, built into a house.

Five stages. Same bricks. Radically different outcomes.

The label at the top reads: DATA.

The label at the bottom reads: EXPLAINED WITH A STORY.

I’ve been thinking about it for three days. Because that image isn’t just about data visualization. It’s about where AI is about to fundamentally break the enterprise open — and why most companies are still living in stage two or three wondering why nothing works.

We’re All Sitting on a Pile of Bricks

Here’s what I’ve seen at every company I’ve worked with — from Verizon’s $260M P&L to REEF’s physical network to what we’re building right now at Experity: organizations have more data than they’ve ever had, and fewer answers than they’ve ever needed.

The data exists. Usually a lot of it. CRM records, EHR transactions, customer interactions, pipeline metrics, operational logs. Thousands of bricks in a pile.

And for years, the solution everyone sold was: sort the bricks.

Build a data warehouse. Implement a BI tool. Create dashboards. Clean the data. Tag it. Organize it into neat little color-coded piles. That’s the “sorted” phase. And look — that’s real progress. I’m not dismissing it. But sorted bricks are still just bricks. They don’t tell you what to build.

Then came the “arranged” phase: analytics. Slicing and dicing. Segmentation. Cohort analysis. More sophisticated sorting with pattern recognition layered on top. A huge improvement. Still not a house.

Then “presented visually”: BI dashboards. Tableau. PowerBI. Data storytelling became a profession. This is where a lot of companies are right now, and they feel pretty good about themselves — because the chart looks clean and the executive presentation has color-coded bars.

But here’s the thing. A bar chart of sorted LEGO bricks isn’t a house. It’s a picture of organized potential.

The Gap

There’s a massive valley between “we have insights” and “we have answers.”

Most organizations I’ve encountered have become very, very good at generating data, organizing data, and presenting data. What they haven’t cracked is explaining data in a way that drives immediate, contextual action.

That last image — the house — that’s not just a better visualization. It’s a completely different product. The same bricks, assembled by someone (or something) that understood what needed to be built, not just what bricks existed.

This is the central promise of AI agents. And most conversations about it are still stuck in the “sorted” and “arranged” stages.

I talk to operators every week who are deploying AI in their businesses. And when I ask what it’s actually doing, the answer is usually some version of: it’s making our existing processes faster. Faster sorting. Faster arranging. Better dashboards. Summaries of reports that used to take hours.

That’s real value. But it’s optimizing the first four images, not building the house.

What Agents Actually Change

The reason the fifth image — the house — is different isn’t because someone worked harder. It’s because someone shifted from organizing information to applying judgment.

Judgment about what to build. Judgment about what matters. Judgment about which bricks belong together to create something useful.

That’s what AI agents are designed to do — and it’s why the agentic AI movement is categorically different from “AI tools that help you work faster.”

An agent doesn’t just retrieve data. It reasons across data. It identifies what’s missing. It takes action based on what it finds. It communicates outcomes in context — not just as another row in a dashboard, but as a decision-ready story.

Think about what that actually means at ground level. At Experity, we’re working on this with Care Agent in urgent care. The data exists — patient visit patterns, operational bottlenecks, billing outcomes, scheduling gaps. For years, that data has been available. Sorted. Arranged. Presented visually.

But a patient who leaves without being seen isn’t prevented by a better dashboard. An agent that detects the early signals of walkout risk and communicates them to the right person in the right moment — that’s a house. That’s data doing actual work.

Same bricks. Entirely different outcome.

The Hierarchy Is a Roadmap

What I love about that LEGO image — beyond how elegantly simple it is — is that it maps almost perfectly to where different organizations are right now in their AI maturity.

Most enterprises are solidly in stages two and three. They’ve made real investments in data infrastructure and business intelligence. They’re proud of their dashboards. They’re not wrong to be.

The leading edge — the early movers in AI-native operations — are beginning to crack stage four. Their AI tools can generate reasonably good narratives from data. They’re surfacing insights. They’re reducing the time between data collection and human comprehension.

But almost nobody is consistently living in stage five at enterprise scale. The companies that figure that out first — agents that don’t just analyze data but build the house, automatically, in context, and communicate it as a story that drives action — those companies are going to have an operating advantage that compounds fast.

The Question You Should Actually Be Asking

Most of the AI conversations I’m in right now are still about stage two and three efficiency: how do we process more data, faster?

Those aren’t bad questions. But they’re the wrong questions if you’re trying to build durable competitive advantage.

The better question is: where in our business are we sitting on sorted bricks when we need a house?

Where do we have data, dashboards, even solid analytics — but the output is still a human staring at charts trying to figure out what to do next? Where is the gap between insight and action still filled by heroic individuals rather than systematic processes?

Because that gap? That’s where agents live.

That’s the space that’s about to get rebuilt from scratch over the next 18-36 months. Not the analytics. Not the dashboards. The actual decision-making layer — the moment between “here’s what the data says” and “here’s what we’re doing about it.”

A Note on What This Isn’t

I want to be clear about something: stage five isn’t magic. The house still requires good bricks.

If your data is dirty — if your CRM is a mess, if your customer records are incomplete, if your operational data is siloed across 11 different systems with no coherent structure — an AI agent won’t fix that. Agents built on bad data don’t build houses. They build unstable structures that look impressive until something breaks.

This is why I push back when people talk about AI as a shortcut around the messy data infrastructure work. It’s not. Stages two, three, and four still matter. You can’t skip to the house without getting the bricks right first.

But — and this is the part most organizations are missing — you can be doing stages two, three, and four with the explicit goal of enabling stage five. The question you’re building toward matters. “How do we get better dashboards?” is a stage three goal. “How do we enable agents to act on this data in real time?” is a stage five goal. They produce different infrastructure decisions, different integration priorities, different data quality investments.

Most companies are optimizing for what they have now instead of what they’re trying to build next.

Build the House

That LEGO image is going to keep showing up in my presentations because it does something unusually well: it makes the abstract concrete.

We’ve been talking about “data-driven decisions” for 20 years. We’ve built enormous infrastructure around making data available, organized, and visible. And we’ve gotten good at it.

But visible data isn’t the same as activated intelligence. A sorted pile of LEGOs is better than a chaotic pile. A bar chart of sorted LEGOs is better still. None of it is a house.

The companies winning the next decade aren’t going to win because they have better dashboards. They’re going to win because they figured out how to put the bricks together in ways that build something real — faster, more accurately, with less heroic human effort filling the gaps.

That’s the game. Stop reorganizing the pile.

Build the house.

Pull up almost any SaaS stock chart right now. Scroll through the list. Figma down 76% from its 52-week high. Trade Desk down 74%. Duolingo. HubSpot. Monday. Atlassian. The names go on and on—and almost every single one is a bloodbath.

Someone shared a “Software Meltdown” tracker on Twitter recently—35+ public SaaS companies, neatly arranged in rows of red. Over 90% of them are trading 30-80% below their highs. Not their all-time highs from 2021. Their 52-week highs. We’re not talking about a hangover from pandemic valuations. This is a fresh wound.

And I keep seeing smart people treat this like a blip.

It’s not a blip.

What the Numbers Actually Tell You

Here’s the thing about financial tables like this one—they’re a Rorschach test. You see what you want to see.

Optimists look at this and say: “Overblown. Markets overcorrected. These are great businesses at discount prices.”

Pessimists look at this and say: “SaaS is dead.”

I look at this and see something in between—but much closer to the pessimist end than most operators want to admit.

These aren’t bad companies. ServiceNow is a generational business. Salesforce built the modern CRM category. Datadog is genuinely excellent infrastructure. These are not fly-by-night SPACs.

And yet: -50%, -40%, -39%. Across the board.

The market is saying something. The question is whether we’re actually listening.

The Business Model Problem Nobody Wants to Name

Here’s my read: the SaaS business model as it was designed is getting stress-tested in real time—and a lot of these companies are failing the test.

The classic playbook looked like this: charge a per-seat subscription, grow revenue by adding seats, expand within accounts by adding modules. Land, expand, retain. Print money forever.

It worked beautifully for about fifteen years.

Then a few things happened simultaneously.

First, buyers got exhausted. The average mid-market company now has 130+ SaaS tools. Nobody knows who’s using what. Finance is running audits. CFOs are cutting anything that can’t prove ROI in a single quarter. The “expand” motion hit a wall.

Second, the macro tightened. When money was free, growth-at-any-cost was the game. Investors rewarded forward multiples. Now they want cash flow. That shifted the entire incentive structure of how these companies operate—and the best ones are adapting fast, but most aren’t.

Third—and this is the one I watch most closely—AI entered the building.

And it’s not leaving.

The AI Tax on Traditional SaaS

I want to be careful here because this argument gets oversimplified constantly.

“AI will kill SaaS” is too reductive. But “AI won’t change the SaaS business model at all” is delusional.

Here’s what I actually believe: AI is applying a structural tax on seat-based SaaS.

Think about it practically. If I can deploy an AI agent that does the work of three SDRs, why am I paying for three SDR seats in my CRM? If I can use AI to automate reporting, why do I need five analysts licensed in my BI tool?

The headcount-to-software relationship is breaking down. And a huge chunk of SaaS revenue is anchored to headcount.

This isn’t hypothetical. It’s already showing up in net revenue retention numbers. Companies that were at 120%+ NRR are sliding back toward 100%. Some below. The expansion motion that powered the entire growth thesis for this category is slowing in real time.

The Survivors vs. The Casualties

Not all of these red numbers are the same story.

There’s a difference between companies that are going through painful repricing and companies that are facing genuine existential pressure.

Look at Zoom—down only 6.65% from its 52-week high in this dataset. Zoom rebuilt itself. They moved aggressively into AI, rethought the product surface, and accepted that they’d never recapture their pandemic multiple. That’s discipline.

MongoDB is up 33% over the last year. Snowflake has a positive year. Confluent is basically flat.

Notice anything about those names? They’re infrastructure. Data and developer tools. Not workflow automation, not productivity apps, not feature-software that AI can increasingly replicate.

The pattern here is pretty clear if you’re willing to see it: software that sits below the workflow—in the data layer, the infrastructure layer, the integration layer—is holding up better than software that sits in the workflow.

Because AI is eating workflows. It’s not yet eating the pipes that data runs through.

What This Means for You

If you’re an operator, you need to make two kinds of bets right now.

Bet one: Audit your stack ruthlessly. If you’re paying for SaaS seats that deliver outputs AI can now produce—automate, consolidate, or cut. Your competitors who are doing this are getting structural cost advantages you can’t close with hustle.

Bet two: Think hard about which tools in your stack have infrastructure leverage. The SaaS products that own your data, your integrations, or your core operational record are defensible. The ones that own workflows—especially repeatable, language-based workflows—are not.

If you’re in SaaS, building SaaS, or selling SaaS, the question you should be asking isn’t “how do we add AI features?” It’s “does our product live above or below the AI automation line?”

That question will determine whether your company ends up in the red column or the green one over the next three years.

A Final Thought

I’ve worked in and around technology long enough to have seen a few of these cycles. The dot-com crash. The 2008 correction. The 2022 SaaS repricing.

Every time, two things happen.

The majority of people insist this is temporary. And then a meaningful portion of the incumbents don’t make it through—not because they weren’t good businesses, but because they couldn’t adapt fast enough to a structural shift that was actually happening.

This feels structural to me.

The companies that make it out the other side will be the ones that got honest about it early.

Are we being honest yet?

Pre-seed. Two people on the team. Except the second person is Claude.

He doesn’t call it a tool. He calls it his technical co-founder.

I laughed when he said it. He didn’t.

That sentence stopped me cold. Because I’ve known this guy for years. He’s not the type to reach for hyperbole. He doesn’t say things he doesn’t mean. So I did what any reasonable person would do when confronted with a statement like that.

I asked him to walk me through his actual day.

What I heard over the next thirty minutes made me put my phone down and stare at the ceiling for a while.

6 AM Standup. With an AI.

Every morning, he starts his day the same way. He pulls up Claude Code, looks at his overnight metrics, and starts talking through what happened while he slept.

“Claude flags what moved, what broke, what needs attention,” he told me. “Then we prioritize the day.”

We.

He said “we” four times before I realized there was nobody else there.

No second desk. No Slack ping from a human co-founder. Just him and a model that, apparently, gives a damn about his retention numbers at 6 in the morning.

I’ve watched a lot of operators describe their AI workflows over the past two years. Most of them describe augmentation — Claude helped me write this, Claude summarized that. It’s a feature. It’s a speed boost.

This was different. He was describing a relationship. A working rhythm. The kind of thing you develop with a person you trust enough to think out loud with every day.

The Deck That Got the Meeting

A few weeks ago, an investor asked him to turn around an updated deck in 24 hours.

If you’ve ever been a founder, you know what that request actually means. It means the investor is interested but hasn’t committed, the current deck isn’t doing enough work, and you have one shot to tighten the narrative, rebuild the financials, and make the case stick — all overnight.

He said “Claude and I went dark for a full afternoon.”

He talked through the narrative. Claude rebuilt the financials, restructured the slides, rewrote the copy.

They sent it at 11pm. He got the meeting.

Now — I want to be careful here, because this is the part where a lot of AI content gets sloppy. Claude didn’t close the deal. Claude didn’t walk into that investor meeting and charm anyone. My friend still had to do all of that. The human judgment, the relationship, the conviction — that’s all him.

But here’s what I keep thinking about: in the old world, a pre-seed founder with no co-founder and no team would have sent a B+ deck because that’s all one person at midnight could produce. He sent an A- deck. Maybe an A.

That gap matters.

The Sparring Partner Nobody Talks About

Here’s the part that got me the most.

He preps for every investor call using Claude as a sparring partner. Not like flashcards. Not like a practice run where you recite your talking points into the void.

Claude plays the skeptical VC.

It asks the hard questions. It pokes holes in the unit economics. It pushes back on assumptions he’s gotten too comfortable with.

“Claude asked me a question about churn last Tuesday,” he said, “that no actual investor has been smart enough to ask yet.”

Read that again.

An AI model asked a harder churn question than the humans who do this for a living.

I don’t say that to embarrass any investors. I say it because it points at something important: when you have a sparring partner who has no ego in the outcome, no relationship to protect, no reason to soften the blow — you get different questions. Better questions. The kind that find the cracks before someone else does.

This is one of the most underrated use cases in the entire AI conversation. Not productivity. Not content. Preparation. Having something in your corner that is constitutionally incapable of telling you what you want to hear.

40 Emails. Every One Different. 11 Replies.

His cold outreach process is where the thing becomes genuinely hard to explain without sounding like a product demo.

He writes one draft. One. Then Claude rewrites it for each prospect — based on their LinkedIn activity, their company’s recent news, what their likely objections are, what they probably care about this quarter.

40 emails last week. Every single one different. He got 11 replies.

For context: the average cold email reply rate in B2B sits somewhere between 1-5%. He’s not operating in that world.

What he’s doing isn’t spray-and-pray at scale. It’s personalization-at-scale, which is a category that wasn’t really available to a solo operator eighteen months ago. You either had the headcount to do real research on every prospect, or you sent generic blasts and hoped volume covered for relevance. There was no door three.

Now there is. And he walked through it.

“I’ve Had Human Co-Founders Who Did Less.”

I asked him, eventually, whether it ever felt strange. Having an AI as his co-founder. Talking to it every morning. Saying “we” without a second human in the room.

He paused for maybe two seconds.

“I’ve had human co-founders who did less.”

I’ve been sitting with that line for a week.

There’s real complexity buried in it. Human co-founders bring things Claude can’t — accountability, shared risk, judgment forged by failure, the ability to look you in the eye at 2am when everything is on fire and say we’re going to figure this out. That’s not nothing. That’s actually a lot.

But he’s not making a philosophical argument. He’s making a practical one. In the day-to-day mechanics of building a company — prep, analysis, writing, research, iteration — Claude Code shows up. Every morning. Does the work. Doesn’t check out. Doesn’t have a competing priority. Doesn’t need equity or a title or a three-month onboarding ramp.

For a pre-seed founder trying to survive long enough to find product-market fit, that consistency is worth more than most people are willing to admit.

What This Is Actually Telling Us

I’ve been in and around growth and operations for 21 years. I’ve watched technology change how teams work dozens of times. Usually the change is additive — new tool, new capability, roughly same organizational structure underneath.

What my friend is describing isn’t additive. It’s structural.

He didn’t hire Claude to help his process. He built his entire operating model around the assumption that Claude would be there, would be capable, and would keep getting better. The team design came second. The tooling came first.

That inversion is new.

Most companies are still in the phase where they hand Claude to a team that was built for a different era. Here’s a guy who built the team for the era we’re actually in.

Two things I keep coming back to:

First, the productivity gap between a founder with a well-developed AI workflow and one without is growing faster than most people realize. He’s not slightly more efficient. He’s operating at a different capability level. And the delta compounds — every week he gets better at working with Claude, the gap widens.

Second, we’re still in early innings on what “AI co-founder” actually means. Right now it means a really good thinking partner and execution accelerator. In 18 months? I genuinely don’t know. The capability curve is steep enough that I’d feel foolish trying to put a ceiling on it.

What I know is this: the question used to be whether AI could do meaningful work. That conversation is over. The question now is whether founders — and operators, and teams — are designing their organizations to take the work seriously.

My friend is. Every morning at 6am.

The rest of us are still arguing about whether it counts.

But in three years, we might be talking about a paralegal shortage. An accountant shortage. A customer service rep shortage—because the people who used to fill those seats figured out what the data already shows and got out early.

Andrej Karpathy (yeah, the former Tesla AI director and OpenAI founder) built something that should be required viewing for anyone with a salaried job. He analyzed every occupation in the US economy—143 million jobs total—and scored each one on AI exposure from 1 to 10, pulling data straight from the Bureau of Labor Statistics Occupational Outlook Handbook. You can see it yourself at karpathy.ai/jobs.

I’ve looked at a lot of “AI will replace jobs” content over the years. Most of it is vague enough to mean nothing. This one hits different.

Because it’s a map. And the pattern on the map is impossible to ignore.

Here’s What the Data Actually Shows

Let me give you the short version before we dig in:

Job Exposure Score

• Plumber 1/10

• Firefighter 1/10

• Carpenter 2/10

• Nurse 2/10

• Surgeon 2/10

• Personal Trainer 2/10

• Chef 2/10

• Teacher 5/10

• Software Developer 5/10

• Real Estate Agent 5/10

• Marketing Manager 5/10

• Accountant 7/10

• Lawyer 7/10

• Financial Analyst 7/10

• HR Specialist 7/10

• Customer Service Rep 9/10

• Cashier 9/10

• Receptionist 9/10

• Bookkeeper 9/10

• Data Entry Clerk 10/10

Look at that distribution for a second. Really look at it.

The jobs with the lowest exposure scores share something obvious once you see it: they require a physical body in a specific location, doing something that requires reading a room, a pipe, a patient. The jobs with the highest scores share something equally obvious: they’re mostly about receiving information, processing it, and outputting something back.

That’s it. That’s the whole thing.

If your job is “take info in, process it, give info out”—you’re in trouble.

If your job requires being physically present or earning human trust face-to-face—you’re probably fine.

Why This Pattern Makes Sense

We built an entire white-collar economy around tasks that AI is genuinely good at.

We called it “knowledge work.” We made it prestigious. We went $200K in debt to get the degrees that unlocked the door to it. And now the machines can do the core of it—not perfectly, not completely, but well enough that the economics of employment are going to shift.

A data entry clerk making $35K processes structured information for 8 hours a day. That is quite literally the textbook definition of what a language model does. It’s not a knock on data entry clerks—that work has real value—but the job description reads like a capability list for a 2022 AI demo.

The counterintuitive flip in this data is who’s actually safe.

The plumber. Exposure score: 1.

Think about what a plumber does. They show up to a house they’ve never seen before, crawl under a sink or into a crawl space, diagnose a problem that’s never identical to the last one, physically manipulate tools in a three-dimensional space, and then interact with a homeowner who’s stressed, skeptical, and making a judgment call about whether to trust them. All in the same visit.

AI cannot do that. Not today. Not in 2027. Probably not by 2030 in any meaningful deployed sense. The bottleneck isn’t intelligence—it’s embodiment, physical dexterity, and presence.

Meanwhile, the lawyer hits a 7/10. Not because lawyers aren’t smart or valuable—they absolutely are—but because a huge portion of legal work is exactly the pattern: ingest documents, identify relevant precedent, synthesize into an argument, output a memo or brief. The reasoning is sophisticated. The underlying mechanics are still information processing.

The Thing Nobody’s Talking About: The Middle Is Where It Gets Complicated

The 1s and 10s are easy to reason about. The 5s are where people are going to get blindsided.

Software developers score a 5. And I’ve already watched this play out in real time.

I talk to engineering leaders every week. The ones at companies actively deploying AI coding tools are telling me they’ve seen a 30-40% productivity jump on boilerplate work. Junior developers who used to spend their first year writing CRUD functions are now doing it in a fraction of the time.

That’s not a 5-year trend. That’s happening right now.

But here’s what’s also true: the senior engineer who can architect systems, who understands which abstractions break under load, who can walk into a room with a VP and translate business requirements into technical decisions—that person isn’t going anywhere. They’re actually getting more valuable, because now they can do more with less scaffolding.

Same with marketing managers. A 5 makes sense. The part of marketing that’s pure information processing—trend reports, campaign performance analysis, first-draft copy generation—is being automated. The part that’s judgment, taste, understanding what a brand stands for in a specific cultural moment, knowing when to break the rules—that’s still human territory.

Teaching is an interesting 5. The lecture component—explaining a concept clearly—AI can do passably. The relationship component, knowing which student is checked out because something’s going on at home versus which student genuinely doesn’t understand the material, adjusting mid-lesson when you read a room wrong—that’s irreducibly human.

The 5s are going to bifurcate. Some of those jobs will effectively become 8s as AI eats the repetitive core. Some will drop to 2s as the human-judgment component gets isolated and protected. Which direction your specific role goes will depend entirely on which half of the job you spend your time on.

What This Means If You’re Thinking About This Seriously

I’m not going to pretend I have a clean three-step playbook here. I don’t. The honest answer is that nobody knows exactly how fast this moves or where it stops.

But a few things seem clear:

Proximity to physical reality is going to be undervalued for a while. There’s a generation of people who were told that getting your hands dirty was for people who didn’t have other options. That narrative is going to flip. The skilled trades have a real labor shortage already. If you’re 22 and good with your hands, you might be sitting in a better position than the 22-year-old who took the paralegal job.

The “information middleman” role is probably where the most disruption concentrates. Not the expert who generates the insight—but the layer between the expert and the decision-maker, whose job is to package and transmit. That layer is thin already and getting thinner.

Human trust is becoming an actual competitive moat. This sounds fluffy until you think about it concretely. The financial advisor who’s been your family’s financial advisor for 15 years—who knew your father, who was there when you got divorced, who you called when your business was struggling—that person isn’t competing with a chatbot. The trust layer is structural, not just relational. It’s built on presence and time and mutual vulnerability. AI can’t fake that. At least not yet.

The $3.7 trillion in wages exposed to high-AI-impact jobs is a macro signal. Karpathy’s tool shows $3.7 trillion in wages across high-exposure occupations. That’s not a rounding error. If even a fraction of that productivity gets absorbed by AI tools over the next decade, the downstream economic effects—on employment, on wages, on where people choose to invest in education—are going to be significant. This isn’t a technology story anymore. It’s a macroeconomics story.

The Question I Keep Coming Back To

I spent a good chunk of my career working in operations—figuring out how to do more with the resources available, which increasingly means figuring out how to deploy technology against tasks that used to require headcount.

And I’ve watched smart, capable people get caught flat-footed by this shift. Not because they were lazy or unaware. But because they were heads-down building expertise in the thing that was working, while the ground shifted under them.

The heat map Karpathy built is the most honest version of that ground shift I’ve seen in a visual format. It’s not a prediction. It’s a description of structural vulnerability based on what these jobs actually require people to do.

The question worth sitting with isn’t “is my job on the list?”

It’s: “Which half of my job is on the list?”

Because that’s where the real work starts—figuring out which parts of what you do are irreducibly human, doubling down there, and being honest about what’s already being eaten.

The map is out. The only move left is to read it.

OpenAI hit that moment last week. And they at least had the decency to be honest about it.

Fidji Simo—the CEO of applications who ran Instacart before Sam Altman brought her in last August—stood in front of OpenAI employees and told them the company couldn’t afford to be distracted by “side quests.” The message was simple: stop chasing everything. Start winning where it actually matters.

What she didn’t say out loud, but what the numbers say clearly: the consumer AI chatbot business they built ChatGPT into is commoditizing so fast it can’t finance the losses required to stay competitive in it.

The Numbers Tell the Story

OpenAI generated $13.1 billion in revenue last year. That sounds like a lot until you see the other column: $8 billion in operating losses.

The internal projections from there are the part that should stop you cold. The company is looking at $14 billion in losses for 2026 alone. Cumulative losses reaching $143 billion by 2029.

Let that sit for a second. $143 billion in the hole before they expect to turn profitable.

For context: Amazon burned through roughly $3 billion over eight years before finding its footing. Uber—often held up as the poster child for burning investor money—ran a $31 billion cumulative deficit before reaching profitability. OpenAI is projecting nearly five times Uber’s hole. And they expect to still be filling it three years from now.

The chart that’s been circulating from media reports plots OpenAI’s projected losses against Amazon, Spotify, Tesla, and Uber. Every other company on that chart looks like a rounding error compared to what OpenAI is planning to spend.

This isn’t a startup burning cash to buy market share. This is a company building something that requires a level of capital so extreme that winning the consumer market—even winning it decisively—might not cover the cost of getting there.

What “Winning” Consumer AI Actually Looks Like

ChatGPT had an 86.7% share of web traffic among AI chatbots twelve months ago. Today that number is 64.5%.

Google Gemini went from 5.7% to 21.5% over the same period.

Those aren’t just competitive losses. That’s the consumer AI market compressing in real time—multiple capable, well-funded products chasing the same users at the same price points. The floor on what a good AI model costs to run keeps dropping (more on that in a minute), which means the product that got people to pay $20/month is facing the same fate as every software category that got commoditized before it.

The consumer subscription math doesn’t work at this scale. It never really did. A $20/month Pro subscriber would need to renew for 7,150 months to cover their share of the $143 billion in projected losses. That’s not a business model—it’s a research project with a checkout page bolted on.

Where the Real Money Lives

Meanwhile, something different was happening on the enterprise side.

Codex—OpenAI’s coding agent product—crossed 1.5 million weekly active users with 20x growth since August 2025. Enterprise seats grew 9x year over year. 92% of Fortune 500 companies are currently using ChatGPT products in some capacity.

The revenue per user in enterprise isn’t just higher—it’s a different order of magnitude. And the switching costs are astronomical compared to a consumer subscriber who can cancel in two clicks and sign up for a competitor the same afternoon.

Simo ran Instacart. She understands unit economics better than almost anyone. When unit economics break in consumer, you go where the margins are real and the customers can’t easily leave. Enterprise software sales have known this for twenty years. It’s why Salesforce isn’t trying to win the consumer CRM market.

Coding agents and enterprise API access are that business. Sora and robotics and whatever else falls under “side projects” are capital incinerators with no visible revenue timeline. Cutting them isn’t a strategic failure—it’s a company finally deciding to be honest about what it’s actually building.

The DeepSeek Problem Nobody Wants to Talk About

Here’s the thing that makes this pivot more urgent than most coverage acknowledges.

DeepSeek—the Chinese AI lab that triggered a global stock selloff earlier this year—has matched GPT-5 level performance at somewhere between one-tenth and one-thirtieth the cost.

One-thirtieth.

That’s not a competitive advantage eroding at the margins. That’s the entire competitive moat for consumer AI products evaporating. If the model quality that justifies a premium subscription can be replicated at 3% of the training cost, the floor on consumer AI pricing is falling faster than anyone expected. You can’t build a $143 billion loss-recovery plan on a product where the commodity version is nearly as good as yours and costs next to nothing.

The only defensible position in a commoditizing market isn’t model quality. It’s integration depth. Switching costs. The enterprise customer who has Codex embedded in their engineering workflow, their CI/CD pipeline, their code review process—that customer doesn’t leave because a cheaper model exists. They leave when something better and more integrated comes along. That’s a much higher bar.

OpenAI is trying to be that product. The pivot to coding and enterprise isn’t abandoning their vision. It’s acknowledging that their vision only survives if they own the infrastructure layer that enterprises can’t easily rip out.

The Part That Doesn’t Get Said Out Loud

OpenAI was founded on a specific mission: AGI for humanity. Ensure that artificial general intelligence benefits everyone. Sam Altman has said versions of this in every major interview for the past five years.

The strategy Simo outlined last week isn’t AGI for humanity. It’s Codex for Fortune 500 engineering teams. Enterprise API contracts. Productivity software for developers who can expense the usage fee.

Those are different companies. Not incompatible—plenty of companies serve both a mission and a market—but the gap between “AGI for humanity” and “the coding copilot preferred by Goldman Sachs’ engineering team” is wide enough to drive a truck through.

I don’t think this makes OpenAI villainous. I think it makes them a company, which is what they’ve been for a while even if the nonprofit origin story lingered in the branding. The $280 billion revenue target by 2030 requires roughly 2.3x growth this year, 2x in 2027, and sustained compounding from a $25 billion ARR base. That math only works if they own the enterprise infrastructure layer. Consumer subscriptions at $8 and $20 a month do not get you to $280 billion. Codex at usage-based pricing across every Fortune 500 engineering team might.

But it’s worth naming what changed. The company that was going to democratize intelligence is pivoting to selling it as a service to companies that can pay for it at scale. That’s not wrong. It’s just honest.

If You’re Watching from the Outside

A few things I keep coming back to:

The enterprise coding market just became the most important battleground in AI. Anthropic’s Claude Code, OpenAI’s Codex, GitHub Copilot—these aren’t just developer tools. They’re the integration points that determine which AI company owns the enterprise relationship long-term. Whoever gets embedded deepest in how engineering teams work builds a moat that won’t matter how cheap the next open-source model gets.

Consumer AI is heading toward the same fate as cloud storage. Commoditized, mostly free at entry level, real revenue only at enterprise scale. The companies still trying to win consumer AI on model quality alone are going to have a rough 2026.

Fidji Simo is probably the most important hire in OpenAI’s history. Not because she’s smarter than anyone else there. Because she’s the one who ran a company with brutal unit economics—grocery delivery, where every order can lose money if you’re not obsessive about the math—and learned how to find the customer segments where the numbers actually work. OpenAI needed someone who’d been inside that kind of pressure. They found one.

The mission statement and the business strategy have officially diverged. This doesn’t mean OpenAI fails. It means they’re becoming a different company than the one they said they were building. That happens. The interesting question is whether the mission catches back up to the business once the business is sustainable, or whether it quietly disappears from the About page over the next few years.

The consumer AI race was always going to end this way. Too many capable competitors, too much capital chasing too few paying customers, and the underlying models getting cheaper and better at a pace nobody can sustain a premium on forever.

What’s surprising isn’t that OpenAI is pivoting. It’s that it took this long for someone to say it out loud.

The question now isn’t whether enterprise AI is the better business. It obviously is. The question is whether OpenAI gets there before Anthropic locks up enough Fortune 500 contracts to make the whole race irrelevant.

They’re three to five years and $143 billion of losses away from finding out.

What are you seeing in how your enterprise is thinking about AI vendor selection? Curious whether the OpenAI/Anthropic competition is actually showing up in buying decisions or if most companies are still hedging across both.

LinkedIn exploded. The AI SDR crowd started sharing the $1,200/month vs. $22,000/month comparison again. “See? The math works. AI SDRs are the future. The human SDR is dead.”

There’s just one problem with that narrative.

Rox AI isn’t an AI SDR.

What Rox AI Actually Is (And Why It Matters)

Rox was founded by Ishan Mukherjee, former Chief Growth Officer at New Relic and co-founder of Pixie (acquired by New Relic). His co-founders include Stanford professor Chris Re — one of the most cited researchers in machine learning.

Their product is an AI-native CRM replacement. A “Revenue Operating System” that sits across the entire deal lifecycle, not just top-of-funnel prospecting. Their customers include Ramp, OpenAI, MongoDB, NVIDIA, Databricks, and Confluent — companies that could buy any AI SDR tool on the market and chose something fundamentally different.

Here’s what makes Rox interesting: they explicitly rejected the AI SDR replacement thesis.

One agent per account. The human always presses send. Augmentation over automation. The AI handles research, context assembly, meeting prep, deal intelligence, and follow-up drafting. The human handles judgment, relationships, and the conversations that move revenue.

Sequoia led the seed. General Catalyst led the Series A and this latest round. These aren’t investors who got confused about what they were buying. They looked at the AI SDR landscape — 50-70% tool churn, the 11x scandal, conversion rates 40% lower than human reps — and bet $1.2 billion on the opposite thesis.

The smart money isn’t validating AI SDR replacement. It’s validating full-stack revenue capacity reform.

The Number Everyone Ignores: 28%

Here’s the stat that should be on every AI SDR vendor’s pitch deck but isn’t:

Sales reps spend only 28-30% of their time actually selling.

Salesforce’s State of Sales report has tracked this for years. The breakdown of what eats the other 70%:

• CRM data entry: 17%

• Internal meetings: 15%

• Email and administrative tasks: 14%

• Prospect research: 14%

• Scheduling: 12%

Top performers manage to claw back a few points — getting to 35-40% selling time. That extra 5-10% translates to 5-8 additional selling weeks per year. It’s the single biggest differentiator between quota-crushers and everyone else.

Now think about what happens when an AI SDR floods that rep’s calendar with 3x more meetings.

The rep who was already spending 70% of their day on non-revenue work now has more meetings to prep for, more CRM records to update, more follow-ups to draft, more internal handoff notes to write. The AI solved the volume problem at the top of the funnel and made the capacity problem in the middle of the funnel actively worse.

Cheaper pipeline flowing into a constrained system doesn’t create revenue. It creates more expensive waste.

The MQL-to-SQL Graveyard

The biggest drop-off in every B2B funnel isn’t lead generation. It’s what happens next.

MQL-to-SQL conversion rates average 15-21% across industries. That means 80-85% of marketing-qualified leads never become sales-qualified opportunities. This is where pipeline goes to die — not because the leads were bad, but because the humans responsible for progressing them don’t have the time, context, or process support to do the work.

Now layer in AI SDR data:

• AI SDR meetings convert to opportunities at 15% vs. 25% for human-sourced meetings

• AI-only pipeline generates 2.6x less revenue than human pipeline in head-to-head tests

• Meeting show rates drop from 70-85% (human-booked) to 40-60% (AI-booked)

The AI SDR conversation obsesses over the cost of generating meetings. Nobody talks about the cost of wasting them.

When your MQL-to-SQL conversion is 15% and your AEs have 28% selling time, adding more meetings to the top of the funnel is like turning up the water pressure on a pipe that’s already leaking from six joints. The water isn’t the problem. The pipe is.

I Run 93 AI Agents. Here’s What I’ve Learned.

I’ve been building what I call a Revenue Nervous System — a network of 93 AI agents spanning marketing, sales, customer success, revenue operations, product, and engineering. Not a hypothetical architecture. Production agents doing real work every day.

Here’s what became clear after the first 90 days: the agents that generate the most value aren’t the ones that replace humans. They’re the ones that reclaim human capacity for revenue-generating activities.

The SDR agent that books meetings? Useful. The agent that auto-enriches every account with competitive intelligence, recent news, tech stack data, and conversation history before the AE’s first call? That’s the one that moves the revenue needle.

The difference is where the agent sits in the workflow:

Layer 0 — Volume generation (AI SDRs sit here)
Automated outreach, meeting booking, initial qualification. This is where 90% of the market is building. It’s the easiest layer to automate and the lowest-value per unit of pipeline.

Layer 1 — AE time reclamation
AI handles CRM updates, research, meeting prep, follow-up drafts, internal reporting. The goal: move AEs from 28% selling time to 50%+. This is where Rox AI is building. This is where the $1.2B valuation lives.

Layer 2 — Pipeline intelligence
AI scores, enriches, and prioritizes pipeline so humans focus on the 20% of deals that drive 80% of revenue. Instead of working 50 meetings equally, the AE works 10 meetings deeply.

Layer 3 — Handoff orchestration
The meeting-to-opportunity conversion gap (15% AI vs. 25% human) exists largely because context gets lost between the AI that booked the meeting and the human who runs it. Intelligent handoff — full context transfer, next-best-action recommendations, conversation guides built from the prospect’s actual engagement history — closes that gap.

You need all four layers working together. An AI SDR without layers 1-3 is a firehose pointed at a bottleneck.

The Three Capacity Layers (Fix These Before You Buy an AI SDR)

If you’re evaluating AI SDR tools right now — or if you’ve already deployed one and the results are underwhelming — here’s the diagnostic framework I use:

1. Measure Revenue Per Hour of Selling Time

Before you add pipeline volume, know your baseline. If your AEs have 28% selling time (roughly 11 hours/week of actual selling in a 40-hour week) and generate $X in pipeline per quarter, your revenue efficiency per selling hour is:

Quarterly pipeline / (11 hours x 13 weeks) = revenue per selling hour

Now model what happens if you increase selling time to 40% (16 hours/week) without adding a single new lead:

Same pipeline / (16 hours x 13 weeks) = you just got 45% more capacity for free

That’s your AE time reclamation opportunity. It almost always has higher ROI than adding more top-of-funnel volume.

2. Audit Your Qualification-to-Close Ratio

If your MQL-to-SQL is below 20% or your meeting-to-opportunity is below 25%, you have a qualification problem, not a volume problem. Adding an AI SDR will make this worse, not better, because AI-generated meetings convert at a lower rate.

Fix qualification first:

• Define explicit handoff criteria between AI-booked meetings and AE follow-up

• Build meeting prep packages that give AEs full context before the first call

• Score pipeline by engagement depth, not just demographic fit

3. Map Your Context Gaps

Where does information get lost in your sales process? Common context gaps:

• SDR-to-AE handoff: The AI booked the meeting but the AE doesn’t know what the prospect engaged with, what questions they asked, or what their actual pain point is

• Meeting-to-CRM: The conversation happened but the notes are thin, the next steps aren’t logged, and the follow-up email doesn’t reference anything specific

• Cross-team: Marketing ran a campaign that influenced the deal but Sales doesn’t know about it. CS has renewal intel that could inform the expansion conversation but it lives in a separate system.

Every context gap is a conversion leak. AI agents that bridge these gaps — transferring context automatically, assembling deal intelligence, generating personalized follow-up — deliver more revenue impact than AI agents that send more cold emails.

What the $1,200 vs. $22,000 Comparison Actually Misses

The comparison is directionally correct. AI SDR platforms range from $500-$5,000/month. A fully loaded human SDR in a high-cost market runs $9,400-$15,400/month (the $22,000 figure likely includes HCOL base + benefits + management overhead + recruiting costs amortized over a 14-18 month average tenure).

But the comparison only measures cost per meeting generated. It doesn’t measure:

• Revenue per meeting: Human-sourced meetings generate 2.6x more revenue

• Downstream capacity cost: Each meeting consumes 2-3 hours of AE time (prep, call, follow-up, CRM update). Lower-quality meetings consume the same time for less return.

• Opportunity cost: An AE spending an hour on a low-quality AI-booked meeting is an AE NOT spending that hour on a high-probability deal that needs attention.

• Domain reputation risk: AI sending 1,000+ emails/day from your domain in an era where Gmail’s Gemini AI is specifically filtering AI-generated outreach isn’t free. Deliverability degradation affects your entire outbound motion, not just the AI channel.

The real comparison isn’t cost per meeting. It’s revenue generated per dollar of total sales cost. When you run that math, the pure AI SDR replacement model usually loses to the hybrid model by 2-3x.

What Happens Next

The AI SDR market is consolidating. The first wave (2023-2024) was about headcount replacement: “fire your SDRs, buy our bot.” That thesis is producing 50-70% churn.

The second wave (2025-2026) is about revenue infrastructure: AI-native CRMs, deal intelligence platforms, capacity reclamation tools. Rox AI, Monaco ($35M from Founders Fund), and the “anti-AI SDR” camp are building here.

The second wave is where the real money gets made. Not because top-of-funnel doesn’t matter — it does — but because top-of-funnel without mid-funnel capacity is a waste accelerator.

Gartner predicts that by 2028, AI agents will outnumber sellers 10:1. But they also predict fewer than 40% of sellers will report that AI agents improved their productivity. That gap — more AI, not more productivity — is exactly the capacity trap.

The companies that win will be the ones that fix the pipe before they turn up the water pressure.

The Bottom Line

Rox AI is worth $1.2B because they’re solving the hard problem: making human sellers more effective during the 28% of their day they actually spend selling.

The AI SDR vendors fighting over meeting volume are solving the easy problem.

If you’re a revenue leader evaluating AI for your sales org, start with one question:

What is my AEs’ revenue per hour of actual selling time, and how do I double it?

The answer to that question will determine whether AI SDRs create ROI or just create more meetings that nobody has time to convert.

The $1,200 AI SDR is real. The $22,000 human SDR is expensive. But the 70% of AE time spent on non-revenue work is the most expensive thing in your entire GTM motion.

Fix that first. Then add volume.

UserGems is reporting 50-70% annual tool churn rates on AI SDR platforms. That’s not normal SaaS churn. That’s roughly double the turnover rate of the human SDRs these tools are supposed to replace. Companies are buying, deploying, watching pipeline numbers wobble, and ripping the thing out before the first contract renewal.

Gartner predicts 40%+ of agentic AI projects will be abandoned by end of 2027. S&P Global’s 2025 survey found that 42% of companies abandoned most of their AI initiatives — up from 17% just a year earlier.

This isn’t a blip. It’s a pattern.

The 11x Scandal Should Have Been a Wake-Up Call

If you want to understand where this market really is, look at what happened with 11x.

This was the AI SDR darling. $50M Series B from Andreessen Horowitz at a $350M valuation. Harry Stebbings on the cap table claiming ~$25M ARR. TechCrunch cover stories.

Then it came apart. TechCrunch exposed that 11x was claiming customers it didn’t have — logos like ZoomInfo and Airtable used without consent. Former employees told reporters the company was losing 70-80% of customers that walked in the door. Inc. Magazine called it AI’s potential “Theranos moment.”

Independent analysis estimated the real valuation at ~$31M versus the $350M they raised at. That’s paying 116x actual revenue.

11x wasn’t just a bad company. It was the logical endpoint of a market where everyone was buying the hype faster than they were measuring the results.

I’ve Seen This Movie Four Times Before

I’ve been in go-to-market for 21 years. Every “automation replaces humans” cycle follows the exact same script:

Phase 1: Massive hype. Impressive demos. Cost savings on a slide deck. Venture money floods in.

Phase 2: Bolt-on deployment. Companies attach the tool to broken processes and expect magic. Nobody redesigns the workflow.

Phase 3: Volume goes up, quality craters. Pipeline looks full on the dashboard but conversion rates quietly drop. Closers start complaining about lead quality.

Phase 4: Correction. The best companies redesign the workflow. Everyone else churns and blames the tool.

Outsourced call centers in 2004. Marketing automation in 2012. Chatbots in 2018. AI SDRs in 2024-2026.

The pattern is identical every single time. We’re somewhere between Phase 2 and Phase 3 right now.

The Numbers Don’t Lie — But the Vendors Cherry-Pick

Here’s what the head-to-head data actually shows when you strip away the marketing:

The AI SDR booked more meetings. The human SDR generated 2.6x more revenue.

Every AI SDR vendor shows you the volume slide. Nobody shows you the revenue-per-meeting slide. Because the story it tells isn’t the one they’re selling.

One company I’m advising ran a 90-day controlled test:

• AI-only pipeline: 847 meetings booked, 11% opportunity conversion

• AI + human hybrid pipeline: 312 meetings booked, 38% opportunity conversion

The hybrid generated 2.3x more revenue from fewer meetings. Read that again. Fewer meetings, more money. Because meeting quality compounds and meeting volume doesn’t.

The Email Deliverability Crisis Is Making It Worse

Here’s something most AI SDR vendors won’t tell you: the email infrastructure they depend on is actively working against them.

• 45-47% of global email traffic is now classified as spam

• 16.9% of commercial emails never reach their destination

• Gmail’s Gemini AI is specifically filtering AI-generated outreach

• Microsoft rolled out stricter enforcement in May 2025 (bounces above 2%, complaints above 0.3% = domain reputation damage)

Average B2B cold email reply rates dropped from 6.8% in 2023 to 4-5% in 2025. The 10x volume advantage of AI SDRs is running headlong into inbox providers who are 10x better at flagging automated messages.

You’re not just fighting for attention anymore. You’re fighting for deliverability. And an AI sending 1,000 templated emails a day from a burned domain isn’t a growth strategy — it’s a self-inflicted wound.

The $75K vs. $25K Math Is Wrong on Both Sides

The cost comparison that launched a thousand pitch decks is misleading in both directions.

The human SDR actually costs more than $75K. Fully loaded — salary, benefits, management overhead, tools, office space, recruiting costs given 30-40% annual attrition — you’re looking at $110K-$139K per rep.

The AI SDR actually costs more than $25K. When you add implementation ($5K-$15K), data enrichment subscriptions, email infrastructure (domain warming, deliverability tools), CRM integration work, and the human time required for daily tuning and supervision, real TCO runs $35K-$65K.

The gap is real. But it’s not the 3:1 ratio the vendors show. It’s more like 2:1. And when the AI generates 2.6x less revenue, the math flips entirely.

What the Winners Are Actually Doing

The companies getting this right — and they exist — aren’t having the “replace vs. keep” debate. They’ve moved past it.

SaaStr’s case study is the most documented example. They deployed 20+ AI agents, went from 8-9 human salespeople to 1.2 humans plus AI. Results over 8 months: $5M additional pipeline, $2.4M closed, deal volume doubled, win rate doubled.

But the details matter more than the headline. Jason Lemkin — who’s arguably the biggest AI SDR bull in the market — said this: “If you hook up an AI SDR and go away and do nothing, you will get nothing. Zilch. Nada.”

Their inbound AI agent alone needed 47 iterations to stop being too aggressive on pricing. They have a dedicated AI operations person doing daily tuning. It works because they treat AI SDR as an operations challenge, not a plug-and-play product.

45% of sales teams have already moved to hybrid models. The data on hybrid is consistently strong:

• 30% improvement in conversion rates

• 20% pipeline growth

• 60% cost reduction versus all-human teams

• 2.8x more pipeline from augmentation versus replacement

The pattern: AI handles research, personalization, timing, initial touchpoints, and follow-up sequencing. Humans handle judgment, relationship-building, complex objections, and the conversations that actually move deals.

The Three Mistakes Killing AI SDR Deployments

After watching this play out across 14 companies, three failure modes account for the vast majority of the churn:

1. Bolting AI onto broken processes

If your ICP is undefined, your messaging is generic, and your handoff to AEs is broken — AI won’t fix any of that. It’ll scale the dysfunction at 10x speed. “Garbage in, garbage out” is not a metaphor here. It’s a daily pipeline review.

2. Measuring volume instead of revenue

The company that books 847 AI meetings and celebrates is the company that churns in 6 months. The company that measures meeting-to-opportunity conversion, opportunity-to-close rate, and revenue per meeting actually understands whether the tool is working.

3. Expecting zero maintenance

Every successful AI SDR deployment I’ve seen requires daily tuning. Not weekly. Not monthly. Daily. Prompt adjustment, domain monitoring, reply analysis, false positive filtering. The “set it and forget it” promise is the biggest lie in the market.

What Happens Next

Gartner predicts that by 2028, AI agents will outnumber sellers 10:1. McKinsey says 70% of routine sales tasks will be automated by 2030. This technology isn’t going away.

But the standalone AI SDR product — the one that promises to replace your BDR team for $25K/year with no workflow changes — is going to consolidate, rebrand, or die.

What survives will look more like this:

The $250K SDR. The hybrid rep who uses AI for 60-80% of operational tasks and applies human judgment to the 20-40% that actually closes revenue. They’ll earn 2-3x what today’s SDRs make, and they’ll be worth 10x.

AI-native outbound workflows. Not bolt-on tools, but redesigned processes where AI and humans each do what they’re best at. The workflow is the product, not the agent.

Revenue-measured, not volume-measured. The metric shift from “meetings booked” to “revenue generated per outbound dollar” will kill most current AI SDR vendors’ value propositions.

The Bottom Line

The AI SDR isn’t a strategy. It’s a tool inside a strategy. And if you can’t articulate the difference, you’re about to become a churn statistic.

The companies winning right now share one thing: they redesigned their top-of-funnel workflow before they added AI to it. They didn’t automate the existing process. They built a new one where AI makes human reps 3x more effective instead of trying to make AI do a human’s job at 40% quality.

If you’re evaluating AI SDR tools right now, ask the vendor one question: “What’s the meeting-to-opportunity conversion rate for your average customer?”

If they pivot to volume metrics, you have your answer.

What’s your experience been — AI SDRs delivering, or disappointing? Hit reply or drop a comment. I’m collecting real stories for a follow-up piece on the specific workflow designs that are actually working.

Sources & Further Reading:

• UserGems: Are AI SDRs Really Worth It in 2026?

• TechCrunch: 11x Claiming Customers It Doesn’t Have

• SaaStr: 6 Months of AI SDRs — $1M+ in 90 Days

• Gartner: AI Agents Will Outnumber Sellers 10x by 2028

• Leads at Scale: AI Won’t Replace Your SDR Team — Here’s the Data

• Inc: AI’s Theranos Moment — The 11x Scandal

• McKinsey: Sales Automation — Boosting Revenue and Reducing Costs

• GrowthUnhinged: 2026 State of AI for B2B GTM Report

45,000 tech workers lost their jobs in Q1. Over 9,200 of those layoffs were directly attributed to AI and automation. Amazon alone accounts for 30,000 — flattening management layers, removing entire functions.

And if you listen to the narrative, it sounds clean. Efficient.…

Read more

Meanwhile, something far more consequential happened quietly in the background.

AI moved off the server and onto the desktop. Memory became persistent by…

Read more

JPMorgan Chase is using agents to produce investment banking decks in 30 seconds.

Thirty seconds.

I don’t care where you l…

Read more

This week was different.

Every major AI platform shipped autonomous agents to real users. Not demos. Not waitlists. Not “coming soon” blog posts with concept videos. Production software, in people’s hands, doing work without being told each …

Read more

Not a chatbot playground. Not a tech demo. A full-blown Reddit-style forum called Moltbook where AI agents autonomously post, debate, upvote, and organize into communities. They created 2,300+ sub-communities. They built governance structures. They policed content quality.

Humans? Welcome to…

Read more