6/2/26: Inside Perplexity's Revops, 3 AI Skills Replacing Admins
Another week in GTM AI land and today we have a killer deep dive into Perplexity with their Head of Enteprise Ops and Systems, a good friend Nate Follen who is one of the best operators and Revops leaders in the space. I have had the privilege of seeing his work up front when he was at Ramp and now seeing what he is doing, is mindblowing.
As usual, we have lots of goodies, please make sure to read everything as my goal is to give as much value as possible. We have also most recently uploaded a TON of amazing content on the paid side, go check it out:
Let’s get into the podcast!
You can go to Youtube, Apple, Spotify as well as a whole other host of locations to hear the podcast or see the video interview.
The RevOps leader who stopped hiring
“Every time I think I need to hire someone, I just solve it with AI instead.”
Nathan Follen said that to me this week, and I have been chewing on it ever since. Nathan leads go-to-market systems and ops for Perplexity Enterprise. Before that, he was a big reason RAMP scaled like a rocket. So when he tells me he is running RevOps with a team of agents instead of a team of people, I do not roll my eyes. I take notes.
He showed me 3 skills he built in the last 2 months. Work that used to eat his team hours every week. Now it runs by itself, every single day, whether his laptop is open or sitting in a drawer.
Here is what he is actually doing.
1. Voice of Customer that runs itself
Most teams treat voice of customer like a quarterly project. Nathan turned it into a living dashboard that refreshes daily.
The build was almost embarrassingly simple. Two prompts and an API key into his call recording tool, Momentum.io synced with Salesforce. From there the agent does the work a data team used to:
Auto-tags discovery calls vs follow-ups from the context of the conversation alone. No admin. No manual fields.
Surfaces the week’s themes: pricing objections, feature requests, rollout blockers, champion quotes.
Tells him what the product team should do about each theme, and what enablement should build.
Then it did something I did not expect. It pulled the top 20 customer quotes, grabbed the clips using the new Momentum.io SmartClips product, and edited a 2-minute customer sizzle reel with music. Inside the same tool. The thing marketing usually waits two weeks for.
Why it matters: Your customers are telling you exactly what to build and how to sell it. The bottleneck was never the data. It was the hours to process it. That bottleneck is gone.
2. The weekly deck that builds itself
Nathan runs a weekly RevOps go-to-market call. The deck for it used to take an hour to build by hand.
Now a skill kicks off every Thursday on a schedule and does the whole thing:
Pulls live numbers from Snowflake and project updates from key Slack channels.
Pings the sales team to clean up stale pipeline before the data gets pulled.
Sets a cron job to re-check that Slack thread, grab whatever people added, and drop it into the final slides.
One hour of work, every week, reduced to a notification that the deck is ready. The agent even decides which buried Slack updates the whole team needs to see.
Why it matters: Reporting is the tax RevOps pays to do its real job. Nathan stopped paying it.
3. A CRM that cleans itself at night
This is the unglamorous one that quietly runs everything.
Account ownership, Salesforce hierarchies, mismatched domains and contacts. All of it gets cleaned on a nightly batch job instead of a pile of real-time flows that break the moment you scale. When Computer catches something it should fix, it runs the cleanup, sometimes directly and sometimes through a tool like Polytomic or Hightouch.
Then the part that earns trust: a monitoring agent reads every Slack channel and his inbox, and DMs him each error, ranked by severity, with the fix already attached.
Why it matters: Nathan said the speed unlock was not the building. It was the confidence to build fast because he knows the system will catch its own mistakes.
The shift underneath all three
Here is the line that got me out of my chair. Nathan runs a daily standup with his agents. He asks them what they worked on and what they could do better.
The old RevOps job was to find the top two priorities and protect focus. The new job is to broaden the aperture and run dozens of projects in parallel, because a v1 ships in the time you used to spend arguing about whether something was worth doing.
He put it perfectly. If you hire two people to do the same thing with agents, you get a mess. If you give people their own surface area and let each one manage a team of 10, 20, 100 agents, you hire them on the spot.
People keep calling this the future. Nathan is doing it on a Tuesday.
Nate Follen: Pull Quotes & Learnings
On hiring vs building
“[Every time I think I’m going to hire someone, I just solve it with Perplexity.]”
“In some cases, you don’t need expensive software to do fit-for-purpose things. Building something out internally with Computer or other agentic tools is the right solution.”
“If you hire two people and tell them to do the same thing with agents, it’s going to be a mess. But if both have different responsibility areas, and each can manage a team of 10, 20, 100 agents, then absolutely hire that person right away. Especially if they’re curious and willing to manage an agent team.”
On confidence and decisions
“Things like territory carving and total addressable market used to take a headcount or two to really nail. Now I feel really confident in those decisions using AI.”
“The accuracy in search, and then the memory, makes it much more accurate.”
On the orchestration layer
“We’re the orchestration layer across 400 different tools. If there’s a very good tool for something, we want it to be agentic so we can make modifications and monitor it. We leverage the best of what each tool is built for, and orchestrate between them.”
“The connectors are a huge game changer, and it’s a flywheel. The better the tools get and the more accessible they are through agents, the better for us.”
On Voice of Customer
“Instead of needing a Salesforce admin or a data team to analyze whether something’s a first call, it parses that based on the context of the call itself.”
“The ‘what do we do about it’ has been the game changer. Sometimes viewing a dashboard isn’t really actionable.”
On the mindset shift
“The job of RevOps used to be: find the top two things to work on, stay focused, get them done. With agents, you can broaden that aperture.”
“In the time you tried to push back on a project, it could have been a version 1, completed, to see if it works.”
“We can run a lot more projects in parallel and test a lot more things, especially on the marketing side, with less resources.”
On managing agents like a team
“I asked all my skills and agents to do a daily standup: what did you work on, and what should you do better?”
“Every week I ask: where have projects stalled, and where should I focus that would have the biggest impact?”
On speed and accuracy
“A couple of minor inaccuracies about someone’s company or role can kill a deal early.”
“The monitoring is automated. That confidence, that we can change things quickly and we’ll catch the errors, makes it a lot easier to build fast.”
“It’s solving for organizational change and making that painless.”
“These last couple of months have been unbelievable. It’s a different world than it was a year ago.”
How Nate Thinks (the learnings)
1. Orchestrate, don’t replace. He doesn’t rip out his stack. He sits an agentic layer on top of 400 tools, uses each for what it’s best at, and orchestrates between them. The agent is the conductor, the tools are the orchestra.
2. “What should we do about it” beats any dashboard. A dashboard reports. Nate’s agents recommend: what the product team should build, what enablement should fix, what marketing should say. Insight without a next action is just decoration.
3. The buy-vs-build line moves every month. His test isn’t features. It’s “what works and what we can maintain.” Expensive software earns its place by depth and support. Everything else is a candidate to build fit-for-purpose. Nate still went out and grabbed Momentum.io because he knows the limits of what AI can or cannot do.
4. Accuracy plus memory is the moat. Tell the system once that “sellers” means the AE team, and every future answer gets sharper. Context compounds. Bad data breaks every agent downstream, which is why CRM hygiene runs nightly.
5. Broaden the aperture. Old RevOps protected focus by killing projects. New RevOps ships v1s. A version 1 is cheaper than the meeting where you debate whether to build it.
6. Manage agents like a team, with rituals. Daily standup with his agents. Weekly project triage. Zero-lead-leakage checks. The work shifts from doing the task to running the team that does it.
7. Hire for new surface area, not duplicate work. Two people pointed at the same agent-driven task = chaos. One curious operator running 100 agents on their own surface area = leverage. Hire the agent-managers.
8. Monitoring is what unlocks speed. The build wasn’t the hard part. The confidence to build fast came from an agent that DMs him every error, ranked by severity, with the fix attached. Safety nets make speed possible.
9. Most friction is internal. The thing slowing the team down usually isn’t the customer. It’s the internal process. Make organizational change painless and the team moves at a speed that looks unfair.
The Perplexity Computer RevOps Playbook
Build the 5 agentic workflows that let a Perplexity RevOps leader stop hiring and start orchestrating. Setup, exact prompts, and a 7-day rollout. Steal all of it.
Part 1: What Perplexity Computer actually is
Perplexity Computer is what they call a general-purpose digital worker. You give it a goal in plain English. It figures out the steps, picks the right tools, does the work, and hands you the finished thing. Most AI tools give you a summary or a plan you then go build yourself. Computer delivers the actual artifact: a built dashboard with a shareable link, a cleaned dataset with charts, a research report with citations, a finished deck.
Four things make it different from a normal chatbot, and all four matter for the workflows below.
It runs on many models, not one. Computer sits on top of 19+ frontier models and routes each piece of your task to the model best suited for it. Claude Opus handles the heavy reasoning and orchestration. Gemini runs deep research. Others handle long-context recall, images, and video. You write one prompt. It assembles the team. When one model is down or weak at something, it does not take you down with it.
It works in Tasks, not chats. A Task is a job, not a conversation. When you submit a prompt, an orchestrator breaks the objective into subtasks, assigns each to the right model, runs them in parallel, and compiles the result. You can start a Task, immediately start another, close your browser, and come back to both finished.
It remembers. Computer keeps context across sessions and learns your preferences and standard workflows. Tell it once that “sellers” means your AE team and “CS” means these five people, and every future question gets more accurate. Memory is also why continuing a Task is cheaper than starting a fresh one.
It runs on a schedule, asynchronously. Set a Task to run once, daily, weekly, monthly, or yearly at a trigger time. It runs in the cloud whether your machine is on or off. Each run saves a thread you can go back to. This is the feature that turns a one-time task into an employee who shows up every morning.
Pricing, so you can plan. Pro is $20/month with 4,000 credits. Max is $200/month with 10,000 credits. Enterprise Max is around $325 per seat per month with admin controls, custom connectors, and security and governance. Start on Pro to learn it. Move up when a workflow proves its value.
Part 2: The 30-minute setup that makes everything else work
Do not skip this. The people who say Computer is “just okay” almost always skipped the setup and then wondered why it felt generic. Three steps.
Step 1: Connect your tools
Connectors give Computer real read-and-write access to your actual data, not a summary of it. There are 400+ built in. The ones that matter for GTM:
Salesforce and HubSpot for CRM data and actions
Snowflake, BigQuery, or Databricks for your warehouse
Slack for team comms and notifications
Gmail and Google Calendar for inbox and meetings
Your call recording tool (Gong, Momentum, Fireflies) by connector or API key
Linear, Jira, or Asana for project status
Click Connectors in the sidebar, find the app, click Enable, complete the OAuth login. About a minute each.
If a tool you use is not in the list, you can still connect it. Provide an MCP (Model Context Protocol) server URL and Computer talks to your proprietary CRM, custom analytics server, or private API. This is exactly how Nathan connected his call recording tool: no native connector, just an API key and a shared skill that told Computer how to use it. Enterprise admins can share custom connectors across the whole org.
Step 2: Create your first Skills
A Skill is a saved set of instructions that auto-activates when Computer recognizes a matching task. Think of it as a job description you write once and never repeat.
Without Skills, you re-explain your brand, your formatting, and your reporting structure on every single task. With Skills, you explain it once and it sticks. Computer ships with built-in Skills for Slides, Research, Research Report, and Chart. To make your own: click Skills in the sidebar, click Create skill, and upload a .md file.
Skills stack. A research Skill can hand off to a report-formatting Skill, which hands off to a slides Skill. One prompt runs the whole pipeline.
My rule, stolen from Nathan: if you have explained the same thing to Computer twice, it should be a Skill.
Step 3: Set Custom Instructions
Skills fire for specific task types. Custom Instructions apply to every task, all the time. Keep them under 1,500 characters. The single most valuable one I have found, and the one that will save you the most money:
Always come back to me and clarify any misunderstandings, challenge my thinking to make sure you are clear on the stated outcome, and create a brief plan before you build anything.
That one line forces Computer to confirm what you actually want before it spends credits running the wrong job. Add your context too: who your team is, what “done” looks like, your tone, your no-go zones.
Part 3: How to not waste money (the credit economics)
Three habits separate people who love Computer from people who churn.
Continue Tasks, do not restart them. Continuing uses persistent memory and is always cheaper. Start a new Task only when the objective genuinely changes.
Control model routing on big jobs. By default Computer often reaches for the most capable, most expensive model. For routine work, tell it which model to use. Save the heavy reasoning models for the hard parts.
Make it plan before it builds. The custom instruction above is not just for quality. A clarifying question costs almost nothing. A wrong 2,000-credit Task costs a lot.
Treat credits like a budget and these workflows pay for themselves in the first week.
Part 4: The 5 workflows (copy the prompts)
Each one below has the same shape: what it does, why it matters, the exact prompt to build it, and how to schedule or share it. Swap in your tool names where I use brackets.
Workflow 1: Voice of Customer engine
What it does: Reads every sales call, tags the call type, surfaces the week’s themes, tells you what to do about each one, and finds your best champion quotes. Refreshes daily.
Why it matters: Your customers are handing you your roadmap and your messaging on every call. The only thing that ever stopped you from using it was the hours. Computer removes the hours.
Build prompt:
You have access to my [call recording tool] via the connected API key and to Salesforce. Build me a Voice of Customer dashboard that refreshes daily.
Steps:
Pull all call transcripts from the last 7 days. Use Salesforce to enrich each call with account name, deal stage, and the title of who we met with.
Auto-tag each call as a first/discovery call or a follow-up based on the content of the conversation, not a CRM field.
Group what you find into these sections: Pricing and packaging, Feature requests, Objections and risks, Positive signals and champions, Rollout and use cases.
For each theme, write: (a) the key takeaway in one line, (b) 2 to 3 exact customer quotes with the account name, (c) what the product team should do about it, (d) what enablement or marketing should build.
Output as a clean dashboard with a shareable link. Remember that “sellers” means my AE team and “CS” means my customer success team for all future questions.
Then turn it into a daily Task: “Run this every weekday at 7am and post the summary to our #voice-of-customer Slack channel.”
Bonus, the move that wowed me: Add a second prompt. “From this week’s calls, pick the 20 strongest customer quotes, pull the video clips, and edit a 2-minute sizzle reel with background music.” Computer has video models built in, so it can actually produce the reel.
Workflow 2: The weekly deck that builds itself
What it does: Builds your recurring meeting deck end to end. Pulls live data, chases stale inputs, and assembles the slides on a schedule.
Why it matters: Reporting is the tax your ops team pays to do its real job. This stops the bleeding of an hour or more every week, per recurring meeting.
Build prompt:
Build me a weekly RevOps go-to-market deck and save it as a reusable Skill called “Weekly Deck Prep.”
Each run should:
Pull current pipeline and revenue numbers from [Snowflake/Salesforce]. Include month-over-month and week-over-week trends.
Scan these Slack channels for project updates: [#channel-1, #channel-2]. Decide which updates the whole team needs to see and summarize them by project.
Pull open and recently closed items from [Linear/Jira] for a quick burndown.
Post a message to #sales-team asking reps to update any stale in-month pipeline and to reply in thread with anything they want included in the deck.
Set a cron job to re-check that thread 24 hours later and fold their replies into the final deck.
Build the deck in our format: title slide, pipeline summary, trends, project status, risks, asks. Output slides with a shareable link.
Schedule it: “Run Weekly Deck Prep every Thursday at 8am.”
Workflow 3: Nightly CRM hygiene plus an error-watcher
What it does: Cleans your CRM on a nightly batch instead of brittle real-time automations, and DMs you any errors with severity and a fix.
Why it matters: Bad data quietly breaks every other agent you build. And the real unlock Nathan named was confidence: he builds fast because he trusts the system to catch its own mistakes.
Build prompt (hygiene):
You have access to Salesforce. Every night, run a CRM hygiene job and report what you changed.
Check account ownership against our rules of engagement: [paste your ROE in plain English]. Reassign accounts to the correct owner based on logged activity, so the rep actually emailing an account owns it.
Validate account hierarchies and flag or fix mismatched domains and contacts.
For anything you cannot safely auto-fix, list it for my review with the recommended action.
Prefer a clean batch approach over real-time triggers. If we need data moved between systems, use [Polytomic/Hightouch] and tell me what you set up.
Build prompt (the watcher, this is the trust layer):
Create a monitoring Skill. Every morning, scan my key Slack channels and my inbox for errors, failed syncs, and broken automations. DM me a single summary ranked by severity (critical, high, low). For each issue, include what broke, where, and the exact steps to fix it.
Schedule both: hygiene nightly, watcher each morning before you log on.
Workflow 4: Pre-call prep that flips discovery
What it does: Before every meeting, sends you a one-pager on who you are meeting, why they care, and what to show them.
Why it matters: A couple of small inaccuracies about someone’s role or company can kill a deal early. Accurate prep, done for you, means you walk in already halfway through discovery.
Build prompt (save as a Skill called “Pre-Call Prep”):
Each morning, scan my Google Calendar for today’s external meetings. For each one, use Salesforce, the web, and LinkedIn to build a one-page brief and DM it to me as a PDF.
Include: who I am meeting and their role, their company and what they do, recent news or signals, the deal context from Salesforce, the 2 to 3 use cases that land best for someone in their role, what to make sure I mention, and one genuine personal connection point if you can find one.
Schedule it: “Run Pre-Call Prep every weekday at 7:30am.”
Workflow 5: The daily agent standup and weekly project triage
What it does: Your agents report to you. Each day they tell you what they did and where they are stuck. Each week they tell you where projects stalled and what to focus on for the biggest impact.
Why it matters: This is the actual mindset shift. You are not doing the work anymore. You are managing a team that happens to be made of agents. You need a standup just like you would with people.
Build prompt (daily):
Every morning, have my active Skills and scheduled Tasks report a standup: what each one worked on yesterday, what it completed, what it is stuck on, and one thing it could do better. Then read my Slack and email and suggest where I should add a new agent or Skill to remove a bottleneck.
Build prompt (weekly):
Every Friday, review my core Slack channels and project tools and tell me: where have projects stalled, where is work getting dropped, and what are the 3 things I could focus on next week that would have the biggest impact on revenue and the team. Also run a zero-lead-leakage check: which leads or deals need follow-up and have gone quiet.
Schedule both. This is the closest thing to a chief of staff you can buy for the price of a Pro seat.
Part 5: Turn any workflow into a shareable Skill
The reason this compounds is that good workflows get shared. A Skill is just a Markdown file. Here is a template you can paste, edit, and upload under Skills > Create skill.
# Skill: [Name]
## When to use
[Describe the trigger. Example: "When I ask for a weekly RevOps deck
or it is Thursday morning."]
## Inputs and tools
- Salesforce (accounts, opportunities, activities)
- Snowflake (revenue tables)
- Slack channels: #channel-1, #channel-2
## Steps
1. [First step, be specific]
2. [Second step]
3. [Output format and where to deliver it]
## Rules and definitions
- "Sellers" = the AE team
- "CS" = customer success
- Always confirm the plan before building
- Output format: [dashboard / PDF / slides / Slack message]
## Done looks like
[One or two sentences describing a great result.]
Write it once. Share it with your team. Now everyone’s “rockstar rep workflow” is everyone’s workflow.
Part 6: Your 7-day rollout
You do not boil the ocean. You build one thing a day.
Day 1: Set up Pro. Connect Salesforce, Slack, your warehouse, and your call tool. Write your Custom Instructions (use the clarify-and-plan line).
Day 2: Build Workflow 4 (Pre-Call Prep). It is the easiest win and you will feel it in tomorrow’s meetings.
Day 3: Build Workflow 1 (Voice of Customer). Let it run once, then refine the sections.
Day 4: Build Workflow 3’s watcher (the error monitor). Trust comes before automation.
Day 5: Build Workflow 2 (Weekly Deck). Point it at your real meeting.
Day 6: Build Workflow 5 (the standup). Meet your team of agents.
Day 7: Turn your two best builds into shared Skills using the template above. Send them to one teammate.
Seven days. Five workflows. One genuinely different way of working.
My challenge to you
You do not have to believe agents will run RevOps. You just have to test it once. Pick the single workflow above that would save you the most time this week and build a v1 today. Not a perfect version. A v1. You will learn more in one hour of building than in a month of reading takes about AI.
The companies that learn to orchestrate this year will move at a speed that looks unfair to everyone else. The ones that wait will spend next year trying to catch up to a team a tenth their size.
So here is my challenge to you: build one. This week. Then come tell me what you made.
I hope this saves you the hours it saved Nathan. Go build something.
mk? mk.
Coach