Topline · 21 Jun 2026 · From the week of 15 June
The $1M Employee Is Here: Why ClickUp Replaced 22% Of Its Team With AI | Gaurav Agarwal, COO @ ClickUp
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Gaurav Agarwal, COO of ClickUp (a company above $300 million in ARR), joins hosts Sam Jacobs, AJ Bruno and Asad Zaman to explain ClickUp's decision to part ways with 22% of its workforce while rolling out pay packages of up to $1 million a year for individual contributors who stay. He describes a central Foundry team that builds AI tooling while each function's top performers automate their own workflows, with examples from marketing, data, creative and sales. The episode also covers talent bifurcation, meeting recording and candor, the collapse of adjacent roles, and whether value accrues to model companies or the application layer. Sponsor reads, a quiz segment and a bulls-and-bears debate are present but are not summarised here.
For founders
- Gaurav's view is that AI will do the job itself, with humans becoming managers and trainers of AI rather than using AI as a sidekick.
- ClickUp's model is a central Foundry team that builds the tooling, while two or three top performers per team build the agents that automate that team's own work.
- Gaurav said ClickUp chose entropy first, letting people experiment like pirates, and plans to bring in someone to structure the work later.
- Gaurav said ClickUp is still working out how to measure AI-driven outcomes, using the lowest-level leading KPI tied to the main revenue KPI as the test before rewarding people disproportionately.
- Gaurav said that if models commoditise, value shifts to the orchestration layer within the application layer, while Sam called reliance on a single model vendor a key-man risk.
For revenue leaders
- Gaurav said he does not think sales gets the same leverage from AI as engineering does, so ClickUp is still working out how to pay sellers who use AI.
- A two-person marketing team at ClickUp built a four-agent campaign workflow that Gaurav said delivers 70 to 100 campaigns a week, compared with five to 10 for a human.
- ClickUp's sales leader and ops person built a coaching bot that gives reps individual feedback on discovery calls, which Gaurav plans to extend across the deal cycle.
- Gaurav suggested removing meeting note takers from conversations about comp, individual feedback and customer-sensitive topics, because people speak less openly when recorded.
- Gaurav said that when a company says a salesperson produces 10 times their OTE, he finds it somewhat inefficient and asks why the company is not hiring more.
What was said 26, most useful first
A friend's analysis at a large bank found the smallest mortgage admin SOP had 13 steps, and the best model's accuracy at 15 steps was under 35%. Listen
A host relayed the experience of a friend who runs AI at a large bank and was asked to automate administrative divisions. The friend concluded that individual tasks can be automated but nothing at job-replacement level outside engineering. Gaurav responded that his experience at ClickUp was significantly different.
“the smallest SOP for anything was 13 tasks, 13 step SOP. And if you take the best model in the world, the most expensive, the best model, its accuracy at 15 steps is less than 35%.”
At ClickUp, a two-person marketing team built a four-agent campaign workflow that Gaurav said delivers 70 to 100 campaigns a week. Listen
Gaurav described separate agents for analysing data, writing emails, creating audiences and sending through Outreach, with the whole chain built by two people. He said a human would typically produce five to 10 campaigns a week. He said ClickUp is now looking at how to add email capacity through tools like Smartlead and Outreach.
“this entire workflow is what delivers about 70 to 100 campaigns a week. Typically, a human would have done about five to 10 campaigns a week.”
Listen to the episode Pipeline & demand generation Link to this Report a problem
Gaurav described ClickUp's Foundry model, where a systems team builds the tools and top performers in each function build their own automations. Listen
The Foundry team provides the building blocks, including a vibe-coding tool, a database, a place to build agents and agent observability. Two or three top performers per team then automate that team's workflows using those tools. Gaurav said his data team stopped doing analysis and now creates context in Hex for agents instead.
“So the harness is built by the Foundry team, as we call it. But the way harness is used to automate your team's workflow, that's been done by each of the team's top performers, like two people, three people per team”
For the layoff decision, ClickUp looked at who was leaning into AI and whether they could get there, because AI amplifies whatever skill a person already has. Listen
Gaurav said the decision was leadership-led rather than manager-led. It looked at who wanted to lean into AI versus who resisted, and whether each person could get there. He said weak writers using AI produce garbage emails. In engineering the top 20% gain large multiples, while a group within the bottom 80% uses AI to write more sloppy code and drive up token costs.
“It is an amplification of whatever you have.”
Listen to the episode Hiring & team building Link to this Report a problem
Sellers cannot 100x themselves with AI the way engineers or artists can, with a seller possibly reaching 2x to 5x. Listen
He compared AI leverage to outliers like Taylor Swift and Drake, and said good engineers can 100x themselves. He said a seller cannot go from a million-dollar quota to a hundred-million-dollar quota, and that they might manage 2x, 3x or 5x. He described the leverage as jagged and bifurcated across roles.
“there's not too many jobs, like a seller can cannot 100x themselves. You can't go from getting a million dollar quota to 100 million dollar quota.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
ClickUp's marketing creative team uses AI to reverse-engineer competitor ads and generate variants, starting with static ads and moving to UGC and video. Listen
He said the system looks at competitor and top B2C ads, identifies the psychological reasons they work, applies those to ClickUp's ads, creates variants and sends them to Facebook. He said the creative team produces about 600 assets a week, and that the same approach is planned for UGC and then video using Runway.
“then take those ads and then apply it to ClickUp's ads and then create variants and then pump straight into Facebook.”
Listen to the episode Positioning & marketing Link to this Report a problem
ClickUp is beginning to use AI for capital allocation by having one agent recommend, another counter, and a third summarise for a human decision. Listen
He described a pipeline in which the third agent learns from whether the human accepts or rejects each recommendation and why. He said the approach is used for direct media capital allocation, and that it is still early.
“If you build an agent that makes a recommendation, then you build another agent that counters the recommendation.”
Gaurav suggested removing meeting note takers from conversations about comp, individual feedback and customer-sensitive topics because people speak less openly when they are recorded. Listen
He said he has had people ask him to turn off note takers so they can speak openly, particularly in sales. He said most meetings are not affected because people have become blind to note takers, but that tricky human, comp and customer conversations should not be recorded.
“those conversations should not be recorded because those are tricky human related comp related customer related conversations”
Listen to the episode Leadership & culture Link to this Report a problem
As AI collapses roles, the winner is the specialist with several spikes, and that the one with the best taste and drive takes over adjacent departments. Listen
He said generalists will not take over, but specialists with more than two or three spikes will absorb those spikes. He used the example of a growth person who also understands data, sales ops, sales and success. He said which spike eats which is yet to be determined.
“the one with the best taste and the drive to work and learn and improve, eats up adjacent departments.”
Listen to the episode Hiring & team building Link to this Report a problem
AI will do the job itself, and humans will become managers and trainers of AI. Listen
Gaurav said his mindset moved from using AI as a sidekick to keep existing structures to AI doing the work. He argued that humans will build AI to do the job, and that AI will do it better than an 80th percentile human. He described the resulting human role as managing and training AI.
“AI will do the job better than an 80th percentile human. And then our jobs become managers and trainers of AI.”
ClickUp tries to limit each agent to three to five tasks. Listen
He described breaking a workflow into smaller processes, with each agent handling a narrow job, rather than building one agent to do everything. He presented this as the current practice at ClickUp, which he said is still learning as it goes.
“we try to limit an agent to three to five things that they should try to do, not more than that.”
In a zero-to-one AI push he wants entropy first, with bottom-up experimentation, and structure added later. Listen
Six months into ClickUp's push, with the last three more aggressive, Gaurav said the company does not yet know what it is doing. He preferred a pirate-ship phase of experimentation before bringing in someone to structure the work like a naval fleet. He acknowledged that this approach has no structure, but said he thought top-down direction alone would not work for this kind of change.
“what I need right now is I need entropy... let's go be a little bit like a pirate ship.”
Listen to the episode Leadership & culture Link to this Report a problem
ClickUp is redesigning compensation to pay more for employees who use AI to build agents that do work. Listen
He said ClickUp is rolling out new bands that reward people who build digital workers, and that as one person automates their job and then adjacent jobs, a few people may end up orchestrating agents that do the work of many. He said this is still being figured out, and that engineering is easy to justify while sales is less clear.
“we want our top employees who are using AI to build digital workers or like build other agents that do the job. They should be paid higher”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
He does not think sales gets the same leverage from AI that engineering does. Listen
He said that while engineering is an easy case for higher pay, sales could be easy to justify but he does not think sellers get the same AI leverage. He said ClickUp is still working this out as part of its compensation changes.
“I don't think sales gets the same leverage out of AI the way engineering does.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
ClickUp's compensation framework starts from the lowest-level leading KPI that is directly tied to the main revenue KPI. Listen
He said the company is trying to measure whether a person who can do the work of five people creates incremental business value. For each department it is looking for a leading KPI that connects to revenue, and if AI moves that KPI, the people who made it possible would be rewarded disproportionately. He said this framework is still being built.
“what's the lowest level leading KPI that we know is directly tied to the main revenue KPI.”
Listen to the episode Metrics & finance Link to this Report a problem
He finds it inefficient when a company says a salesperson produces 10 times their OTE in revenue. Listen
He said he does not think it is a positive sign, and that he asks why the company is not hiring more. Another participant agreed that this leaves money on the table.
“I actually don't even think it's a really positive sign nowadays when companies say i have a salesperson who does like. 10 times their OTE in revenue, I find that somewhat inefficient.”
Listen to the episode Sales team, hiring & comp Link to this Report a problem
ClickUp's sales leader and ops person built a bot that coaches reps individually after discovery calls, and Gaurav plans to extend it across the deal cycle. Listen
Gaurav said the bot gives each rep personalised feedback on what they could have done better in a discovery meeting, which he likened to having a coach after every call. He plans to extend it across the deal cycle. He illustrated the idea with a hypothetical note to Sam about failing to push on pain.
“they built a coaching bot that coaches reps individually on a discovery meeting and what you could have done better.”
Listen to the episode Sales process & deals Link to this Report a problem
AI error tolerance depends on the use case: one or two bad ads out of 100 is acceptable, but a 5% error rate on claims would put an insurer out of business. Listen
He said ClickUp's marketing can tolerate a small share of weak output because it is one of many options. He contrasted this with payroll, mortgages and insurance claims, where errors are not acceptable. He said this jaggedness creates confusion in the market about whether AI works.
“If you think about marketing as a use case, let's say out of those 100 ads that it gives you, one or two are really dumb. That doesn't matter as much”
Working in public in shared channels gives AI full context and avoids repeated pinging between people. Listen
He said ClickUp's rule is that work discussions should happen where all stakeholders can see them, rather than through one-to-one messages that get relayed back and forth. He said this helps AI have full context, and that it is less important for private conversations.
“it's better to work in public so all your stakeholders can have the context versus me pinging you, you pinging Sam, Sam pinging Ajay, then Ajay pinging me back about the same thing.”
Listen to the episode Leadership & culture Link to this Report a problem
He has not seen anyone on his team leave to become a mercenary rather than a missionary, and expects AI to create more mercenaries. Listen
He said he is not seeing as much of a divide as the question suggested. He expects more mercenaries to emerge as jobs are reduced. He expects those mercenaries to democratise excellence by delivering previously frontier capabilities, such as agentic campaign workflows, to companies of all sizes, similar to how Mailchimp and Squarespace spread earlier tech.
“I have no single person on my team leave because they want to become a mercenary versus a missionary.”
Listen to the episode Leadership & culture Link to this Report a problem
Gaurav predicted that top operators will increasingly sell outcome-based automation independently while leverage is high, with prices falling later as in software. Listen
He said some people will choose flexibility and pay from outcome-based work, while others will choose career growth and equity inside companies. He expects prices for these services to come down over time, as happened in software, but said the current phase is a great time for this model.
“Eventually, you'll have many people saying I can automate your workflow for accounting, for example, that the price would come down again.”
Listen to the episode Strategy & market Link to this Report a problem
Value will accrue more to the application layer than to the model layer if models commoditise. Listen
He compared models to electricity that is priced by time and access, with orchestration in the application layer doing computation to reduce cost and deliver value. He said the exception is if one model pulls far ahead, which he expects to find out next year.
“So I do think value goes to the application layer significantly more than the model layer.”
Listen to the episode Strategy & market Link to this Report a problem
Relying on one model vendor creates key-man risk. Listen
He cited Anthropic having to pull Fable, calling it Mythos, and the emotional maturity of leaders like Dario and Sam Altman as reasons for concern. He said doing everything with Claude makes him nervous, and that specialisation gives a comparative advantage, so he expects better applications from focused vendors.
“feels like a lot of key man risk for one specific vendor.”
Listen to the episode Strategy & market Link to this Report a problem
A company's value proposition should not change daily, even as features change often. Listen
He argued that companies release functionality rather than new platforms every day, and that features do not fundamentally shift the value proposition. He said a company should keep a cohesive message while adding capabilities, and only occasionally add a new arm that changes positioning.
“you should not be adjusting your value proposition on a daily basis.”
Listen to the episode Positioning & marketing Link to this Report a problem
Gaurav predicted that the companies that own the user relationship will win in the near to mid term, because they can build feedback loops to train models. Listen
He said that at some point there will be limited training data available and models will peak, so whoever owns the user can use their interactions to improve models. He described this as the near-term winner, with a more uncertain outcome if AGI arrives.
“whoever owns the user will be able to build feedback loops to train their models better.”
Listen to the episode Strategy & market Link to this Report a problem
He thinks AI education is one of the easiest things to learn, and that he sees a lot of resistance and inertia. Listen
Responding to a host who described employees asking for more AI enablement, Gaurav said he has had similar requests. He said more enablement can be done, but AI is one of the easiest things to learn: teenagers learn it from YouTube, or people can just ask the LLM. He said there is a lot of resistance and inertia and people are struggling.
“AI education is one of the easiest to learn. There are teenagers learning AI from just watching YouTube videos.”
Listen to the episode Hiring & team building Link to this Report a problem