When Your Best Hire Asks for a Compute Budget
Soon the best people you try to hire won't ask about headcount. They'll ask: "What's my monthly compute budget?" That's the line that reframed how I think about the next few years of work — and it comes from a recent interview with Dmitry Shevelenko, Chief Business Officer at Perplexity.
The argument underneath it is simple and a little vertiginous: if a single curious person can orchestrate enough AI to do the work of a department, then the unit of an organization stops being the employee and starts being the employee plus their compute. Budgets shift from software seats to something closer to AI payroll. These are my notes from the interview, plus where my own thinking landed after.
A quick note on sourcing: the figures and case studies below are recounted by Shevelenko in the interview and I haven't independently verified them. Treat them as illustrative claims from someone with a clear seat at Perplexity, not audited results.
Every employee becomes an executive
The core reframe is to stop treating AI as software and start treating agents as digital workers. Once you do, the job of a human changes shape. You're no longer the person doing the data-gathering and the execution — you're the person who sets up the workflow, kicks off the agent, and validates the output. You become, functionally, a manager of a small AI staff.
Shevelenko's framing is that everyone becomes an executive, and gets judged the way executives already are: on the ROI of the resources they deploy. A VP is measured on how well they spend a headcount budget. Tomorrow's knowledge worker gets measured on how well they spend a compute budget.
The unit of work is shifting from "what can you do?" to "what can you orchestrate?"The leverage numbers he cites are the part that makes it concrete. Perplexity reportedly runs its month-end financial close with one human overseeing the AI, and serves 50,000 enterprise customers with a go-to-market team of seven. Whether or not those exact figures hold, the direction is the thing: tiny teams doing what used to take an org chart.
This rhymes hard with something I wrote earlier — that you should stop treating AI as a tool and start onboarding it like a digital employee. The compute-budget framing is the same idea seen from the finance side: if it's labor, it belongs on a payroll line, not a software-license line.
The job becomes orchestration, and prompting is management
If execution moves to the agents, what's left for the human? Shevelenko's answer is error validation and taste — and I think he's right.
Agents still hallucinate and still make logic errors. So the durable human skill becomes the ability to stress-test output: to build heuristics for where a model is likely to be wrong, and to catch it before it ships. That's not a technical skill so much as a judgment skill.
The second half is feedback. When an agent fails, it's usually a failure of management — the same way a junior employee who delivers the wrong thing was often briefed badly. Learning to give concise, actionable correction is the new core competency.
My takeaway: "Prompting" is a bad word for this because it sounds like a syntax trick. It's really delegation. The people who are already good at handing work to a junior — clear brief, tight feedback loop, knowing what "done" looks like — have a transferable superpower. The people who hoard execution because "it's faster if I just do it myself" are going to struggle.
AI collaboration is multiplayer, not a solo chat window
One of the sharper product observations: AI shouldn't force humans into isolated, one-on-one chat windows. It's most useful where teams already work — Slack, Microsoft Teams, shared channels.
The reason is social, not technical. When AI lives in a shared channel, people watch how others prompt it. They learn each other's workflows, jump into a thread to ask a follow-up, and hand a half-finished AI task off to a teammate. The skill of managing agents spreads by osmosis instead of by training doc.
This is also Perplexity's answer to an adoption problem I'll get to below: the fastest way to teach a hesitant employee to delegate to AI is to let them see their boss do it in public.
Agents will delegate to sub-agents
The architecture Shevelenko describes for harder tasks isn't one big model — it's a planning model acting as a manager that breaks a complex prompt into steps and farms them out to specialized sub-agents. An orchestra, not a soloist.
What makes this matter for work is the trust horizon. Right now we hand agents tasks measured in minutes, maybe a day. As context windows and reasoning improve, that horizon stretches: a task you'd trust an agent to run for a week, and eventually a project it runs for a quarter before it needs you. The length of leash you're willing to give is the real measure of how capable these systems have become.
If you want the far end of that curve, it's the same dynamic Anthropic describes when AI starts writing most of its own code — the human moves up the ladder from doing, to reviewing, to setting direction.
Curiosity and agency become the moat
Here's the conclusion that I keep turning over. When AI can perfectly execute a task — write the SQL, pull the data, draft the report — then execution stops being a competitive advantage. It's a commodity.
So what's scarce? Knowing what to do. The bottleneck for the future knowledge worker isn't capability, it's the idea — the curiosity to ask the right question, the initiative to start the right project. Agency becomes the rarest asset a person has.
When execution is free, the entire game is deciding what's worth executing.My takeaway: This is oddly hopeful and a little brutal at the same time. Hopeful because curiosity is human and can't be outsourced. Brutal because there's no longer anywhere to hide — you can't be "the person who's really good at the spreadsheet" when the spreadsheet builds itself. The people who were always asking "why are we even doing it this way?" just became the most valuable people in the building.
What Perplexity learned by watching users
The second half of the interview is less manifesto and more field notes — how Perplexity actually observes people using AI, and how that changes what they build. Four observations stuck with me.
Executives are the best early adopters
After rolling out their internal "Computer" agent broadly, Perplexity's most active users were CEOs, business owners, and senior executives — reportedly 92% of them became weekly actives. The reason is exactly the orchestration point above: executives are already trained to delegate. Handing work to an agent is the same muscle as handing work to a team. Regular knowledge workers struggled more, because delegation isn't yet a habit for them.
That's the real reason for the Slack and Teams play: meet the hesitant adopters where they are and let them learn delegation by watching the people who are already fluent in it.
Agentic commerce is overhyped
Perplexity found users love using AI to research a high-value purchase — but they don't want the AI to click "buy." People are loss-averse about their own money; they want to manually confirm the final transaction so nothing gets messed up (wrong flight, wrong address). And modern checkout — Apple Pay and friends — is already nearly frictionless, so there's little pain to remove at the final step.
So Perplexity stepped back from one-click agentic checkout and leaned all the way into being the best tool for the research phase, leaving the execution click to the human.
My takeaway: This is a clean example of resisting a hype cycle by actually watching behavior. "Agent buys your stuff" demos beautifully and dies in production, because the friction was never at checkout — it was in the research. Build for the painful part, not the impressive one. It's the same lesson as "sell the shovels, charge for the service": solve the boring, valuable problem.
Advertising destroys trust in AI answers
Perplexity's read is that the moment an AI product carries ads, users assume the answers are corrupted — biased toward the highest bidder — even if the organic results are untouched. Because they see themselves as being in the trust and accuracy business, they've refused an ad model in the core experience. The judgment is that preserving trust is worth more than the ad revenue.
Server-side agents hit a "context wall"
Cloud-run agents keep failing at real web tasks: they hit paywalls, get blocked by captchas, and can't access a user's logins, cookies, and local state. Perplexity's fix was Comet, an agentic browser that runs on the client side so the agent inherits the user's actual session — cookies, logged-in state, preferences. The handoff is designed to be invisible: you ask "Computer" a question, and if it needs live, authenticated web access, it quietly hands off to Comet in the background instead of opening fifty tabs on your screen.
Super-users will pay enterprise money for raw compute
The pricing observation ties back to the opening. Software was historically sold per seat — a flat $20/month. But Perplexity noticed individual "pro-sumers" and small-business owners burning so much compute to automate their work that their run-rate value was reaching $20,000–$30,000 a year.
So they're embracing usage-based pricing — selling compute credits — alongside flat subscriptions. The logic: if one person can extract the ROI of an entire department from the AI, they'll happily pay department-level prices for the compute. Perplexity is actively encouraging this with programs like its "Billion Dollar Build," nudging users to build whole businesses on Perplexity's compute as the backbone.
This is a pricing story but it's also a go-to-market-as-engineering story: they watched a usage signal in the data, found the power users paying 1,000x the list price in disguise, and built a pricing model to match the behavior instead of forcing the behavior into the model.
Frequently Asked Questions
What does "compute budget" mean for employees?
It's the idea that as AI agents do more of the execution, workers will be allocated a budget of AI compute the way they're currently allocated headcount or software seats. Instead of asking "how many people are on my team?", a high-leverage employee asks "how much compute can I deploy?" — and gets judged on the return they generate from it. Dmitry Shevelenko of Perplexity frames this as every employee effectively becoming an executive who manages resources rather than just doing tasks.
Why does Perplexity think agentic commerce (AI checkout) is overhyped?
Because users want AI to do the research on a purchase but want to click "buy" themselves. People are sensitive about spending their own money and want a final human check to avoid errors like booking the wrong flight. On top of that, modern checkout (Apple Pay, saved cards) is already low-friction, so automating the final click removes very little pain. Perplexity focuses on the research phase instead.
Why won't Perplexity put ads in its AI product?
Their position is that the moment an AI product carries advertising, users assume the answers are biased toward whoever paid the most — even if the organic results are untouched. Because Perplexity sees itself as being in the "trust and accuracy business," it judges that preserving user trust is worth more than the revenue advertising would bring.
What is Comet and why does it run client-side?
Comet is Perplexity's agentic web browser. Server-side agents that run in the cloud often fail at real tasks because they hit paywalls, captchas, and can't access a user's logins. By running on the user's own machine, Comet inherits the user's cookies, login state, and preferences — so the agent can actually complete authenticated, multi-step web tasks.
What skills matter most in a world of capable AI agents?
Two stand out from the interview: error validation (developing the taste and heuristics to catch where an agent is wrong before its output ships) and management (giving concise, actionable feedback to correct an agent, the way you'd manage a junior employee). Above both sits curiosity and agency — when execution becomes a commodity, the scarce skill is figuring out what work is worth doing at all.
The bottom line
The future of work in this telling isn't AI replacing humans wholesale. It's a hybrid: small, intensely curious teams orchestrating enormous amounts of compute to run long-horizon projects. Budgets migrate from software seats to AI payroll. And the people who win are the ones best at managing a digital workforce — which, conveniently, is mostly the old skill of good delegation pointed at a new kind of worker.
The part I can't shake is the inversion at the end: for years we optimized people for execution. The whole point of school and most of the point of early-career work was to get fast and reliable at doing the thing. If Shevelenko is right, that's the part that just got commoditized — and the soft, unteachable stuff we always undervalued, curiosity and taste and the nerve to ask a better question, is the only thing left that's scarce.
The full interview is worth the watch. If the future-of-work angle interests you, this also pairs well with my notes on why AI is a digital employee, not a tool and what happens when AI starts writing its own code. I'd love to hear where you land — does the "compute budget" framing feel real to you, or like a few years early?