Harvard Business School and Perplexity published a field study this week with a number in it that is easy to quote and easy to misread. Across a 90-day window running February 27 to May 27, Perplexity’s autonomous agent — the product it calls Computer — averaged 26 minutes of work per session. Its conversational search product averaged 33 seconds. A 48× gap.
The efficiency claim is bigger still. Working from more than 84,000 agent sessions across 18 domains, the researchers matched roughly 10,000 near-identical query pairs — cosine similarity above 0.99 — and found that agent-plus-human cut estimated task time by 87% and estimated cost by 94% against search-plus-human. Users complained less, too: a meaningful-dissatisfaction rate of 1.3% versus 2.9%.
None of that is the finding that matters. This one is: about 23% of the tasks people handed the agent were tasks those same people had never once asked search to do.
Speed changes the clock. Scope changes the job.
The paper maps each query to what it calls a task statement — the underlying job behind the request, tied to occupational categories. Agent queries crossed occupational lines more often than search queries, 59% versus 50%, and demanded higher-order cognition more often, 76% versus 55%. People were not simply doing their existing work faster. They were reaching for work they had not previously attempted: the marketer building the model, the analyst drafting the deck, the founder running the competitive scrape.
That is the part with career consequences. If a tool only makes you faster, your job description holds and your output rises. If a tool widens what you can credibly attempt, the boundaries between roles go soft — and whoever moves first into the newly reachable work resets the baseline for everyone else’s job description.
Our take: Read the 87% as marketing and the 23% as strategy. The time-and-cost figures are modeled estimates, and Perplexity co-authored a paper about Perplexity’s own product — directionally interesting, not settled. The 23% is a different kind of number: a behavioral observation about what users actually chose to do, and the one you can act on this week. Pick the task you have been outsourcing, deferring, or quietly avoiding because it sits outside your skill set. That is where the leverage is. Speed on work you already do is a raise you have to go ask for. Scope on work you couldn’t do before is a promotion you can simply take.
The cost nobody prices in
Twenty-six minutes of autonomous work is also twenty-six minutes you did not supervise. Every headline finding here concerns what the agent produced and how satisfied users felt — not how often the output was correct, and not who is accountable when it isn’t. Google’s robotics release this week made the same point in a different domain: the demo is the easy half, and the success-rate table is the document worth reading. An agent that hands you a finished spreadsheet in an unfamiliar field is also handing you the one document you are least equipped to check.
Which is the honest version of the scope story. Widening what you attempt only pays if you widen your judgment at the same rate. Otherwise you have not expanded your job — you have expanded your exposure. The delegation question is not what the agent can do; it is what you can still verify.
What to watch
- Your own 23%. List the tasks you attempted this quarter that you would not have attempted last year. If that list is empty, you are using an agent as a faster search bar.
- Verification cost. Time saved in production reappears as time spent in review. Budget it on purpose or it quietly eats the gain.
- Vendor-authored research. This will not be the last study in which the company shipping the product co-signs the paper measuring it. Read the method section and ask who benefits from the headline.
- Role boundaries in job postings. When adjacent skills move from nice-to-have to requirement, the scope shift has been priced in — and the window for taking it for free has closed.
The headline says an agent is 48 times more patient than a search box. The useful sentence is quieter: people started asking it for things they had never asked for before.