Harvard and Perplexity just put a number on something most of us only felt: an AI agent did 26 minutes of autonomous work per session, while plain AI search did 33 seconds. On matched tasks, that swap cut time by 87% and cost by 94%. Here is what the study actually found, and what it means for how you work.
For the last two years, the pitch around AI at work has mostly been vibes. People say it saves them hours. Founders post screenshots. Everyone nods. But hard, production-scale evidence on how AI agents change the actual shape of knowledge work has been thin.
On June 8, 2026, that changed. Perplexity, working with researchers at Harvard Business School, published a study analyzing real usage data from two of its own products: regular AI Search, and Perplexity Computer, the autonomous agent that launched in February. The paper is titled How AI Agents Reshape Knowledge Work, and the full technical version sits on arXiv.
Here is the thing. This is not a survey where people self-report how great AI is. It is built on roughly 84,000 agent sessions and millions of search interactions, compared head-to-head. That makes the findings worth slowing down for.
Let’s break down what they found, why it matters for anyone who does thinking for a living, and what you can do with it this week.
The Setup: Assistant vs Agent, Measured Side by Side
The study draws a clean line between two ways of using AI.
An assistant answers questions. You ask, it responds, and then you go do the work: open the tools, gather the files, write the doc, check the output. The AI sits between your intent and the finished thing, but you still carry the load.
An agent does the work. You describe an outcome, and it plans the steps, runs them, asks for input when it needs it, and hands back something finished. In this study, Perplexity Search played the assistant and Perplexity Computer played the agent.
The researchers used a smart trick to compare them fairly. They found Search sessions and Computer sessions that started with near-identical opening queries, treating each pair as the same task attempted two different ways. Across 10,000 matched pairs, they could measure what actually changed when you hand a task to an agent instead of doing it yourself with answers in hand.
Finding 1: The Autonomy Gap Is Enormous
This is the headline number, and it is genuinely striking.
On matched tasks, Computer performed an average of 26 minutes of machine work per session, versus 33 seconds for Search. That is a 48-fold increase in autonomous work on effectively the same request. At the median it was 9 minutes versus 14 seconds, roughly 40 times more.
What is the agent doing for those 26 minutes? Searching, browsing, writing, editing, running code, and checking its own intermediate results. The work you would normally do by hand after Search gives you an answer, the agent just does.
It also reaches across your connected apps. Computer made at least one external connector call in 7.9% of sessions, against 1.8% for Search, and averaged roughly 12 times more connector calls overall. In plain terms, it does not just run longer, it pulls data and takes actions across the services you have plugged in, instead of leaving that to you.
And here is the part that matters: longer autonomy did not mean lower quality. Measuring dissatisfaction by what users did on their next turn, Computer’s meaningful dissatisfaction rate was 1.3% versus 2.9% for Search, a 55% reduction. More autonomy, fewer complaints.
Finding 2: The Time and Cost Savings Are Not Subtle
To turn autonomy into a dollar figure, the researchers compared two setups on the same tasks. In the first, Search handles the research and you do the execution by hand. In the second, Computer runs the whole workflow and you scope the task and review the output.
Since you can’t directly watch how long a task would take a human, they triangulated using three independent methods: a tool-based estimate, an estimate generated by a large language model, and 25 interviews with real Computer users about their pre-agent workflows.
The tool-based estimate landed here:
| Approach | Average task time | What you do |
|---|---|---|
| Search + Human | 269 minutes | Research with AI, then execute everything manually |
| Computer + Human | 36 minutes | Scope the task, let the agent run, review the result |
That is an 87% reduction in task time. When the researchers priced human labor using U.S. Bureau of Labor Statistics wage data and added in model cost, the total task cost dropped by 94% on average. The cost savings beat the time savings because human hours are expensive and machine compute is cheap. For context, the study put Computer’s model cost at roughly $4 to $10 per task, while Search runs around $0.05.
The pattern held across all 18 domains studied, with 79 to 92% time savings and 87 to 96% cost savings. Programming was the most extreme: 596 minutes for Search + Human versus 48 minutes for Computer + Human, a 96% cost reduction. The other estimation methods pointed the same way, with the user interviews showing a median speedup of 25 times.
What this really means: the expensive part of knowledge work was never the thinking. It was the manual execution wrapped around the thinking. That is the part agents are eating.
Finding 3: People Take On Bigger, Stranger, Harder Work
This is the finding I find most interesting, because it is not about speed at all. It is about ambition.
When the same people used Computer instead of Search, the kind of work they attempted changed in two directions.
Horizontally, they crossed into other professions. Computer users worked outside their primary occupation 59% of the time, versus 50% for Search. A marketer attempts light coding. A founder attempts design work. The agent handles the specialist execution they could not do alone, so they stop outsourcing it.
Vertically, they reached for harder thinking. Scored against Bloom’s classic taxonomy of cognitive complexity, 76% of Computer queries required higher-order cognition, versus 55% for Search. The biggest jump was at the very top: 50% of Computer queries were “Create” level tasks (building something new), against 26% for Search. Computer tasks also pulled on more knowledge domains per query and bundled more interconnected subtasks into a single request.
The cleanest signal: about 23% of Computer queries involved a specific task that the same users never once sent to Search. These agent-only tasks clustered in software and web development, documentation, and data visualization. The summary the researchers offer is sharp: Search explains, Computer produces.
The Real Shift: You Stop Being the Operator
Put the three findings together and a single change emerges. When a system can search, browse, code, edit files, connect to your apps, and deliver a finished thing, your bottleneck moves.
You spend less time operating the workflow and more time specifying goals, supplying context, checking outputs, and asking for the next thing. In the study’s words, the user moves from operator to supervisor.
That sounds small. It is not. If one person with an agent can complete work that used to need three different specialists, the long-run story is not really about saving 20 minutes. It is about how roles get defined and how teams get built. The researchers are careful to flag this as an open question, not a prediction, but the direction is hard to miss.
How to Put This Study to Work This Week
Reading research is fine. Changing how you work is better. Here is a practical way to test the core finding for yourself, whether or not you use Perplexity specifically. The same logic applies to any capable agent, including Claude Cowork or ChatGPT’s agent modes.
- Pick a task you normally grind through manually. Not a question you need answered. A deliverable you need produced. A competitor analysis, a formatted report, a small data cleanup, a draft landing page. Something with multiple steps.
- Write the outcome, not the prompt. Instead of “what are good project management tools,” try “research the five most-used project management tools for remote teams in 2026, pull current pricing and the top complaints from review sites, and format it as a comparison table I can share.” Describe the finished thing.
- Hand it to an agent and walk away. The whole point is autonomy. Resist the urge to hover. Let it run, and note when it pauses to ask you something. Those pauses are the checkpoints, not failures.
- Time the old way honestly. Before you judge the result, estimate how long this would have taken you start to finish, manually. That number is your real baseline, and it is usually bigger than you think.
- Review like a supervisor, not a doer. Your job now is to check the output, catch errors, and ask for extensions. This is the muscle the study says matters most. Build it deliberately.
- Try one task outside your lane. If you are non-technical, attempt something technical. If you are technical, attempt something creative or strategic. This is where the “scope expansion” finding becomes real for you personally.
Do this three or four times and you will feel the operator-to-supervisor shift in your own work, which is far more convincing than any chart.
The Honest Caveats
The researchers are refreshingly upfront about the limits, and you should be too before treating these numbers as gospel.
It is early, and early adopters are not normal. The people using an autonomous agent three months after launch skew heavily toward AI-native users. Their baselines and habits may not match the broader workforce.
The time estimates are estimates. You cannot directly observe how long a task would have taken a human, so the savings rest on assumptions. The three methods agreeing is reassuring, but the exact magnitudes should be read as approximate, not precise.
It only sees inside one ecosystem. The data captures activity within Perplexity’s products. Work you do in other tools is invisible to it, which limits the full picture of how anyone actually works.
Matched sessions are a useful but noisy proxy. Pairing similar queries is clever, but people do not always organize their work into neat sessions, so some real agent tasks have no Search equivalent to compare against.
None of this erases the findings. It just means the right takeaway is directional: agents move work from manual execution to supervision, expand what people attempt, and do it at a fraction of the cost. The precise decimal points will keep moving.
The Bottom Line
The era of arguing about whether AI helps at work is basically over. We now have production-scale evidence that autonomous agents do dramatically more of the execution, at a fraction of the time and cost, without sacrificing quality, and that people respond by reaching for bigger and broader work than they would attempt alone.
The skill that compounds from here is not prompting. It is supervision: scoping a goal well, supplying the right context, and reviewing output with judgment. The study is a snapshot of a transition that is still early. The smart move is to start practicing the supervisor role now, while it is still a choice rather than a job requirement.
If you have a colleague who still thinks AI at work is hype, send them this. The numbers do the arguing for you.
Sources and further reading
- How AI Agents Reshape Knowledge Work, Perplexity Research (June 8, 2026).
- How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope, technical report on arXiv (2606.07489).
- The Adoption and Usage of AI Agents: Early Evidence from Perplexity, Harvard Business School Working Paper.
- A New Study from Harvard and Perplexity Finds AI Agents Perform 26 Minutes of Autonomous Work per Session, MarkTechPost (June 8, 2026).
- Occupational Employment and Wage Statistics, U.S. Bureau of Labor Statistics.
Just Being Resourceful! 
Leave a comment