How Computer is Reshaping Knowledge Work

This week, in collaboration with researchers at Harvard Business School, we published a new study looking at how people actually use Perplexity Computer in the real world. And how - in its first three months of existence - it is already changing the shape of knowledge work. Here is a short version of what we found, and why we think it’s worth paying attention to.

Why we did this research

Over the last few decades, computers have moved closer to our work. Each new interface, from the command line to the GUI to the touchscreen, took on more of the task and reduced the burden on the user.

As interfaces became simpler, so did our questions. But users still had to break requests into steps, enter them manually, and piece the results back together.

Perplexity Computer changes that. It orchestrates over 20 frontier models across files and tools, turning a request into teams of agents that understand the goal, divide the work, and act asynchronously. The user asks for the outcome, not the process.

We wanted to know what happens when people actually start working that way.

What we found

1. Users are asking for work, not answers. 

Search provides answers. Computer executes them. On near-identical requests, Computer ran an average of 26 minutes of machine work compared to 33 seconds for Search - a 48x increase on the same task. Those minutes go into searching, editing, running code, and reaching into connected tools - Computer made about twelve times as many calls to external apps and connectors per session as Search. User effort is concentrated at both ends: setting the objective and reviewing the result. The middle is handled by the agents.

2. It is meaningfully faster, and cheaper.

Across 10,000 matched pairs of tasks - the same kind of request made on both Search and Computer - Computer cut estimated time by 87% and cost by 94%. A workflow that took 269 minutes dropped to 36. 

The savings held across 18 domains: software, finance, marketing, healthcare, legal, education, and more. In user interviews, the median person reported their work going 25x faster. Some reported speedups over 300x.

3. The asks got bigger, and broader.

With execution handled, the barrier to pursuing ambitious projects drops. 23% of Computer requests were for things people had never asked Search to do - clustering in software, web development, documentation, and data visualization - tasks, not questions.

Half of all Computer requests asked Computer to produce something - twice the share on Search - and 76% required analysis, evaluation, or creation rather than lookup, versus 55% on search. 71% involved abstract, non-routine work, compared to 53% on Search. A typical Computer task spanned 2.4 areas of expertise, whereas Search drew on only 1.74, and was nearly three times as likely to need 3 or more domains.

People also worked outside their own field. Across 8,000 users from eight occupation groups, people did work outside their primary occupation 59% of the time with Computer, versus 50% with Search. 

4. Quality did not drop.

Taking on harder work across different domains did not lower quality. Multi-model orchestration lets Computer match each step to the model best suited to it, so a request that spans research, analysis, and design is handled by the correct model at each stage rather than one model for all of it. This routing allows the work to remain at the same quality bar even as the requests got more complex.

Why this is interesting

AI is fast. That part is expected.

The more interesting finding is what people chose to do with the time back. When execution gets cheap, the calculus changes. Work that wasn't worth doing before is suddenly worth doing. Work that used to require a specialist is something you can take a credible swing at yourself.

The bottleneck in knowledge work was rarely finding information; it was capacity to execute. Perplexity Computer takes on this capacity, and as a result, people are aiming higher. They move from operator to supervisor - spending less time running the workflow and more time setting the goal and checking the output.

What's next

We are still early. Computer is only three months old, people are getting accustomed to working with it, and the workflows and usage will undoubtedly adapt over time.

But this data already demonstrates individual-level shifts in behavior. 

How it aggregates at the team and organization level is the open question, and the one we are most interested in next.

You can read more about this research, including the full methodology, charts, and caveats, in the technical blog and research paper.