Perplexity launched its AI agent product, called Computer, in late February 2026. It is a tool that takes a goal you describe, breaks it into steps, and executes those steps on its own using software, web browsers, connected apps, and file systems. You give it an outcome; it figures out the process.
What makes the Harvard study unusual is that it compares Computer directly against Perplexity's own search product, using matched pairs of tasks from the same users. If someone searched for "competitive pricing analysis for our market" and also ran that task on Computer, both sessions appear as a pair. The team found 10,000 such pairs, which is enough to draw real conclusions.
The raw time gap is large. Computer runs 26 minutes of machine work per session, against 33 seconds for search. That is a 48-to-1 difference in how long the system is actually working on your behalf. But the more revealing number is what happens when you count in the human doing the remaining work. A person using search to complete a task takes about 269 minutes in total. The same person using Computer takes 36 minutes. The agent does not just speed up one slice of a task; it collapses the whole workflow.
The cost comparison shows something most people get backwards. Computer costs $4 to $10 per task in AI model charges. Search costs about five cents. On a per-task basis, the agent looks expensive. But when you break it down by individual step, the agent costs $0.16 per step while the search-plus-human approach costs $2.05. The agent handles more steps per task; the human pays for fewer of them with their time. The 94% cost reduction in the study accounts for the wage cost of the person doing the work, which is where most of the money actually goes.
The study identifies a useful rule of thumb: if a task will take your staff under 20 minutes to complete manually, the agent is not worth the per-task cost. If it will take longer, the agent saves money.
The finding with the longest implications is about what people use agents for once they get comfortable. Users do not just speed up the same tasks. They start delegating harder ones. Seventy-six percent of tasks sent to the agent required higher-order thinking (analysis, evaluation, creating something), compared to 55% for search. The share involving creative or synthesis work nearly doubled. Tasks also crossed occupational lines more often: the agent routinely handled work that would normally require a researcher, an analyst, and a strategist in sequence. One person, one goal, handed off entirely.
PWC's workforce research flags a consequence of this: as agents take on mid-level execution work, the premium shifts to people who can direct them well. The skill is not doing the task; it is specifying the outcome clearly enough that the agent can run with it.
There is a practical note on the product itself. Perplexity Computer is available only on the company's most expensive subscription tier, at $200 per month for individuals. Enterprise access runs $325 per seat. It is not a casual tool; it is priced for heavy professional use.
There are real things to watch before scaling any agent like this. The work involves credentials, sensitive data, and decisions about what the agent is allowed to do unsupervised. Questions about audit trails, data visibility, and what happens when the agent makes an error have not been fully resolved by any provider yet. Starting with a single, bounded workflow is the right approach: pick one process, run it, verify the output before it goes further.
The study also notes something that should reassure people worried about agents replacing search entirely: Computer adoption actually increased daily search usage by a small but measurable amount among the same users. The two tools are filling different roles, not competing for the same one.