Who is adopting AI agents in 2026, and what they actually do with them

When the HBS Working Knowledge story ran the headline number, the room reacted the way rooms do when a category finally has measurement. Knowledge workers are the heaviest users of AI agents, and 36% of what they send to those agents is some form of productivity work. Academics and finance are 10% of adopters, and the top six use cases add up to almost all of the activity that gets logged.

This matters because the AI agent story has been a vibes story for two years. The HBS writeup of a working paper from Perplexity co-authored with HBS economist Jeremy Yang is the first clean look at who is on the other side of the chat window and what they are typing. It is also a useful corrective for anyone who thinks the agent story is mostly about replacing customer support reps or writing marketing copy.

The adopter profile, narrower than the marketing

Perplexity's Comet browser and its Comet Assistant agent are the data source. That has obvious limits. Comet is a single product from a single company, the user base skews toward people who already like trying new AI products, and adoption is concentrated in higher GDP countries with higher education attainment. Yang and his co-authors say that out loud. The findings describe Comet users, not all agent users.

Even with that caveat, the shape of the adopter base is informative. Digital technology is the largest career cluster at 28% of adopters. Academics and financial workers are 10%, and marketing, design, and entrepreneurship are 5%. The next three buckets are education, healthcare, and software engineering, and after that the long tail of everything else. The profile is consistent with what the OpenAI enterprise sales team and the Anthropic sales team have been saying in their respective earnings calls for two quarters. Early adoption is dominated by people whose day job is moving information around, not people whose day job is moving boxes around.

The interesting question is not whether knowledge workers are the heaviest users. It is whether that pattern holds as the products get cheaper and the user interfaces get friendlier. If it does, the labor story is mostly an information work story for the next two to three years. If it does not, the labor story gets much messier much faster.

Where agent adoption shows up, and where it does not

The use case breakdown is the part of the data that should be quoted most often. Productivity and workflow tasks are 36% of queries. The Yang paper lumps that bucket together, but the examples in the writeup are document and form editing, account management, email management, and spreadsheet and data editing. Learning tasks, including watching class videos and summarizing key content, are 21%. Media and entertainment, mostly browsing and posting, is 16%. Shopping and commerce is 10%. Travel and leisure, mostly flight and hotel booking, is 7%. Career related searches and applications are 7%. The remaining 3% is the long tail.

Two patterns are worth flagging. First, the dominant activity is not search and it is not chat. It is delegated execution. The Comet example Yang opens the article with is a Boston to San Francisco trip booked in one prompt. That is closer to handing a task to an assistant than it is to searching the web. Second, the mix is heavily skewed toward cognitive work, not transactional work. That is consistent with what we have been seeing in the agent benchmarks and in the live ops work at our comparison of the best AI coding agents in 2026. The agents that ship in 2026 are better at research and synthesis than they are at production transaction work, and the usage data lines up with where the products actually work.

Travel and shopping together are only 17% of activity. That is a surprise to anyone who has read the agent press releases. The narrative out of OpenAI, Google, Microsoft, and the agent startups is that consumer transactions are the killer app. The HBS data does not support that narrative yet. The reason is probably that the agent products are not yet reliable enough for a flight booking that the user cannot afford to mess up. The reliability gap shows up directly in the usage data.

The 21% learning share is also a signal that the agent market is being driven by information discovery as much as by information creation. Exa's recent $250 million raise to become the search engine for AI agents lines up with that read. Agents need to pull live web data, structured sources, and document collections before they can synthesize anything, and the data plumbing underneath the agent layer is starting to look like its own venture category.

Adoption is concentrated in higher GDP countries and countries with higher education attainment. The paper does not break that out further, but the practical implication is that the early agent market is not a global consumer market. It is a wealthy, educated, knowledge work market. That is consistent with the cost structure of the current products. The frontier model APIs that power most agents are priced for a small number of high value tasks, not for the long tail of consumer queries. The market is shaped by the price, and the price is shaped by the cost of running a frontier model on a hard task.

The income skew also has a policy implication Yang flags directly. If uneven adoption widens existing disparities, the policy question is no longer whether to regulate the agents. It is how to subsidize access to them. That is a different conversation from the safety and alignment conversation that has dominated the policy debate so far, and it is one that regulators in the European Union, the United Kingdom, and several US states are starting to have.

The Yang paper also points at a second order design implication. If a website is being visited by an agent more often than by a human, the right interface is not the right interface for a human. This is a real product problem for anyone who runs a SaaS application with a public surface. The agent first interface is going to look different from the human first interface, and the agent first interface is going to need a new set of design patterns that the current generation of product teams has not built yet.

The most important caveat is what the data does not show. The Yang paper measures who is using Comet and what they are typing. It does not measure whether the task got done, whether the user was happy, whether the agent saved time, or whether the user would have done the task at all if the agent was not available. The right way to read the data is as a map of intent, not a map of productivity. A 36% share of productivity queries does not mean 36% of work hours are getting done by an agent. It means 36% of the things people try to delegate to an agent are in the productivity bucket.

There is also a selection bias that runs in both directions. The most enthusiastic agent users are overrepresented in the Comet data, because people who try a new product and stick with it are the people who find it useful. The people who tried a Comet trial and bounced are not in the data. That makes the data a strong map of what agent power users do, and a weak map of what the average knowledge worker would do if the product were free and built into the browser they already use.

Finally, the data is one product, one company, one country mix. The pattern of who uses agents might look very different in the next data set, especially as the user base of ChatGPT, Claude, and Gemini in the workplace starts to be measured properly. The HBS data is the cleanest single look we have so far, and that is what makes it worth writing about, but it is the beginning of a research thread rather than the end of one.

What the data means for builders and policy teams

The HBS data is the cleanest reason yet to focus the next phase of agent product work on knowledge work, on productivity, and on workflows that can be completed end to end with the current model capability frontier. The bad news is that the current frontier is not yet reliable enough for the long tail of consumer transactions. The good news is that the productivity bucket is large enough to build a real business against, and that is the bucket where the model capability is already strong enough to ship a product that holds up under daily use.

For policy teams, the data is the cleanest reason yet to start the conversation about subsidizing access. The agents are going to be useful first to people who are already in the top income and education deciles, and the gap between the top decile and the rest of the population is going to widen unless someone acts on it. For product teams, the data is the cleanest reason yet to start designing the agent first interface, because the people using the agent today are the people who will be using the agent heavy products in two years.

The HBS study is not the final word on who is adopting AI agents in 2026. It is the first word backed by real data at scale, and that is exactly the kind of data product teams and policy teams need to read carefully right now.