The State of AI Agents
for Business

For most people, AI at work still means a conversation. You type a question and get an answer back. It's useful, and it's the way most people first met AI.

The businesses in this report work another way. We studied every piece of work their agents did over five weeks. What we found looks like a small organization at work: agents that start their own work, hand tasks to each other and bring people in when a decision is needed.

This report tells that story in seven chapters: where agents are today, the work they deliver, how they work together, how the work splits between people and agents, what the best-run agents show is possible, what the work runs on, and where it goes next.

The most telling number in this report is about who decides when the work begins.

Half of all agent work ran on a routine. Someone decided once that an agent should prepare a brief every morning, check the CRM every hour or send a report every Friday, and the agent has done it ever since. Delegation that stays delegated. About two in three routine runs happen outside weekday office hours, in the evening, overnight or at the weekend, so the routine keeps working after you log off.¹

A third was started by another agent. One agent finished their part and handed the next part on. For every piece of work a person started by message, agents started about two and a half for each other.

When people picture an AI agent at work, they usually picture a sales agent sending cold email. That job is in the data. It sits beside a much wider spread of work.

Businesses build agents around the work they actually have: the inbox and the pipeline, finding leads, code and deploys, the calendar, content and SEO, and finance. We think each agent works best with one clear job, handing the rest on to an agent built for it.

The ready-made agents tell a story of their own. When businesses hire an agent off the shelf, they start with the work closest to revenue: half of those hires are for marketing and SEO, and 43% for sales.

Between them, a business's agents reach a long way into the work.

Put several agents in one team, and the work starts to flow between people's agents.

The agents that do the most are the ones a whole team can call on. In teams, the quarter of agents shared with human teammates did half of all the work. They were also handed 58% of the work other agents started. One person's research agent can finish a brief, and another person's sales agent can pick it up.

In the companies we studied, people and agents work in the same documents, tables and pages. The split of the work is striking. Agents fill in the research, keep the records current and publish the pages. People set the plan, make the final edit and decide what goes out.

Look at where people still type, and you find where judgment lives.

Docs are where people step in most, because docs hold the plan, the pitch and the policy. Tables and pages are where agents carry almost all of it, because that's the work of keeping things current.

The people in these companies moved up, from doing the work to directing it. Every person directs more work than they could do alone.

On those days, agents watch an inbox all day, keep a CRM current every hour, or work through a list of prospects one by one. What sets the busiest days apart is how the agent is set up.

Most of them have a routine that runs without being asked, and a steady flow of work handed over by other agents. The gap between an agent that waits for a message and an agent that runs all day comes down to setup, and any business can close it.

On their busiest days, more of an agent’s work arrives from other agents: 42% is handed over by another agent, against 34% on an average day. People start about one piece in seven either way.

Behind every finished piece of work is a model doing the thinking, and a lot of reading before any writing starts.

When the work goes out in their name, businesses pick the most capable models available. Frontier models such as Claude Opus 5.5 and GPT-6 Astra ran 9 in 10 pieces of agent work.

Agents read far more than they write. For every word an agent writes, they work through about four hundred, most of it their own working context, re-read at every step. A single agent step that uses a business tool typically processes 1.4 to 3 million tokens, five to ten times more than a step of plain conversation.

Five weeks of data can't predict a decade. It can show which way the work is moving. These are our predictions for 2027, labeled as predictions.

Some things stay with people: direction, judgment calls on money and pricing, relationships with customers and partners, and accountability when something goes wrong.

To our knowledge, this is the first report to measure, from real activity data, how agent work begins inside businesses (a routine, another agent, or a person typing) and who makes the changes to the shared work. Others have counted how many agents companies run. We looked at how the work moves.

What we measured lines up with what the people who study business most closely expect.