When you type a goal into a chat window, it comes back as a plan or a draft. You may have a framework in place, but there’s still work to be done. Give that same goal to an AI agent – one that can reach the files and the data – and it comes back as a finished task. At our organization (One Model), that difference explained more of the change in how our work got done than any new tool we rolled out in the period we examined: July to October 2026.
Agent-run work can only happen in tools that support agents. When the share of our typed work run by AI rose, I assumed at first that people had simply moved work out of chat. Some had. Our product, AI Impact, splits the change in two: work moving between tools, and change in how people used the tools they already had. About a third of the rise came from the work moving between tools. Two thirds came from a change in how people used the tools. People who already had an agent let it run more of their work.
Figure 2.1: Of the 3.9-point rise in the AI-run share of typed work between July to August and September to October, 2.6 points came from behavior inside the same tools.
That changes where I'd put the effort. Most AI budgets go on access: more licenses, more products. Access wasn't our constraint. In July, 71 of the 94 people active that month had already run at least one agent task. By October it was 79 of 104. What moved was how much of their work they were willing to hand over.
Figure 2.2: People active each month, and those who ran at least one agent-run task.
Sales is a good example. In July, about 46% of our sales team's AI work happened in agent-capable tools. By October it was 67%, and the share of their work run by AI went from 26% to 36%. Engineering, which got there first, sat at about 58% the whole time. Customer success still does about three quarters of its AI work in chat, and very little of it is run by AI. Same company, same tools available, three different patterns.
How people ask decides who does the work. In September and October, requests that stated a goal made up about 45% of typed activity, and the AI produced almost all of that output. Questions made up about 20%, and there the person did almost all of the work.
Figure 2.3: How the request was made against who produced the output, September to October.
But a goal on its own wasn't enough. In October, goal-style tasks without an agent behind them reached a finished result 67% of the time. Tasks the person steered step by step finished 80% of the time, and tasks an agent ran finished 76% of the time. Handing over the outcome without giving the AI the means to act on it produced more half-finished work than either of the alternatives.
Figure 2.4: Share of tasks reaching a resolved result, and modeled net hours of work per task, by step, October.
When an agent did run the task, it delivered far more. Our model of the work estimates about 0.87 hours of finished work per agent-run task in October, against 0.33 for a task the person directed step by step. So if you want people to delegate more, start with what the AI can reach and act on. Telling people to write better prompts comes second.
When AI runs more of the work, the person's job moves to setting the goal and checking the result. When we looked at this in September, people accepted 99.6% of the actions agents proposed. I can't tell from that number whether it reflects trust that has been earned or approval without reading. Nobody can. What I can see is whether the work held up: in our product we track whether tasks reach a resolved result and how often the next request redoes the last one, alongside who ran them.
Agents also work when people don't. In four weeks from late September, for every ten 15-minute blocks a person spent working with AI, an agent spent about three more working alone, some of them overnight. Someone owns that output. In most organizations nobody has written down who.
Delegation is a skill, and it moves at different speeds for different people. Of the 60 people active in every month from July to October, 19 raised the share of their work run by AI by ten points or more, 31 stayed within ten points, and 10 went down. Some of that is people changing projects. Some of it is people learning to hand work over. Those 19 are who I'd ask to show everyone else how they work.
If you are responsible for how AI is changing your workforce, I'd stop asking which tools people have and start asking how much of the work AI runs, where that is moving and where it isn't. The answer will differ by team more than you expect.
AI Impact measures this from the work itself, task by task, and shows who runs it, how it was asked for and whether it finished.