Will AI Take My Job? What the Numbers Say

Cohere Labs published a dataset on September 3, 2026 that gives an unusually concrete answer to a question usually answered with adjectives. Researchers collected 696,291 published AI agent tools from 123,069 public servers and matched each one against the 923 occupations in O*NET, the U.S. government's occupational database. Of those 923 occupations, 419 had no agentic tool activity of any kind.

What the research actually measured

The Agentic Task Ecosystem dataset applies a strict test. A tool counts only if it completes a recorded work task from start to finish, rather than helping a person do one or handling a single step in a process a human coordinates. Under that rule, about one tool in forty, or 2.6 percent, qualifies. Those tools cover 1,380 task statements, roughly 15 percent of the software-performable work catalogued in O*NET.

Three caveats matter before anyone quotes this at a staff meeting. It is a supply-side map of what developers published on public servers, not evidence that any employer has adopted these tools. Private tooling built inside companies is invisible to it. And the researchers are explicit that the dataset says nothing about reliability, deployment, or whether a tool holds up over a long task, noting that their count can only miss tools rather than overcount them, so 2.6 percent should be read as a floor.

One finding cuts against the usual story. Among the tools that group into recognisable categories, a large share is infrastructure for running agents, plus a small amount of genuinely new work, most of which is the work of managing agents. Automation is generating its own occupations while it absorbs parts of others, which we wrote about in the new jobs AI is creating.

Your job is a bundle of tasks

Almost every job is a bundle of twenty or thirty distinct tasks, and they are not equally exposed. A bookkeeper who spends 40 percent of the week typing numbers off invoices and 60 percent handling exceptions, chasing clients and explaining variances has one highly exposed task and a lot of work that is not. What changes is the ratio, not the existence of the role.

Five traits make a task more exposed. It repeats in the same shape. It happens entirely on a screen. It lives inside one or two systems rather than crossing many. Its output can be checked quickly by someone who knows what right looks like. And a mistake is cheap to correct.

Five traits push the other way. The task needs someone physically present. Somebody has to be accountable, sometimes by licence or statute. The inputs are ambiguous and have to be interpreted. It requires getting agreement between people who want different things. And an error is expensive, public, or both.

Score your own week against those ten and you will get a better answer than any list of at-risk professions. Our job exposure tool runs the same analysis by occupation if you want a starting point.

The five-task audit

This takes about an hour and it is worth more than a year of reading forecasts.

1. List every recurring task in your week. Not projects. Tasks with a start and an end, the things you would hand to a new hire on day one.

2. Put rough hours against each. Guessing is fine, but guess before you score the exposure, so the numbers are not influenced by the conclusion you want.

3. Score each task against the ten traits above. Five exposure traits, five protective ones. Net them out.

4. Add up the hours in the exposed column. That percentage is the honest answer to the question in the title, and it is almost always lower than the headline number and higher than people expect for the tasks they dislike.

5. Try to automate the single most exposed task yourself. Spend two hours with whatever assistant you already have. You will learn more about the real limits in that afternoon than from any report, and you will be the person in the room who has actually tried it.

If you run the business, read this differently

Owners tend to read research like this as a headcount signal. It is a poor one. The dataset measures published tools, and the gap between a tool existing and a tool working reliably in your systems with your data is where most AI projects quietly stall.

In the businesses we work with, the first year of savings usually shows up as reclaimed hours inside existing roles rather than removed positions. That is a real gain, but only if somebody decides what the reclaimed hours are for. Time that gets absorbed back into busywork produces no measurable return and makes the next AI proposal harder to fund, which we covered in capturing the hours AI saves.

There is a second-order risk worth planning for. When AI absorbs the routine work, it takes the work junior people learn on, and teams can end up with no pipeline into senior roles. We wrote about both halves of that in hiring seniors while skipping juniors and keeping a team sharp.

What to actually do about it

The people who come out of this well are the ones who learned what their tools get wrong. As output volume rises, the ability to spot a confident error becomes more valuable rather than less, and that skill only comes from using the tools on work you understand well enough to check.

Start with the most exposed task on your list, automate it properly, and take the hours you free up into the work that scored protective. If your role is being redrawn either way, being the person who redrew it is a better position than being the person it happened to. Our guide to how AI is redrawing job descriptions covers what that looks like across common roles.

Frequently Asked Questions