What the Exposure Map Can Tell You

The headline reaches the office before the new software does.

Someone has posted a chart in the staff chat. Their occupation appears near the top. A large share of its tasks, the chart says, could be affected by artificial intelligence. By lunch, the phrase has become “a large share of our jobs.” By the end of the day, a colleague is wondering whether to cancel a course she has already paid for.

This is a costly little journey for a sentence to take. It begins with a judgment about what a system might do and ends inside a person’s plans for next year.

An exposure map deserves attention. It can identify places where work may change sharply. Yet the map cannot do the rest of the thinking for us. We have to ask what was measured, how the work is actually done, and what happens between a task becoming easier and a household receiving less money. Or more money. That route exists too.

I want people to have a share in productive capital whichever route their own job takes. The case for that share should survive both a frightening forecast and a cheerful one. If we build the case only on a particular unemployment number, we make a durable right depend on winning a guessing contest.

There is a better use for the map. Let it tell us where to look closely. Then follow the work.

Begin with the verbs

A job title hides a crowd of activities.

Consider the person who runs the front desk at a small appliance repair business. She reads messages, books visits, explains charges, checks which parts have arrived, and updates customers. She also notices that two bookings cannot fit into the same morning because the technician must collect a replacement part on the other side of town. A customer insists that the machine was repaired last month. The record says last year. Someone needs to find out what happened without making the customer feel foolish.

“Administration” fits much of this work. It explains very little of it.

Write the verbs instead. Read. Compare. Book. Explain. Check. Decide. Each verb still needs an object and a setting. Compare what? A clear invoice with a clear order, perhaps. Or a vague complaint with a history scattered across three systems. The same verb can conceal very different demands.

An AI tool might turn the customer’s message into a tidy summary. It might draft a reply and suggest a booking. To finish that booking, however, it needs current information about the technician, the part, the route, and the promise already made. Some of these facts may exist only in a colleague’s head. Some may be wrong in the database.

None of this makes automation impossible. It identifies work needed to make automation dependable. The business can connect its systems, improve records, change the booking rules, and give the tool authority within set limits. Those changes cost something. They can also improve the business beyond the original task. A reliable parts record helps people as well as software.

The task changes when its surroundings change. That is why a fixed label such as “safe job” or “doomed job” is often a poor companion for a person trying to plan.

The International Labour Organization’s 2025 exposure index combined task information, worker assessments, expert input, and model predictions. It estimated that one in four workers worldwide held an occupation with some exposure to generative AI; 3.3 percent of global employment was in its highest exposure category. Clerical occupations remained especially exposed. The authors judged job transformation more likely than wholesale replacement because most occupations contain tasks that still require human input.¹

Those are estimates of exposure, with categories defined by the researchers. They are not counts of people who have lost work. Even the highest category does not supply an appointment date with unemployment.

This distinction should sharpen our concern. A person can remain employed while losing hours, learning opportunities, discretion, or bargaining power. Another can keep the same title while gaining better tools and doing work that was previously beyond reach. A headcount alone misses both changes.

Capability has to travel

At the repair business, suppose the system can prepare a sound first reply in seconds. That is a technical achievement. Now watch it travel.

The firm has to obtain the tool and use it. A demonstration viewed by the owner is not adoption. A subscription bought for the office is not regular use. Staff may lack training, access to the right data, or permission to rely on the result. They may also find the product disappointing. A purchase is a decision to try, not evidence that the trial worked.

Next comes reliability in the actual setting. The cheerful demonstration used complete information. Tuesday’s customer supplies half a model number and a photograph of the wrong label. Can the system ask a sensible follow-up? Does it know when the information is inadequate? Can staff discover an error before it becomes a promise the firm must honor?

Then the gain has to survive the whole process. A reply that takes less time to draft may take more time to check. Or it may genuinely save time after checking. The answer can vary by message. Routine appointment reminders and disputed repair guarantees need not share one rule.

Once the gain is real, the business has a choice about its use. It could answer a backlog, improve service, pursue more customers, reduce overtime, or reduce staffing. These are different uses of the same released capacity. They also have different effects on the people doing the work.

Demand enters here. If faster service brings enough additional business, the firm may need more technicians and perhaps more office staff. If the local market is already well served, extra capacity may sit idle. Lower prices might attract customers who had been putting off repairs. They might also mostly shift customers from a rival. Growth at one business does not tell us the net change across all businesses.

Finally, whatever money is created must be divided. A tool vendor may charge a fee. Customers may receive lower prices. Owners may receive higher profits. Workers may receive a raise, shorter hours with maintained pay, or nothing beyond a fuller schedule.

There are decisions at every stage. Some are made by the people in this office. Others depend on markets, contracts, public rules, and whoever owns the systems the office must use. A model’s ability to draft a reply cannot settle those decisions.

Follow the sequence slowly enough to see it: a capability becomes a workable process; the process changes costs or quality; demand responds; people with different rights make claims on the result. Several links may move at once. None should be silently assumed.

A real gain can sit beside a real failure

We have evidence that AI assistance can improve work, including work outside a laboratory.

Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a staggered rollout of an assistant among 5,172 customer-support agents. Their published 2025 study estimated an average increase of about 15 percent in issues resolved per hour. Less experienced and lower-skilled agents gained more. The strongest agents had smaller speed gains and small declines in quality. This was a particular support system in a particular organizational setting, not a universal percentage for every office.²

The result gives us a good reason to take the opportunity seriously. It also suggests a practical question for a workplace trial: who benefits, on which tasks? An average can hide the person who has been helped most and the person whose work has become worse.

A separate experiment led by Fabrizio Dell’Acqua involved 758 consultants. On tasks within the tested system’s capabilities, AI users completed 12.2 percent more tasks and took 25.1 percent less time, with higher quality. On a task chosen to lie beyond those capabilities, AI access reduced the chance of a correct answer by about 19 percentage points. The researchers called this unevenness a jagged technological frontier.³

The awkward part is that a boundary may run through one person’s afternoon. A system can help create an excellent presentation and mishandle the analysis on which its recommendation rests. It can make the wrong answer easier to read.

That observation should improve adoption. It should lead us to test the kinds of work that matter, retain checks where errors are costly, and change the process when evidence warrants it. It should not become a permanent excuse to deny people tools that help them.

Nor should one impressive result become permission to remove every check. Speed is an achievement when the work reaches an acceptable destination. Sending a customer a polished account of a repair that never happened is merely a faster way to create another task.

There is room here for ambition and discrimination. We can want much more capable AI and insist on knowing what the current system can do. In fact, wanting progress gives us a reason to measure its weak points accurately. A false success report leaves the weakness in place.

The job can change without disappearing

The office now handles routine messages quickly. What becomes of the person whose day once contained many of those messages?

One possibility is better work. She spends more time solving unusual problems, following up on repeat failures, and improving the booking process. The business gives her authority to do this and recognizes the change in her role. A tiresome part of the job has shrunk, and her contribution has grown.

Another possibility looks similar in the first week but feels different by the third month. She handles all the difficult customers because the simple contacts have been automated. The screen shows fewer completed cases than it used to, because the remaining cases are harder. Management retains the old target. Her work has become more demanding while the measure makes her appear slower.

A third possibility is that the business needs fewer paid hours at the desk. That change can occur through vacancies left unfilled, reduced overtime, shorter contracts, or layoffs. Different routes matter to different households. A firm announcing no layoffs has not necessarily maintained every route into employment.

We cannot select among these outcomes by reading the occupation’s exposure score. We need information about how the employer is redesigning the role, what demand exists, and what workers can negotiate.

There is also the person hoping to enter the occupation. A task that looks dull to an expert may have been a manageable place for a beginner to learn. If the easy cases vanish, the business needs a different way to prepare someone for hard cases. It may build one. It may decide training is another organization’s problem. The number of experienced workers employed today tells us little about that choice.

This is one reason to resist a narrow contest over whether AI “creates” or “destroys” jobs in the abstract. We need to know which jobs, for whom, over what period, with what pay and path to skill. A new role in a distant city is not an immediate replacement for a lost role where someone lives. Over time, people and institutions can adapt. Adaptation still has to be organized and paid for.

I do not want the map used to frighten workers into accepting any terms. “Your occupation is exposed” can become a cheap negotiating tactic. Technical possibility gives an employer something to investigate. It does not grant moral authority to demand unpaid transition work or announce that everyone should be grateful to remain.

The same map can support a stronger response. If substantial change is plausible, training, consultation, and a share of the gains belong in the plan while the plan is being made.

Make a small map of the work you know

You do not need to rate every activity in your occupation. Start with one process that produces something another person needs: a completed booking, a corrected account, a delivered design, a resolved request.

Trace it from the first input to the point where the next person can rely on it. Ask the colleague who receives the work what tends to be missing. Ask whoever corrects it what failures consume time. Their answers may change your idea of where the work begins and ends.

Mark the parts the tool can already help with under your employer’s approved rules. Beside them, mark what someone must still supply or decide. Keep uncertain parts visible. An honest “we have not tested this” is more informative than a confident color on a chart.

Then identify the condition that would make the process genuinely better. A shorter queue? Fewer repeat visits? Clearer instructions that customers can act on? Choose something close enough to the work that the team can recognize success.

A comparison should include the awkward cases the process will really encounter. If the trial selects only clean, easy requests, say so. It may establish a valuable narrow use. It cannot establish performance on the messier work left outside the trial.

This small map gives workers a way to enter the discussion with knowledge. It also gives a manager something better than a declaration that the whole department is “AI ready.” Ready to do what, for which customer, under which conditions? Those questions make a trial useful.

Do not turn the exercise into a permanent diary of every movement. We are trying to understand a process, not make a person prove the worth of each minute. A short, shared review of a few representative cases can reveal more than a month of unexplained activity logs.

And ask what happens if the trial succeeds. This question belongs near the beginning. If the only stated reward for saving time is a higher target, people have reason to worry about helping. A proposed share in the benefit changes the conversation. It might begin with pay or time, but it should also open the larger question of ownership.

A worker’s right to share in productive wealth should not expire when a particular task becomes easy to automate. Someone outside paid employment should have that right too. The exposure map cannot make this moral choice for us. It can show why leaving income attached to one person’s changing bundle of tasks is a fragile foundation for universal security.

We can use AI to make much more possible. That is the opportunity. We can also decide, ahead of the next dramatic chart, that everyone belongs among the people who own a share of the capacity we are building.

The map shows where the work may move. The claim on the future is ours to write.

Sources and notes