How the exposure map is made

The idea

Any job is really a bundle of tasks. A product manager, for example, writes requirements, sits in customer interviews, argues over priorities, reports the numbers and does a dozen other things in a typical week, and AI is not equally useful for all of them. It is already good at some, helpful but unreliable on others, and no use at all on a few. So rather than asking whether AI will take a job, this site breaks each job into the tasks that fill a week, works out what today's AI can do with each one, and adds the answers back up. That is what a job page shows: your week, task by task. For each task you see how much of it AI can do today, when working that way will be normal in your industry, and why a person is still needed for it.

Three things make the map specific to you rather than generic. The tasks are described and rated separately at each career level, because a junior analyst and a director both “write the performance review” and yet are doing quite different work. Every task carries its share of a typical week, so a task that takes a fifth of your time counts for more than one you touch once a month. And where a task runs through your industry's own systems or rules, its date is shifted by how quickly your industry takes up new tools, which turns out to vary by years.

A mid-2026 assessment

The ratings describe what AI could do in the middle of 2026, counting only AI that was actually in use at the time. They are a snapshot, not a forecast. When a job page shows a date, it is an estimate of how long until working that way is normal where you work. It says nothing about what AI will be able to do later on.

Catalog version pilot-trackA-v2, published 2026-09-07.

What today's AI can do with a task

Take one task at one level, say a mid-level product manager writing a requirements document. The question the raters answer is this: if you gave this person the AI that works today and let them use it, how much of the time the task takes would still be theirs? The answer is a percentage, and the page turns that percentage into one of four labels. The two outer labels are for the extremes, where one side does nine tenths or more of the work. The two middle labels cover everything in between, and simply say who does the larger share.

Almost entirely by AI

It is already ordinary practice for AI to do nine tenths or more of this task, and the person handles the exceptions.

Mostly by AI, with help from you

A working system can already do more than half of it. The person's part is to direct the system, review what it produces and deal with what it gets wrong.

Mostly by you, with help from AI

The person still does more than half of the work, but AI takes a real share of the time, whether that is the drafting, the searching, the summarising or the checking.

Almost entirely by you

Nine tenths or more of the time this task takes still belongs to the person. Today's AI adds little to it.

These percentages measure time, not importance. A director's performance review still needs the director's own judgment about the person. But if AI gathers the evidence from a year of email and writes the first draft, and the director's part becomes reading, correcting and signing, then most of the hours have moved even though the judgment has not.

When it becomes common practice

The second line on every row is a date, and it answers a different question: not what AI can do, but when working the way leading practitioners do will be common practice. The raters answer it first for the industries that are adopting AI fastest, and the page then shifts the answer for your own industry, which the next section explains. Every date is stored as one of three ranges (this year, one to two years, or two to four years) and, once shifted, it is printed as a plain range of years.

What the date is not is a forecast of when AI will be able to do more of the task than it can today. Nothing on this site makes that forecast.

Your industry

Industries take up new ways of working at different speeds, and for AI the gap is measured in years. The US Census Bureau asks a large sample of businesses every two weeks whether they used AI in any part of the business in the past fortnight, and publishes the answers by sector and by the size of the business. Since the question was last reworded, in November 2025, the share of all firms saying yes has climbed from about 17 percent to about 22 percent, and the information sector, which covers software, internet services and media, leads at over 40 percent. Hotels and restaurants are under 10 percent.

Those are counts of businesses, and most businesses are tiny. You are a person, and half of all private-sector jobs are at firms with 500 or more people, where AI use runs far ahead of the sector's average: among firms with 250 or more employees, seven in ten in information and four in ten in retail and construction say they use AI. So the first thing we do is weight each sector's readings by where people actually work. On that basis Information stands at about 64 percent, and the country as a whole is climbing at about 13 points a year.

Then the arithmetic is one division: how far a sector sits below the leader, divided by the national pace, is how many years it runs behind. Each industry in the table below takes the number of the official sector the Census Bureau groups it under (Banking belongs to Finance and Insurance), because the survey measures at that level and we do not claim finer precision than the data supports. Where the Census suppresses a sector's large-firm readings, the table says so and uses the plain count of businesses instead. Government and farms are not surveyed at all, so those industries are shown as a typical employer. A typical employer, which is what you see if you pick no industry, is the average of every sector weighted by how many people work in it: about 3 years behind the leader. It sits there because most people work in hospitals, shops, restaurants and factories, not in software.

Your industry does not move every date, only the ones it can actually hold up. The reason is in the same surveys. While about one firm in five says it uses AI, about four workers in ten say they use generative AI at work, and in the industries that lead, seven in ten do (the Federal Reserve's April 2026 review, drawing on the Real-Time Population Survey). People are adopting faster than their employers, and that can only be true of work that travels with the person and the tool rather than with the company's systems. A marketing manager at a hotel uses the same writing assistant as one at a software company; the hotel's reservations clerk does not get a new property system because the software company did.

So each task in the catalog carries one of three labels, applied by a fixed rule from the task's own description and checked against a hand-labelled sample. Portable work is done with general-purpose tools and the person's own knowledge, so a campaign brief written with AI arrives at a hotel when it arrives at a software company; 228 of the catalog's 310 tasks are portable and the industry adds nothing to them. System-bound work runs through the employer's own software and data, such as matching transactions, posting a journal entry or executing a trade, and waits for those systems to change. Regulated work is signed by a person because an external rule says so. The industry's years are added to those two kinds (76 system-bound and 6 regulated tasks), and a row that moved says so.

Census readings 2025-11-03 to 2026-08-09 (20 fortnights); table computed 2026-09-08. Weights: Census Statistics of U.S. Businesses (employment by firm size) and BLS employment by sector (August 2026). Sources: US Census Bureau, Business Trends and Outlook Survey, AI use by sector (May 2026); Federal Reserve, Monitoring AI adoption in the U.S. economy (FEDS Notes, April 2026); US Census Bureau working paper CES-WP-26-25, The microstructure of AI diffusion (2026).

Putting the two together

Take one row on the Product Manager page at Mid-level IC: Write a PRD. The raters rated it Mostly by AI, with help from you, and they put the date at which leading practitioners work that way at 1–2 years. It is portable work, a document written with general tools, so the row reads “will become common practice in 1–2 years” in Software & Technology, Banking, Hospitals & Health Systems alike, and for a typical employer.

Now a row on the Financial Advisor page at Mid-level IC: Run a tax-loss harvest, rated Mostly by AI, with help from you with the leading practitioners' date at this year. This one runs through the firm's own systems, so the industry's years are added. In Software & Technology, which is among the fastest, nothing is added and the row reads “will become common practice this year”. In Banking, which runs about a year behind, the row reads “will become common practice in 1–2 years”. In Hospitals & Health Systems, which runs about 3 years behind, the row reads “will become common practice in 3–4 years”. If you have not picked an industry, the page uses a typical employer's 3 years, so the row reads “in 3–4 years”.

Notice what changes and what does not. The label stays the same in every industry, because today's AI can do the same share of the task wherever you work. Only the date moves, and only on the tasks the industry can actually hold up. Where a task lands for you is the raters' date, plus your industry's years if the task runs through your industry's systems.

Your role

For every task, the raters also name the single strongest reason a person is still needed for it, and the page prints that reason as Your role together with how durable it is. A durable reason is one that does not weaken as AI improves. Being the person in the room, for example, is not going to stop mattering, whereas a reason that rests only on what AI cannot do yet is the least durable of all.

Carries the live conversation (very durable)

The real-time contact between people is the work itself, whether that is the meeting, the call or the interview.

Holds the relationship and the trust (very durable)

A relationship, or trust earned over time, carries the work, and a system could not stand in for the person who holds it.

Is there in person, hands on (very durable)

The work needs someone present, hands on, or coordinating things on site.

Owns the decision and signs it off (somewhat durable)

A named person has to own the judgment and stand behind it. This reason is about who signs, not about who does the work, and the section below explains why that distinction matters.

Does the check policy or contract requires (somewhat durable)

Company policy, an audit process or a contract requires a person to check the work or sign it off.

Gives the sign-off the law requires (somewhat durable)

The law, or a regulator, requires that a person does it.

Brings the unwritten company knowledge (less durable)

The task depends on knowledge that is not written down anywhere: who really decides what, what was tried before, and what the words mean in this particular company.

Does the part AI can't do yet (not durable)

The only barrier is technical, because AI cannot do this well yet. Nothing inherently human protects the task, which is why this is the least durable reason of the eight.

Why a senior version of a task can be more AI-led than the junior one

Switch a job page from Senior IC to Manager and you will see some tasks become more AI-led, which looks backwards at first. The reason is what we call the review-and-sign rule. A senior IC who writes the team's requirements is doing the work, but a manager who approves what the team wrote is directing it, reviewing it and signing it off. Signing is accountability, and accountability says who is responsible, but being responsible does not by itself take hours. So a version of a task whose human part is review and sign-off is rated as mostly by AI, with the human's role printed as Owns the decision and signs it off. Where the senior version of a task really does involve more human work, such as live negotiation, reading the room or settling a dispute between colleagues, it is rated that way instead.

Levels

Every job has the same five career levels: Junior IC, Mid-level IC, Senior IC, Manager, Director. They are defined by what a person at that level decides on their own, who reviews their work, how novel the cases they handle are and how far the consequences reach, rather than by years of experience or job title. The same task is described and rated separately at each level where it exists, because the junior version and the senior version of a task are different work, and each job page carries a one-line description of what most defines each level in that particular job.

Share of the week

Each task's share of the week is the raters' estimate of how a typical week at that level was spent before AI. That is the starting point every rating is measured against. The shares add up to 100 within a level. The forty-hour week at the top of a job page is simply these shares added up by label and turned into whole hours: 38 percent of the week becomes 15 of the 40 blocks. The “about N years” beside it is the average of the dates on the tasks, each date taken at the middle of its range and each task weighted by its share of the week, rounded to the nearest half-year. Tasks whose date is open-ended or not in sight are left out of that average.

The development chips

Eight tasks at every level carry a chip, and they are the eight tasks in your Development Plan: the ones where you have most to gain from getting better, chosen by a fixed rule from the same ratings the map shows. The chip says which kind of task it is. Get better at the tasks that stay human marks work that is still done mostly by a person, for a reason that will not weaken as AI improves, which makes it the strength worth building on. Get better at directing AI marks work that AI already does most of, and does well, so the skill that matters is learning to direct it. On the map the eight come first, the one changing soonest at the top. How the plan is made.

How the ratings are made

Everything on a job page traces back to one rating per task per level, and a rating has three parts: the share of the task's time that stays with the person when today's AI is used, the date when leading practitioners will work that way as a matter of course, and the strongest reason a person is still needed for the task. Here is how one rating is produced, step by step.

1. Break the job into tasks. Every job is split into the tasks that fill a week, each named in plain words.
2. Describe each task at each level. The same task is written up separately for every level, because the work is different.
3. Three raters, three angles. Each rater gives the share of time that stays with the person, a date, a reason, and the strongest real-world evidence it can name. The three angles are what is technically possible, what is actually deployed somewhere, and what organisations permit.
4. Two skeptics. One argues only about the dates. The other argues against overrating how much senior work needs a person, and only speaks at the senior levels.
5. Take the middle answer. For the share, the rating is the middle of the three raters' numbers. For the date, the skeptic who argues about dates gets a vote too, and the rating is the middle of the four answers. For the reason, it is the answer most raters gave.
6. Adjudicate where they disagree. If the raters are far apart, or a skeptic dissents, an adjudicator reads the reasoning on each side and decides. In the current catalog about four in ten ratings went this way.
7. Calibrate the whole job. A final pass reads every rating in the job looking for a consistent bias in one direction, and re-rates whatever it flags.
8. The row on the page. The label, the date shifted for your industry, and the reason.

None of this is written by hand, and no rating is ever edited by hand. When one is wrong, we fix the rule or the brief that produced it and run the pass again for every row, so the same rule applies to every task in the catalog.

What this site does not do

It does not forecast what AI will be able to do in the future. It does not infer your level from your title; you pick it yourself. And it does not keep a record of who you are. A visit leaves one kind of record, and it is not tied to you: the job, level and industry that were shown, and which level or industry was picked. There is no account, no cookie and no identifier attached to it. If a page fails to load, the site keeps a note of which page broke and what went wrong, so we can fix it; that note carries nothing about you either, not even your location.