The Average AI Risk Hides Who Carries the Loss

The slide is still on the wall after the all-staff briefing ends.

Our AI risk: 25%.

The number hangs above a rural office, a city office, and the names that do not appear on the slide: contractors, temporary workers, people on reduced hours, people working remotely, people whose accommodation is tied to the old workflow.

Do not accept the average as your outlook.

Write it down.

Then take it apart before it starts choosing who is expected to absorb the change.

What the Average Hides

Start with the ILO’s global index.

It places 1 in 4 workers in occupations with some generative-AI exposure.

It puts 3.3% of global employment in its highest exposure gradient, the category intended to identify the greatest potential for automation of occupational tasks.

Keep both figures.

Do not let the first impersonate the second.

Some exposure is broad contact with a technology.

The highest gradient is a narrower classification of technical potential.

Neither is a redundancy count.

Neither tells you whether your employer deploys a system, cuts hours, expands output, stops hiring, or turns routine work into review work.

Exposure tells you where to look.

It does not tell you who carries the loss once the organization looks.

That is what the average on the wall is doing.

It makes a distribution look like one shared weather event.

The rural office can face a different labour market from the city office before the tool is switched on.

The assistant whose entry tasks disappear can face a different future from the senior employee whose task becomes review authority.

The contractor can be outside the denominator altogether.

The percentage has not erased those conditions.

It has hidden them.

The Gradient

Do not turn a group pattern into a personal forecast.

In high-income countries, the ILO places 9.6% of female employment and 3.5% of male employment in its highest exposure gradient.

That is a classification pattern shaped by occupational segregation, including clerical, administrative, and business-support work.

It is not a claim that a model is technically better at replacing women than men.

It is not a distribution of gains.

The same deployment can reduce administrative tasks, raise output targets, intensify monitoring, create an advanced-user role, or move supervisory work to another occupational stream.

Ask who receives the new tool access.

Ask who receives paid training, pilot roles, promotion, and a share of the productivity gain.

Separate ILO research across 84 countries found female-dominated occupations almost twice as likely to be exposed as male-dominated ones: 29% against 16%.

That is a reason to inspect occupational ladders and selection criteria.

It is not permission to write off a demographic group.

Country structure changes the result too.

The IMF’s six-country comparison found women more exposed in almost all cases, but India was an exception, because women’s employment was more concentrated in agriculture.

Do not repeat a gender claim without its country, technology, occupational mix, and definition of exposure.

The same discipline applies to age.

The IMF found no straightforward association between age and AI exposure in its six-country sample.

In Japan, OECD evidence on conventional automation instead found workers aged 16–34 in jobs at substantially higher risk than older workers.

Different technology.

Different country.

Different entry routes into work.

The average is not neutral because it has no detail.

It is incomplete because detail is where the cost lands.

Complementarity

Write the word complementarity beside exposure.

It means the work can become more productive with the technology, because the worker still supplies judgment, authority, relationships, context, or accountability.

It does not mean safe.

It does not mean well paid.

It names a different possible relationship to the same tool.

The IMF’s UK comparison makes the distinction visible.

More than 50% of workers with a college education or higher were in high-exposure, high-complementarity jobs.

The corresponding figure for workers with middle-school education or below was 17%.

That does not prove a degree buys security.

It shows that access to the version of work that uses a tool can be distributed differently from exposure to the tool.

In the same comparison, high-exposure, low-complementarity work ran from 26% for the lowest education group to 36% for the college group.

The difference was under 10 percentage points.

Read the two findings together.

Task disruption can be widely spread.

The route from disruption into better-paid complementary work can be concentrated.

Do not ask only whether the tool can do this.

Ask who gets to approve it, correct it, explain it, sell it, and carry responsibility when it fails.

Those are not decorative tasks.

They are where an organization concentrates authority, pay, and a path forward.

Where the Office Is

Put the rural office and the city office in separate cells.

European data put about 4% of city workers at high automation risk, compared with 7.5% in towns and semi-dense areas and 9% in rural areas.

The OECD attributes much of that gap to education and occupational structure.

It also states that this measure does not incorporate recent generative-AI developments that may affect highly educated city workers.

So do not turn the 9% rural figure into a claim that every rural worker is more exposed to every new system.

Use it to insist on a local account.

Averages inside a single province vary as widely, from one metropolitan labour market to a rural one.

That is why a company-wide figure cannot describe the occupational composition of every site.

Now inspect adjustment capacity.

A small manufacturing town can have a similar exposure rate to a city district and still have fewer employers, longer travel distances, thinner training provision, and less ability to absorb displaced workers.

A high-exposure city occupation can sit inside a thick labour market with many adjacent employers.

The exposure rate is the warning light.

The number of places you can sell the next version of your work determines how hard it is to leave the old one.

Do not call that difference resilience.

Call it labour-market structure.

The Six Columns

Take out a sheet of paper.

Make six columns.

Occupation. Write the actual work group, not a morale-boosting title.

Pay band. Write the minimum-to-maximum base-pay range, the midpoint if you know it, and whether bonuses, overtime, or stock are carrying the total.

Education or license. Write the credential, authorization, or demonstrated skill that gates entry to the next version of the work.

Location. Write the site, the commuting territory, and whether plausible employers or paid training exist nearby.

Task exposure. Write lower, middle, or higher. Use the real task bundle, including inputs, exceptions, checking, customer recovery, and work created after automation.

Likely complementarity. Write lower, middle, or higher. Mark whether the group controls judgment, relationships, approvals, tools, or a route into the redesigned role.

Make one row for the rural office.

Make another for the city office.

Make another for the workers missing from the slide’s denominator.

Do not compress them back into an average when the sheet is full.

Then mark bargaining power beside the six columns.

Low bargaining power can mean a thin local market, contingent status, a low pay band, no paid release for training, no recognized credential, or a rule that keeps you out of the redesign meeting.

Do not treat that list as a character test.

It is a list of conditions that make it cheaper for an organization to make one group absorb the change than another.

Any group with high exposure and low bargaining power needs a separate transition plan.

That plan needs a local vacancy check, paid and accessible training, a path into complementary work, income protection, and a record of who receives the new roles.

An all-staff average cannot do that job.

The Pay Band

Now put wages on the sheet.

The usual company answer is that everyone benefits from productivity.

Do not accept that sentence without a distribution.

In the second quarter of 2026, management and professional occupations had a $1,697 median weekly wage.

Production, transportation, and material-moving occupations had a $968 median.

That gap is between occupation groups, before industry, employer, benefits, location, and rank divide the people inside them further.

An identical task redesign does not arrive at an identical pay position.

Across all U.S. full-time workers in the same quarter, the first quartile was $850 and the third was $1,915.

That $1,065 spread is most of the median itself.

The national median is not your household plan.

Nor is an occupation title a pay position.

Computer systems analysts in May 2025 had a median annual wage near $105,000 in insurance carriers and near $108,000 in computer-systems design.

Across industries and percentiles the same detailed occupation ran from about $66,000 at the 10th percentile in colleges and universities to about $175,000 at the 90th in aerospace manufacturing.

One title.

Different industry.

Different rank.

Different room to absorb a pay cut, reject a bad offer, or wait for a better one.

Ask whether a proposed role is near the 25th percentile, the median, or the 75th for that occupation in your area.

Ask what happens to health coverage, pension contributions, commute, hours, security, and progression.

Do not let a base salary replace the compensation account.

Benefits and wages do not move together, and in real terms one can rise while the other falls in the same year.

Cash, benefits, prices, and bargaining position are four numbers.

The consequence of a task change is not one number either.

The Bottom Cell

Inspect who can take the training offered after the announcement.

An open course is not a transition plan.

In an OECD example from Lithuania, 51% of workers in low-automation-risk jobs participated in training, against 22% in high-risk jobs.

That is what happens when time, pay, equipment, accessible delivery, employer recognition, health, care, and travel are treated as somebody else’s problem.

Count enrolment if you need to.

Then count paid release time, completion, credentials, placement, earnings, and the people who could not take part.

Course seats are not transition results.

Inspect accessibility before the workflow is locked in.

The U.S. Department of Labor’s AI and Inclusive Hiring Framework sets out 10 focus areas for employers assessing, acquiring, or deploying AI hiring technology.

Use that procedural standard to ask whether the system works with screen readers and captioning, whether pace and monitoring accommodate variable performance, whether alternatives exist, and whether an automated decision can be challenged by a person.

Across European OECD countries, an estimated 16% of jobs held by employed people with disability could disappear through automation over the next 15 to 20 years, and a further 32% could change substantially.

That finding is tied to concentration in monotonous and repetitive tasks, not to disability as an inherent feature of work ability.

The same deployment can reduce a task and remove access to the job that remains.

Keep those outcomes in separate columns.

If the company says an average adjustment is fair, ask where the adjustment lands.

A 5% average adjustment can leave the bottom cohort unchanged if the money goes to people already near the top.

The average did not make the loss equal.

It reported the sum after the unequal allocation.

Study the allocation.

And do not buy a course tonight because a shared number on a wall has been made to feel like your personal deadline.

Study the bottom cell first.

Minimum Viable Action

Before you sleep tonight, write your employer’s average number at the top of a sheet.

Write your pay band, location, and task mix beside it.

Circle the information the number omitted.

Then write one sentence: This average describes ____; it hides _____.