The Last 10%: Why 'AI Exposure' Is a Starting Line, Not a Layoff Forecast

Everyone is asking the wrong question about AI and your job. The headline says your role is "90% automatable" — and your stomach drops. But here is what the data actually implies: if a machine does 90% of your work, someone still has to do the last 10%. And that last slice might be the most valuable, most human, and most secure part of your entire career.

That is the argument at the heart of a new essay by economist Joshua Gans of the University of Toronto's Rotman School, published in July 2026 by the Economic Innovation Group. [Fact] Gans takes the now-famous line — that AI could automate 90% of what we do — and turns it from a doom headline into an economic question worth answering.

Technology changes tasks, not job titles

Start with a correction that changes everything. [Claim] Technology rarely deletes a whole occupation in one stroke. It changes who — or what — performs particular activities inside that occupation. Your job is a bundle of tasks, and AI arrives task by task, not title by title.

This is why "Is my job safe?" is almost the wrong question. [Fact] Gans, building on the task-based framework of economists Daron Acemoglu and Pascal Restrepo, argues the useful question is narrower: which tasks get automated, and what happens to the ones left standing? Two workers with the same job title can face completely different futures depending on how the automatable tasks are distributed across their week.

Consider an accountant. If AI absorbs the routine bookkeeping and reconciliation, what remains is exception-handling, judgment calls, and client conversations — the expert 10%. If AI instead took the client-relationship work and left the data entry, the story would be reversed. Same title, opposite outcome.

The same logic reshapes a paralegal's day. Automate the document search and citation-checking, and what remains is case strategy, client intake, and the judgment about what actually matters — work that is harder, not easier, to hand to a machine. A high exposure number, in other words, tells you nothing about which of these two futures you are walking into.

When 90% is gone, who does the last 10%?

Here is the pivot. [Fact] Gans reframes the 90% claim as a genuine question: under what conditions does the final 10% become more valuable, versus simply meaning fewer people are needed?

There are at least four reasons the last 10% can expand rather than vanish.

Focus effects. [Estimate] When routine work disappears, freed attention flows to the hard parts. A radiologist relieved of routine triage can spend more time on the genuinely difficult, ambiguous cases — the ones where a missed call changes a life. The remaining work does not just survive; it gets better.

Expertise recalibration. [Claim] Whether automation removes routine tasks or expert tasks decides everything. Strip out the routine, and the expertise bar rises — fewer workers, but higher-paid and harder to replace. Strip out the expert bottleneck, and the bar falls, opening the field to more people. The direction is not predetermined.

Organizational redesign. [Estimate] Firms rebundle. Surviving tasks get recombined into broader roles, or split off into new specialist positions — think of "AI model governance" as an emerging slice of clinical and legal work that simply did not exist as a job before.

Demand elasticity. [Fact] When a service gets cheaper and better, people buy more of it. Bank tellers are the classic case: ATMs automated the routine cash-handling, yet teller employment did not collapse — banks opened more branches and reshaped the role. Lower cost expanded demand, and demand expanded work.

But Gans is honest about the other path. [Fact] A lawyer whose document review is 90% automated might simply end up with less legal work to do — and firms needing fewer lawyers to do it. The last 10% only protects you if demand, bundling, and expertise line up in your favor. The outcome is genuinely uncertain, and that uncertainty is the point.

An exposure score is a starting line, not a forecast

This is the most useful idea in the essay for anyone reading a scary statistic about their profession. [Fact] An AI "exposure" or "applicability" score tells you where a technology can touch a job. It does not tell you how many people will be employed, at what wage, doing what.

[Claim] Converting an exposure number into an actual employment outcome requires investigating things payroll data hides:

• Are the exposed tasks peripheral, expert, or genuine bottlenecks?
• Does automation raise or lower the barrier to entry?
• Will freed attention improve the quality of what is left?
• Will firms rebundle tasks into broad roles or narrow specialties?
• Will cheaper, better service expand demand enough to keep hiring?

[Estimate] Gans warns that average exposure can hide a job's architecture. Two occupations with identical exposure scores can have completely different fates because the exposure sits in different places in the task bundle. And payroll records miss the quiet shifts: a software developer whose AI assistant cuts documentation time may show no wage change at all, while the actual texture, intensity, and composition of the work has been transformed.

His conclusion is one worth taping to your monitor: "An exposure measure is a starting point for these questions, not the answer."

What this means for your career

If you take one thing from Gans, take this: a high exposure score is an invitation to investigate, not a sentence to accept.

Look at your own week as a bundle of tasks. Which parts are routine enough for AI to take? What would be left? Is the remainder the expert 10% that gets more valuable when the routine falls away — the judgment, the relationships, the ambiguous calls — or is it work that also shrinks when volume drops? That honest audit tells you more than any headline percentage ever will.

And notice the timing hidden in all this. [Claim] The reskilling window opens before the layoff window. The task-by-task nature of AI means change arrives gradually, one activity at a time — which is precisely the room you have to move toward the expert 10% while it is still forming. On our occupation pages you can see which specific tasks in your field carry the highest AI exposure — and that map is exactly where the last-10% strategy begins.

The number on the chart is not your destiny. It is the first line of a question only you can finish answering — with courage rather than dread.

Sources

• Joshua Gans, "The Last Ten Per Cent: How task-level AI exposure fits into a jobs forecast," Economic Innovation Group, July 2026. Working paper (PDF) · Analysis: agglomerations.eig.org/p/the-last-ten-per-cent

Update History

• 2026-07-21 — Initial publication. Summary and analysis of Joshua Gans's task-level exposure framework and the "last 10%" thesis.

---

AI-assisted analysis. This article was drafted with AI assistance and reviewed by a human editor. The 90%/10% framing and the six investigative dimensions are drawn from Joshua Gans's July 2026 essay for the Economic Innovation Group; the interpretation for workers is our own.