Does a high AI exposure score mean a role should be cut?

No. Exposure measures how much of a role's task mix could be affected by AI — it says nothing about whether the right response is elimination, augmentation, or growth. In practice, high-exposure roles held by highly connected people are often the ones an organization can least afford to lose; that's why exposure should always be read alongside network position and business criticality, never alone.

What is an AI exposure score, technically?

The published research approach (Eloundou et al., 2023) decomposes occupations into tasks using the U.S. Department of Labor's O*NET database, rates each task for whether large language models could materially reduce the time it takes, and rolls the task ratings up to an occupation-level score. OrgTool ships seed scores indexed by O*NET SOC code following this method, computed locally, and lets organizations override them with their own assessments.

Where should a team start?

With measurement, not org charts: score the current org's roles for exposure, look at where exposure concentrates by function and layer, and only then model redesign scenarios. Restructuring before measuring is how organizations cut the wrong roles and keep the wrong work.

AI restructuring

How AI restructuring works: from exposure scores to a defensible plan

OrgTool Learn · 9 min read · Updated July 2026

Every major reorganization in 2026 is, one way or another, an AI reorganization. Boards are asking the same two questions: where does AI change the work, and what should the organization look like afterward? Most companies answer with intuition — a leadership offsite, a target headcount, a spreadsheet. This article describes the measurable version of that process: how AI restructuring works when it's done as analysis rather than assertion.

Step 1 — Measure exposure at the task level, not the job level

The foundational insight of the research literature is that AI doesn't affect jobs; it affects tasks. The landmark study here is Eloundou, Manning, Mishkin & Rock (2023), "GPTs are GPTs", which decomposed U.S. occupations into their constituent tasks using the Department of Labor's O*NET database and rated each task for whether large language models could materially reduce the time required to complete it at equivalent quality. The headline finding: around 80% of the U.S. workforce could see at least 10% of their tasks affected, while roughly a fifth could see half or more of their tasks affected.

The practical consequence: an "AI exposure score" for a role is a roll-up of task-level ratings — not a judgment about the person, and not a prediction of layoffs. Two roles with the same title can carry very different real exposure depending on how the work is actually distributed. Good practice is to start from published occupation-level baselines (indexed by O*NET SOC code) and then let the people closest to the work adjust them.

Step 2 — Separate exposure from response

Exposure is a measurement; the response is a design decision. For any high-exposure role there are at least four distinct responses, and conflating them is the single most common failure of AI restructuring:

Automate — the task volume genuinely disappears into tooling; capacity is reduced or redeployed.
Augment — the person keeps the role, AI absorbs the routine share, and the role's output expectation rises.
Elevate — the routine work goes, and the role is redefined upward around judgment, relationships, or oversight of AI-produced work.
Reskill — the current work contracts faster than the person can grow within it, and a deliberate pathway to adjacent work is the plan.

An honest AI restructuring plan tags every affected role with one of these responses, with a named owner and a date — not a blanket percentage cut spread evenly across departments.

Step 3 — Redesign the work before the org chart

The most common sequencing error is drawing the future org chart first and rationalizing the work afterward. The defensible sequence is the reverse: once you know which tasks change, you know which roles change shape; once you know which roles change shape, spans and layers follow. A service function whose routine tier is heavily automated doesn't just get smaller — its remaining work is typically more escalation-heavy, which argues for narrower spans and more senior ICs, not a thinner copy of the old pyramid. Delayering decisions made without this work-level analysis are cost exercises wearing a transformation costume.

Step 4 — Model scenarios, not edicts

Serious restructuring is comparative: at minimum a baseline (today), a conservative case, and an aggressive case, with headcount, cost, span-of-layer geometry, and exposure distribution computed for each. The discipline that matters is determinism — the numbers a scenario produced in March must be reproducible in September, when someone asks why a decision was made. If your analysis lives in a spreadsheet that six people have edited forty times, you don't have a record; you have an argument waiting to happen.

Step 5 — Check the network before you cut

The org chart shows reporting lines; it does not show how work actually flows. Organizational network analysis (covered in depth in our ONA guide) matters doubly in AI restructuring because the two lenses compound: a person with high AI exposure and high network centrality is simultaneously the role most likely to change and the person most expensive to lose. Reduction plans that ignore this reliably sever the informal connections that made the old organization work, and then attribute the resulting slowdown to "change fatigue."

Step 6 — Make the selection defensible

When an AI restructuring reaches actual reductions, it becomes a legal event. Selection criteria — however strategically motivated — produce statistical patterns across protected groups, and those patterns are testable: the EEOC's four-fifths rule as the primary screen, Fisher's exact test for significance, WARN Act notice thresholds by jurisdiction. The time to run those tests is while drawing the selection, not after the list is final. (Both are covered in this library: the 80% rule, explained and the WARN Act working guide.) A plan whose strategic story and legal record are the same document is a plan that survives scrutiny.

Key takeaways

Measure at the task level; score roles by roll-up, starting from published O*NET-based baselines.
Exposure ≠ elimination: tag every affected role automate / augment / elevate / reskill, with an owner.
Redesign the work, then the roles, then the structure — in that order.
Model at least three scenarios and insist on reproducible numbers.
Cross exposure with network centrality before anyone's name is on a list.
Run adverse-impact and WARN analysis during selection, not after.