AI Job Exposure: What a 78-Role Model Can Tell Us

The most popular question about AI and work is also the least useful one: which jobs will disappear? A job is not a single unit of production waiting to be automated. It is a negotiated bundle of tasks, skills, tools, relationships, permissions, and accountability. AI can make one part of that bundle abundant, increase the value of another, and leave a third almost untouched.

That is what our analysis of 78 professional roles is designed to show. It does not forecast redundancies. It asks a narrower question: given the skills that characterise a role, where does today's AI appear to reduce scarcity, where does it increase the holder's leverage, and where is the answer still mixed or durable?

Within this model, the result challenges the easy story that technology roles are safe because they build the technology. The sharpest dividing line runs through software engineering itself. Routine implementation and well-specified digital output face downward pressure; architecture, applied modelling, product judgement, and system-level decisions gain leverage. The hypothesis behind the map is that AI reprices tasks and skills inside roles before it eliminates occupations as coherent blocks. The map does not observe market prices; it proposes where that pressure may appear.

This article interprets those findings and explains how an organisation can use them without turning a research signal into a workforce verdict. The exposure map is the separate interactive tool: it contains the complete, searchable ranking and the four-part profile for every role.

What the score measures — and what it does not

The public file analysed here is role_exposure.json, version 2026-06-28. It describes itself as an indicative assessment, not measurement, and contains no population figures. The related role taxonomy groups roles by shared skill patterns, with selected strategic gaps deepened editorially; its role names and skills are also described as indicative.

The public artefacts do not include the underlying source records, the per-skill labels and weights, the annotation protocol, or calibration evidence sufficient to rebuild the upstream mapping independently. We can reproduce the net-pressure arithmetic and ranking from the four published shares; we cannot independently recreate those shares. The map should therefore be treated as an editorial exposure proxy, not a validated occupational benchmark.

At a high level, the project notes say the model starts with a skill profile for each role. Skills are weighted by their prominence in that profile, so a prominent skill contributes more than a peripheral one. Each skill is then assigned one of four directional judgements about the current AI frontier:

Commoditize: AI can produce a substantial part of the output cheaply, quickly, and at broadly acceptable quality. Scarcity moves away from performing the task itself.
Amplify: AI makes a capable practitioner more effective, but the value still sits in their judgement, framing, integration, or accountability.
Mixed: the skill contains material elements of both. AI may automate execution while making review, exception handling, or specification more important.
Durable: current AI changes little of the skill's practical value, often because the work is physical, contextual, relational, regulated, or responsibility-bearing.

We aggregate those weighted judgements into four shares for each role. Net pressure is the commoditize share minus the amplify share. A positive number means more of the profile leans towards commoditisation than amplification; a negative number means the reverse.

That subtraction is intentionally simple. It makes the direction legible, but it does not make the number a probability, a wage forecast, or a percentage of a job that will vanish. The mixed and durable shares remain visible because two roles can have the same net score for very different reasons. Rounding may also make the displayed shares differ slightly from 100. The ranking is useful for comparison and investigation; a one-point difference is not a meaningful claim of precision.

In the current dataset, 53 of 78 roles have positive net pressure, 18 have negative net pressure, and 7 are exactly neutral. Positive and negative are directional assessments; they are not forecasts of employment loss or growth.

The unit of analysis is the role profile, not an individual. The dataset contains no population counts, salary data, adoption rates, performance ratings, or evidence about a particular employee. It also reflects the coverage and granularity of the underlying skills. A missing or overly broad skill can change the profile. Those limits are not fine print: they determine which decisions the score may legitimately support.

The most exposed roles share an output shape

Data & Business Process Analyst (+36)
Early Career & General IT Roles (+33)
Translation & Editorial Content (+27)
Office Administrative Assistant (+25)
Software Development Engineer (+22)

The leaders are not united by sector or prestige. They are united by the shape of their output: structured analysis, documents, translations, routine code, tests, and administrative artefacts whose acceptance criteria can often be stated in advance.

Data & Business Process Analyst leads at +36: 45% of its skill mass is classified as commoditising and 10% as amplifying. Translation & Editorial Content sits at +27, and Office Administrative Assistant at +25. This is consistent with what generative systems do well: transform one digital representation into another, create a plausible first version, classify information, and reproduce established patterns at very low marginal cost.

But even the upper end resists a simple automation story. Early Career & General IT Roles scores +33 while 48% of its profile is durable. The high net comes from a large gap between commoditising skills (35%) and amplifying ones (2%), not from a claim that the whole role is automatable. Entry-level work is particularly exposed because many of its traditional learning tasks — drafting, documenting, looking up, configuring, and fixing standard cases — can now be accelerated or bypassed. The durable half does not disappear; the organisational problem is that the route by which a novice learns to perform it may weaken.

This is the first practical implication. Exposure can appear not only as job loss, but as a broken apprenticeship ladder. If AI produces the first draft, the junior may ship more while receiving fewer opportunities to build the judgement previously acquired by making and correcting that draft. Workforce planning must therefore ask who performs the remaining work in three years, not merely how many hours the tool saves this quarter.

The most amplified roles retain the right to decide

Product Manager (-41)
Applied Machine Learning Engineer (-21)
Software Systems Design & Implementation (-20)
Operations & Delivery Management (-13)
Full Stack Software Developer (-13)

At the other end, Product Manager has the strongest negative pressure at −41: only 1% of its skill mass is commoditising, compared with 43% amplifying. AI can summarise interviews, generate options, draft specifications, and analyse feedback. It cannot own the trade-off between customers, technical constraints, commercial timing, and the cost of being wrong. The faster option generation becomes, the more valuable disciplined selection can become.

The same pattern appears in Software Systems Design & Implementation (−20) and Applied Machine Learning Engineer (−21). Both use AI-intensive tools. Neither derives its value primarily from producing an isolated artefact to a complete specification. Their work includes choosing boundaries, designing evaluation, reconciling constraints, diagnosing interactions, and accepting consequences when a system fails.

This does not make management or architecture inherently safe. Vague co-ordination, status reporting, and generic documentation can be commoditised too. The point is more specific: where a role holds real decision rights and integrates ambiguous evidence, AI tends to increase the number of options it can consider. Where a role mainly converts a specification into a standard output, AI competes more directly with the production step.

The line runs through software engineering

“Is it a coding job?” is therefore the wrong diagnostic. In the same broad family, Software Development Engineer scores +22 and Software Development Engineer in Test +19, while Full Stack Software Developer scores −13, Application Software Developer −8, Software Systems Design & Implementation −20, and Applied Machine Learning Engineer −21.

The spread between the most exposed and most amplified roles in the “writing and shipping code” family is 57 points. That variation is more informative than the family's average. It tells us that a title-level label such as “software” hides the decision-relevant difference.

Why can apparently similar roles land on opposite sides? Their skill mixes are different. A profile dominated by routine implementation, standard testing, documentation, and well-bounded delivery gives current models more substitutable work. A profile weighted towards system design, problem framing, evaluation, stakeholder trade-offs, and applied model decisions gives a capable practitioner more leverage. Language, framework, and IDE are secondary to the structure of the work.

This also supplies a useful counterexample to the claim that the most technical person is always the safest. Technical depth only protects scarcity when it remains tied to a hard problem, a consequential judgement, or a difficult integration. Once expertise has been packaged behind a reliable interface, routine consumption of that expertise can reprice quickly. Conversely, a less technical role may remain durable when it owns trust, access, physical intervention, or a decision that cannot be delegated to a model.

Exposure is not employment loss

An exposure score describes technical proximity between AI capability and part of the work. Employment is the outcome of a much larger system. At least six forces intervene:

Adoption: a capability in a model is not yet a deployed, secure, integrated workflow.
Economics: cheaper output can reduce labour per unit while increasing demand for many more units.
Work redesign: automated tasks may be removed, expanded, recombined, or shifted to another role.
Quality and liability: acceptable drafting quality may still require costly review, accountable sign-off, or recovery from rare failures.
Institutions: regulation, collective agreements, professional standards, and customer expectations affect the pace and form of adoption.
Complementarity: making one task cheap can increase the value of bottlenecks around it — judgement, distribution, proprietary context, physical execution, or trust.

That distinction is also central to the strongest external benchmarks. The ILO–NASK refined global index combines task-level data, worker input, expert review, and AI-assisted scoring across the international occupational classification. Its 2025 analysis places one in four workers in an occupation with some GenAI exposure, but only 3.3% of global employment in the highest exposure gradient. Its principal conclusion is transformation rather than wholesale replacement, because most occupations still contain tasks requiring human input.

The newer OECD AI exposure measure, published in 2026, takes another approach. It maps AI capabilities across nine cognitive, social, and physical domains to occupational requirements and calculates a capability gap. AI is currently closer to routine information processing, administrative, and codifiable work, and further from contextual judgement, interpersonal understanding, complex decisions, and responsibility. The OECD explicitly treats real labour-market effects as contingent on adoption, regulation, organisational change, and social choice.

Our map is not a miniature version of either index. It uses skill-profile weights, a different four-part judgement, and no employment population data. The scores are not directly comparable. The value of comparison is methodological: three different lenses all warn against translating technical exposure directly into headcount loss.

What “repricing” means in practice

We use repricing as an analytical metaphor, not as an observed labour-market result. The dataset contains no wages, vacancy volumes, hiring demand, or transaction prices. It proposes where the relative scarcity of a task or skill may be changing; it does not demonstrate that it has changed, and it does not mean that an employee's salary will fall by the net score.

When a credible first draft moves from four hours to four minutes, the ability to produce that draft becomes less scarce. Value may move to specifying the problem, supplying proprietary context, detecting subtle errors, integrating the output, or taking responsibility for the result. The role may shrink, expand, or simply change its internal composition. The final employment effect depends on volumes, operating design, and who captures the productivity gain.

This is why an apparently “exposed” role can remain important and an “amplified” role can still lose positions. A company may increase its analytical workforce because cheaper analysis unlocks demand. It may reduce product roles during a strategic contraction even though AI amplifies product judgement. Exposure describes a pressure on the work; it does not override the business cycle, corporate strategy, or managerial choice.

Five legitimate enterprise uses

Used with discipline, the map can improve the questions an organisation asks.

Prioritise workflow discovery. Use high exposure to choose where to investigate task changes, then observe the actual work rather than automating the title.
Stress-test role architecture. Find profiles that combine highly commoditising and durable work. They are candidates for redesign, not automatic deletion.
Design adjacent reskilling. Move from exposed production tasks towards amplified work in the same domain, preserving transferable knowledge and identifying real prerequisite gaps.
Protect apprenticeship. Identify the junior tasks that AI may absorb and create alternative ways to acquire review, diagnosis, and decision experience.
Build scenarios. Combine exposure with adoption evidence, demand, cost, risk, attrition, and capacity to compare interventions. Keep the uncertainty visible.

Five uses should be prohibited. Do not use the score to rank employees, set pay, select people for redundancy, infer an individual's proficiency, or promise a headcount saving. Do not compare business units or demographic groups unless their data coverage is known to be comparable. And do not present a portfolio average to executives without showing the mixed and durable shares underneath it.

An exposure ranking used as a redundancy list is not workforce science. It is a category error with a number attached.

Common failure modes

Title panic. Leaders see “software engineer +22” and launch a generic reduction or training programme. The variation inside software is the finding; collapsing it back to the title discards the evidence.

Precision theatre. A team treats +19 as materially safer than +22, models a payroll saving from the difference, or adds exposure scores from incompatible sources. These numbers order hypotheses. They do not price a balance sheet.

The proxy-to-person leap. A role profile is applied to every incumbent, ignoring specialisation, performance, aspiration, context, and demonstrated skill. Role exposure cannot establish individual replaceability.

Static frontier thinking. The judgement is treated as permanent. Model capability, tool integration, price, regulation, and work practices change. Every exposure assessment needs a date, version, owner, and review trigger.

Ignoring the adoption gap. A demonstration is mistaken for a dependable workflow. Security, data access, evaluation, exception handling, and accountability frequently consume more effort than generation itself.

Reskilling by slogan. Everyone in an exposed role is sent to “learn AI”. Useful reskilling starts with adjacent target work, transferable skills, missing prerequisites, and an opportunity to apply the new capability. Training without changed work allocation creates certificates, not transitions.

A decision process that survives contact with real work

Start with one workflow and one accountable owner, not an enterprise-wide automation percentage.

Name the outcome and boundary. What output is produced, for whom, under what quality, time, cost, and risk constraints?
Decompose the work. With jobholders and managers, identify the main task clusters, exceptions, hand-offs, decisions, and permissions. Job descriptions are inputs, not observations.
Map the required skills. Separate production from specification, review, integration, and accountability. Record the evidence and recency needed for each.
Use exposure as a hypothesis. Examine the four shares and the underlying skills. Ask where AI may commoditise, amplify, or leave a bottleneck—not how many people it can replace.
Run a bounded trial. Compare assisted and current workflows on cycle time, quality, rework, exceptions, worker experience, and risk. Set a stop condition before launch.
Choose an intervention. Automate a task, augment the role, redesign the hand-off, reskill towards adjacent work, change demand, or defer. “Deploy AI” is not a decision.
Review the workforce consequence. Track who gains access to higher-value work, who loses learning opportunities, where workload intensifies, and whether the predicted bottleneck actually moved.
Version the conclusion. Record the model and data date, assumptions, dissent, and next review. Outcomes should update the role profile and reskilling plan.

Consider Software Development Engineer in Test at +19. The wrong response is “testing is 19% automatable”. A defensible pilot would separate test generation, fixture creation, execution, failure triage, release-risk judgement, and quality strategy. AI might make test drafting abundant while increasing the value of evaluation design, observability, systems diagnosis, and accountable release decisions. The reskilling destination is not the vague category “AI skills”; it is the adjacent work that the new workflow makes more scarce and important.

That is the purpose of the map. It does not decide whom an organisation no longer needs. It helps reveal which parts of work are losing scarcity, which are gaining leverage, and where a company must redesign learning before productivity gains hollow out its future capability.

If you need to translate AI exposure into a defensible role-redesign, workforce-planning, or reskilling decision, get in touch to discuss the implementation.

Continue exploring

— Skills Intelligence