Who's Exposed AND Stuck? Reading Manning & Aguirre Against the Manx Workforce

One axis isn't enough

If you've read any AI-impact analysis in the last three years you'll have seen the headline shape: a list of occupations ranked by some exposure score — Frey-Osborne, the Anthropic Economic Index, Felten-Raj-Seamans — with the high-exposure roles flagged as "at risk". It's a useful screen. It's also a partial picture, and on its own it overstates the problem in a particular way.

A heavily-exposed worker who has a degree, a mobile skill set, lots of training pathways and a current of demand pulling them into adjacent occupations is probably going to transition fine. A less-exposed worker with none of those things may be in more trouble than the headline numbers suggest. Exposure tells you what AI can do; it doesn't tell you what the person facing AI can do back.

Manning & Aguirre published NBER w34705 in January 2026 making exactly this point with US data. The paper crosses two questions on a single chart: an exposure axis (how much of the role is automatable) and an adaptive capacity index (how easily the worker — at occupation level — can move). Their finding: of the 37 million Americans in the top exposure quartile, about 26 million also score above-median on adaptive capacity. They'll likely be fine. The remaining ~6 million sit in what they call the vulnerable cell: high exposure, low capacity. About 4.2% of their sample, concentrated in clerical and administrative work. That's the population the policy money should follow.

We've adapted the framework for the Isle of Man. The page is now live at /education/adaptive-capacity. This post explains the logic — what's the same as Manning & Aguirre, what's different, and what the numbers mean for the Manx workforce.

The four-cell map

Every UCM-served Manx occupation lands in one of four boxes:

The vulnerable cell is the policy hotspot. Workers in mobile and overqualified cells will mostly self-rebalance via the labour market. Sheltered cells are quiet — they may eventually face exposure but they aren't in the front line. The vulnerable cell is the bottom-right corner of the chart, the cluster of dots that have nowhere obvious to go.

How we score the two axes

Exposure (0–100) is the same composite we already use across the site — on Demand vs AI Risk, the Shortages page and per-occupation cards on /occupations. It's a weighted blend:

• 55% Frey-Osborne probability of computerisation (the 2013 classic, still the sharpest displacement signal)
• 25% Anthropic Economic Index automation share (what AI can replace)
• 20% inverse augmentation share (what AI can amplify, flipped — high augmentation pulls the score down because amplified work is generally safer)

Same threshold across the site: ≥ 50 = high exposure. So a SOC scoring high here is the same SOC the rest of the platform flags as exposed. Consistency isn't accidental — readers shouldn't have to remember a different cutoff per page.

Adaptive capacity (0–100) is where we depart most clearly from Manning & Aguirre. They build their capacity index from O*NET worker characteristics — educational attainment, age distribution, occupational mobility — at the individual level. We don't have that for the Isle of Man at SOC granularity. So we use UCM pathway data instead, which captures a related but different signal: the ability to retrain into a role, not the ability of incumbent workers to transition out of one.

The five components, weighted-averaged:

• 30% Pathway count — how many UCM parent courses lead toward this SOC. More pathways = more retraining options.
• 20% Annual graduate supply — estimated yearly UCM throughput reaching this SOC. The size of the supply tap.
• 20% Median wage — used as a mobility proxy. Higher-paid roles tend to have more transferable skills (this is from the Manning & Aguirre playbook indirectly).
• 15% Training lag, inverted — fastest qualifying UCM route. Short = high capacity contribution.
• 15% Field diversity — how many distinct occupational fields can route into this SOC. More diverse pathways = broader career-change optionality.

Each component is normalised 0–100 against the IoM observed range, then weighted-averaged. The cell threshold is ≥ 50 for high capacity, same shape as the exposure threshold.

This is a Manx approximation of Manning & Aguirre's framework, not a clean replication. The vulnerable cell on the IoM page should be read as "exposed roles where the retraining infrastructure is thin" rather than "exposed workers who personally can't adapt". A cell-by-cell shortfall in our scoring is a UCM catalogue gap; in Manning & Aguirre's it's a worker characteristic gap. Different problems with different policy responses.

The headline number — and why it isn't what it looks like

The Manx page reports 32.1% of UCM-mapped workers in the vulnerable cell. Manning & Aguirre's US benchmark is 4.2%. That eight-fold gap looks alarming. It's also not directly comparable, and the page is honest about that.

Three reasons the percentages don't line up:

1. Different population. Manning & Aguirre score every American worker in the top exposure quartile. We score the IoM SOCs that UCM offers training for — about 11,800 census workers across 66 occupations, not the whole 49,000-strong working population. The UCM-served subset is by design tilted toward the exposed clerical/admin/finance roles UCM teaches into. If we'd scored the whole Island, the denominator would be different and the share would shift.

2. Different capacity proxy. Their capacity score reflects worker-level mobility (degree, age, prior transitions). Ours reflects system-level pathway depth. The Manx UCM-served workforce skews toward exposed roles where the retraining catalogue is sparse for those exact roles — a structural feature of how UCM grew up around historical Island industries, not a measure of individual worker frailty. So our "low capacity" finding is, mechanically, partly a tautology: UCM teaches certain things; the things UCM teaches are clerical-heavy; clerical roles are exposed; the cell sits high.

3. Different sampling. Their dataset is a national survey with adaptive weighting. Ours is a Manx census slice with administrative data. Different sampling generates different distributions even if the underlying phenomenon is identical.

What the gap does signal — robustly — is that the Manx UCM-served workforce skews toward exposed roles, and the retraining pathway counts we use to score capacity are sparse for many of those exact roles. Read the page as a hotspots map, not a population census.

Plumbers don't solve the gap in 0 months

The single most useful insight on the page isn't the cell map. It's the lag column.

Adaptive capacity isn't a switch you flip. A vulnerable cell with "low capacity" doesn't become high-capacity the moment a new course launches. The lag column on the table is the fastest qualifying UCM route — the minimum elapsed time between someone starting and them being able to practise the role independently. Plumbing is 36–48 months. Nursing is 36 months. An MBA top-up is 18. Even a short professional certification has its own 6–12 month tail.

Translation: even if every recommendation on the page were funded tomorrow, the supply response is years out. Treat the vulnerable cell as policy debt that will mature slowly, not a problem you can rebudget out of next quarter.

The response-delay timeline visualises this: a curve per cell showing what fraction of the cell workforce has finished a qualifying UCM route by month 1, 3, 6, 12, 18, 24, 36 and beyond. The vulnerable curve is flat across the early months and only really starts climbing after month 18. That flatness is the policy reality.

Years to retrain — the bombshell hiding in the data

The lag column tells you when the first cohort lands. It doesn't tell you how many years it would take to absorb the whole cell at current UCM intake. We added that calculation as a column on the page in a follow-up: Yrs to absorb = censusWorkers ÷ annual graduate output.

A worked example. SOC 2421 — Chartered & certified accountants. About 1,019 IoM workers. UCM annual graduate output reaching this SOC: ~32. Years to retrain the whole population at full intake: 31.8 years.

That's not a misprint. It's not even the worst number on the table. The supply tap, even at its existing size, doesn't physically scale to "retrain a 1,000-person occupation in a meaningful policy window".

The headline strip above the response-delay chart breaks this down per cell — and adds a second, more brutal cut: the share of cell workers whose SOC has zero UCM pathway at all. Those workers aren't in the years-to-absorb number. They're in a different problem altogether — not "slow supply", but "no supply tap exists". On the Manx vulnerable cell, that share is non-trivial, and the policy response is fundamentally different. You can fund cohort scaling in occupations with a pathway. You have to fund catalogue addition before you can do anything for occupations without one.

A Manx-grounded skill-transferability signal

When this page first launched, our adaptive-capacity index was missing Manning & Aguirre's most distinctive component — their task-overlap signal. They ask, for every occupation, "how many adjacent occupations share enough core tasks that an incumbent could reasonably move?" It's derived from O*NET task profiles, which are US-specific.

We've since added a Manx-grounded analogue, on a follow-up commit. Instead of US O*NET, we use the union of skill + knowledge tags observed across IoM job listings — there's a per-job enrichment table populated by the platform, and aggregating it by SOC gives a vocabulary of skills + knowledge for each occupation. We then compute pairwise Jaccard similarity between every SOC's vocabulary; ≥ 20% overlap counts as "skill-adjacent". The transferability score is the count of OTHER SOCs above that threshold.

This is observed Manx work, not theoretical task maps. Some of what comes out of it:

• Most transferable: Construction project managers (129 adjacent SOCs), Project support officers (117), Electricians (117), Sports coaches (116), Vehicle technicians (102). These are roles whose skill bundles are broadly applicable — generic project management, generic technical, transferable trades.
• Least transferable (with skill data): Nursery education teaching (9), Pest control officers (13), Chefs (19), Primary teaching (20), Finance investment analysts (20). These are specialised, domain-bound skill sets that don't cross over to many other jobs.

The new component takes 15% of the capacity score (the other components shifted to make room: pathwayCount 30→25%, annualSupply 20→15%, lag and fieldDiversity each lost 2.5pp; wage held at 20%). The cell distribution moved slightly — vulnerable held at 32.1%, but some overqualified SOCs slid down to sheltered as the re-weighting reduced their capacity score.

Honest limitation on the new component: SOCs with very few advertised IoM listings show "thin" rather than a real score. A SOC with one or two listings has a vocabulary too small to compute meaningful overlap; we floor at 3 observed skill tags. Specialist trades and niche civil-service roles are most affected — pest control above is borderline-thin. When you see "thin" in the table, that's data sparsity, not "this role is genuinely non-transferable".

This component closes the most visible gap between our framework and Manning & Aguirre's. It doesn't make us a replica — they still have richer worker-level signals (education, age, mobility) we can't access on the Island — but it brings their distinctive geometric question ("how many places can this worker step to?") into the Manx index.

Two sub-populations inside "vulnerable"

That's the framework's most useful refinement for IoM policy. The vulnerable cell isn't homogeneous. It contains two distinct sub-populations:

• Slow-pipeline workers — there's a UCM pathway but the graduate flow is small relative to the workforce. Years-to-absorb is high but finite. Policy response: fund cohort growth in existing programmes; co-fund employer apprenticeship scaling.
• No-pathway workers — the SOC has zero UCM courses targeting it. Years-to-absorb is undefined. Policy response: catalogue addition. Until UCM (or another provider) launches a programme, no amount of intake-funding does anything for these workers.

The page surfaces this distinction in two places: the per-row "no pathway" indicator in the absorption-headline strip, and the click-to-expand drill-down on each table row, which shows the actual UCM courses leading toward each SOC. Empty drill-down = no-pathway problem. Full drill-down with high years-to-absorb = slow-pipeline problem.

Honest limitations

We're up-front about what the page can and can't do.

• It's not a clean Manning & Aguirre replication. Their O*NET-derived capacity index measures something we don't have for the Island. Our proxy is closer to "system retraining-into-role capacity" than their "individual worker adaptability".
• The four-cell split is a deliberate simplification. SOCs near the boundaries shift cells if you change either threshold by 5 points. Don't read individual borderline assignments as fixed.
• It's not modelling endogenous behaviour. Real-world reskilling rates are a fraction of "everyone retrains the moment a course exists". The years-to-absorb numbers are an absolute upper bound on supply response, not a forecast.
• Pathway-mapping is keyword-based. The link between a UCM course and a destination SOC is built from a specific-before-general keyword mapper. It's right for most trades and concrete professional roles; it picks the most-common matching role for broad umbrella courses. The drill-down on each row lets you eyeball whether the matches make sense.
• We don't have IoM-specific worker characteristics at SOC level. If we did — IoM census records of educational attainment by occupation, occupational mobility data — we'd build a closer Manning & Aguirre replica. As of April 2026 the data isn't there. The skill-transferability component (added 1 May 2026) closes the most visible gap, but it's still a job-listings proxy rather than a worker-properties measure.

What the page is for

Three audiences, three uses:

• UCM curriculum committee — the no-pathway list is the catalogue-gap shopping list. The high-years-to-absorb list is the cohort-scaling shopping list. Both can drive 2026/27 programme decisions.
• Treasury / DfE workforce planners — the response-delay timeline is the multi-year planning horizon. Reskilling money committed in 2026 affects 2028–2031 labour-market outcomes. The page makes that timeline tangible.
• Manx residents who are, say, mid-career chartered accountants — the per-row drill-down tells you which specific UCM courses lead toward your destination, how long they take, and what the AI exposure of that destination looks like. That's a personal-decision tool, not just a policy one.

Cross-references to walk next:

• /education/courses — every UCM parent course (modules suppressed) with length, primary career destination, AI risk and adaptive-capacity cell badge per card.
• /education/transition-era — the policy argument these numbers support.
• /education/projections — the 20-year simulation that turns this static cell map into three dynamic scenarios with Personal Income consequences.
• /education/shortages — the demand-side companion to this page's supply-side framing.

The short version

Manning & Aguirre's contribution is a sharp one: AI exposure × adaptive capacity beats AI exposure alone. Their headline number — 4.2% of US workers in the vulnerable cell — is conservative because their capacity index is reasonable. We've adapted the framework for the Isle of Man with a different capacity proxy and a different population, so our 32.1% headline isn't directly comparable; it's a hotspots map for the UCM-served workforce, not a census-wide claim.

What survives the adaptation cleanly is the structural insight: exposure alone overstates the problem; capacity alone ignores it; both together identify the population the policy money should follow. The Manx version of that population is the bottom-right corner of /education/adaptive-capacity, with the years-to-absorb numbers and the no-pathway flags telling you what kind of help each SOC actually needs.

The numbers will look different on the Island than in Manning & Aguirre's US data. The framework is the same. The policy reasoning it enables is the same. Read their paper. Then read the page.

— Manning, A., & Aguirre, J. (January 2026). "How Adaptable Are American Workers to AI-Induced Job Displacement?" NBER Working Paper w34705. https://www.nber.org/papers/w34705.