The State of AI Hiring 2026

Named index: AI Jobs Demand Index / AI Hiring Intensity Index The gap this chapter opens: AI vs the broader labour market (Gap 1) — and it sets up the report's running question, narrative vs revealed behaviour (Gap 5).

Every AI hiring report has to clear the same suspicion first: is this a real labour market, or a headline? We answer it before anything else, because the rest of the report — pay, skills, access, employers — only matters if the demand underneath is genuine. It is. Search demand for AI roles has grown +86% in four years while the broader job market barely moved, and independent trackers built on entirely different data point the same way. But "growing" is not one story. The growth is lopsided — Europe is accelerating faster than the United States, a handful of brand-new roles are exploding while last year's hype roles deflate, and the corpus we measure is seeded on AI keywords, which means we are scrupulous about which growth claims we will and will not make.

1.1 Is AI really growing? Triangulating the signal

Our AI Jobs Demand Index — four years of Google search demand for AI job titles, across 18 markets, normalised to June 2022 = 100 — stands at 186 as of April 2026. That is +86% in four years, in a period when total hiring in most of these markets was flat to slightly up.

One index proves nothing on its own, so we triangulate against trackers that share none of our data:

• Stanford AI Index / Lightcast — AI skills now appear in ~2.5% of US postings, with agentic-AI skills up ~280% year-on-year.
• PwC AI Jobs Barometer (2026) — an AI-skill wage premium of ~62% (up from 57% the prior year) and skills changing more than twice as fast in AI-exposed jobs.

Three methods — our search-demand index, Lightcast's skill share, PwC's wage signal — disagree on the number and agree on the direction and slope. That convergence is the finding: AI hiring is diverging upward from the broader market, not tracking it.

Honesty note — why ours isn't a market-share number. Our corpus is seeded on AI keywords, so we cannot and do not claim "AI = X% of all jobs" — that is a market-share statement only a non-seeded crawl (e.g. Lightcast) can make. We measure growth and composition within AI demand, and triangulate share against those who can measure it. Different question, deliberately.

1.2 AI categories momentum: where the growth actually is

Aggregate growth hides a violent reshuffling underneath. Measured as year-on-year momentum (latest quarter vs the same quarter a year earlier), the AI role market is splitting into a rising and a falling half:

The agentic and infrastructure roles — the production layer of the LLM stack — are the fastest-growing demand in the entire dataset; the roles deflating are last year's poster children (prompt engineering) and the pre-LLM classics (NLP, computer vision as standalone titles). This is the report's central tension in miniature: the market is moving from experimenting with models to operating them in production.

To keep "AI jobs" from blurring into "jobs that mention AI," we classify every role into one of four types and carry the distinction through the whole report:

This chapter — and the demand index — is about AI-native roles. AI mentions are leaking into marketing, HR and finance postings too, but those remain a minority inside their functions and are a different phenomenon from the native AI labour market we track here.

1.3 The AI Hiring Intensity Index

Combining the trajectory above into a single citable series gives the AI Hiring Intensity Index — AI demand normalised to a 2022 baseline, reported at the market level (the demand-share formulation). On this basis intensity has nearly doubled, and it has done so unevenly across geographies — the subject of Chapter 2.

Honesty note. We publish the Intensity Index as a market-level demand-share measure, not a within-corpus posting-share — because our seeded corpus cannot anchor an honest "share of all vacancies." A non-seeded baseline crawl would let us add a posting-share cut in a later edition; until then, posting-share movements are reported as directional within the corpus, never as market share.

aka The Misallocation Map of AI Careers

Named indexes: AI Jobs Demand Index · AI Employer Interest Index · Demand × Supply Gap (flagship) The gaps this chapter closes: demand vs supply (Gap 2) and hype vs hiring (Gap 3).

This is the chapter no one else can write. Most reports see one side of the AI labour market — what employers post, or what a survey panel says. We hold both sides on the same keys: four years of what job seekers search for, paired with what employers actually post. Lay them on top of each other and a misallocation map appears. The most-searched AI job in the world is chased 300 times harder than it is posted; the roles employers are quietly posting in volume barely register in search. Candidates are crowding the wrong door. This chapter shows which door.

2.1 The AI Jobs Demand Index — Europe is catching up

The blended AI Jobs Demand Index is at 186 (June 2022 = 100). Split it by region and the headline is not the average but the divergence: European demand has grown +112% (EU15 index 211.7) against the United States' +63% (162.8).

Inside Europe, the catch-up is led from Germany. On an own-baseline per-country index, Germany is up +118%, ahead of the United Kingdom (+81%) and France (+77%) — the thick-basket markets we can measure reliably.

The story is not "Europe overtakes the US in volume" — the US corpus is still larger — but "European demand is compounding faster," which reframes Europe from a follower market into the report's central wedge (Chapter 6).

Honesty note. Per-country growth uses an own-baseline estimator distinct from the index scopes, so its US figure (+85%) differs from the index's US scope (+63%); we quote the index for the US/EU headline and the per-country series only for intra-Europe ranking. Reliable mainly for US/GB/DE/FR; thinner baskets are excluded. Exhibits: chart_01, chart_07, chart_16.

2.2 The AI Employer Interest Index — Anthropic on top

Turn the search lens from job titles to employers and one name leads. Careers-search demand (3-month smoothed, 18 markets) puts Anthropic at ~170,000 searches/month, 1.88× OpenAI (~91,000) and ahead of NVIDIA (~86,000) — all three built from roughly zero in 2022.

This is the demand-side mirror of the Model-War Scoreboard in Chapter 4 (where we measure the supply side — which vendors appear in job descriptions). That two different instruments — what candidates search vs what employers write — both surface the same frontier labs is exactly the kind of cross-confirmation the report is built on.

Honesty note. Ranking is directional: brands carry different numbers of tracked query variants (Anthropic ~3 vs OpenAI ~21), so this measures careers-search interest, not headcount or hiring volume, and must not be read as role-demand. Exhibits: chart_02, chart_13.

2.3 The Prompt-Engineer hype curve

No role captures hype-vs-reality like prompt engineering. Global search demand peaked in April 2023 at ~168,000/month and has fallen to ~58,000 — −65% from peak — and is still declining (−19% year-on-year).

The deflation matters because prompt engineering became the public face of "AI careers" precisely as the actual market moved past it. It is still the most-searched role (§2.5) — which is the misallocation: peak attention, fading reality.

Honesty note. Google merges head close-variants into a keyword's volume, so the curve is robust as a trajectory even where the absolute level is approximate. Exhibits: chart_03, chart_17.

2.4 Momentum & the birth of AI professions

The roles replacing prompt engineering are visible in the momentum table. Agentic AI Engineer demand is up +360% year-on-year, AI Infrastructure Engineer +192% — the steepest curves in the dataset — while the AI-native core (AI Engineer +52%) compounds on a much larger base.

These are professions being born: the search terms barely existed two years ago. The pattern — agentic and infrastructure roles accelerating as prompt/NLP/CV deflate — is the demand-side signature of the stack moving from prototyping to production, the same shift Chapter 4 finds in the skills employers list.

Exhibits: chart_06, chart_08, chart_09, chart_15.

2.5 The Demand × Supply Gap — the unrepeatable artifact (flagship)

Here is the map. For each role we divide search demand (searches/month, 18 markets) by live supply (active AI postings whose title matches the role). The ratio is searches per posting — how hard a role is chased relative to how often it is offered:

The spread is two orders of magnitude. A prompt-engineer seeker faces ~317 searches per live posting; a generative-AI or AI-research engineer faces ~2 — the postings are practically waiting. Cross the gap with momentum (§2.4) and salary (Chapter 3) and each role earns a verdict:

This Opportunity-vs-Hype-Trap matrix is stronger than any "fastest-growing roles" list because it carries the other axis — supply — and it feeds every Role Card in Chapter 5.

Honesty note. Demand and supply join through a manual keyword↔role crosswalk; the ratio is a leverage proxy, not literal vacancies-per-applicant. Thin-volume roles (NLP) swing more. Exhibits: chart_04, chart_10.

2.6 Search demand as a leading indicator (+ seasonality, CPC)

Two closing signals. Seasonality: AI-job search runs hottest in April (index 135) and coldest in December (68) — a near-2× swing that a hiring team can plan against.

Commercial intensity: the priciest AI-job search clicks — prompt engineering chatgpt at ~$108 CPC, openai roles at ~$43 — mark where hiring budgets are concentrated, a forward signal of where postings follow.

Honesty note. CPC is a commercial-intensity / hiring-budget proxy, not a salary. Seasonality is a calendar effect on search, detrended against a 12-month moving average. Exhibits: chart_05, chart_11, chart_14, chart_18.

Named indexes: AI Pay Premium Index · AI Apply Pay Transparency Tracker (flagship) · AI Skill Price Index The gap this chapter closes: AI vs non-AI (Gap 1) and US vs Europe (Gap 4).

Most compensation reports ask one question: how much does AI pay? We ask a second one that turns out to matter more: what are employers willing to put in writing? Across our corpus the two answers diverge sharply — AI roles carry a real and measurable pay premium, but whether a candidate ever sees a number depends almost entirely on which side of the Atlantic the job is posted. In the United States, more than three-quarters of AI postings state pay; across the European markets we cover, fewer than one in five does. Compensation, in 2026, is not just a level. It is a disclosure decision — and that decision is the new battleground, sharpened by the EU Pay Transparency Directive whose transposition deadline has just passed.

3.1 The AI pay premium — and what's inside it

The premium is real. The median advertised mid-point for an AI role is $185,500, against $144,308 for a comparable non-AI role — an AI pay premium of +28.5%.

That single number, though, hides most of the story, because "AI role" spans an intern labelling data and a staff engineer owning a model platform. Decompose the premium by seniority and it resolves into a clean ladder:

The shape matters more than any one rung. AI compensation is back-loaded onto experience: the step from junior to senior alone is worth roughly +46%, far larger than the headline AI/non-AI gap. A report that quotes only "+28.5%" invites the misreading that adding AI to your résumé adds a third to your salary. It does not. The premium is overwhelmingly a seniority and role-mix effect — AI postings skew senior — layered on top of a smaller true skill premium. We keep the two separate throughout this chapter.

Honesty note. The premium is computed on structured (stated) salaries, which are US-heavy; it is a disclosed-pay comparison, not a population-wide wage estimate. We do not decompose role-mix from true premium econometrically here; §3.2 estimates a skill component conditional on the other modelled skills, not net of seniority.

3.2 The AI skill price list (hedonic)

To estimate what a specific skill is worth net of the rest of the modelled stack, we fit a hedonic model: an ordinary-least-squares regression of log(annual pay) on a set of skill indicators (skill dummies only — no seniority, location or employer terms), run on 26,006 US AI postings with stated, non-estimated salaries. Restricting to a single currency removes the exchange-rate confound; each coefficient is the pay difference associated with a skill holding the other modelled skills constant (not holding seniority constant — skills that travel with senior roles will absorb some of that signal).

Skills that command a premium (conditional pay lift, 95% CI):

Skills that carry no standalone premium — the commodity layer:

The list reads like the report's whole thesis compressed into one regression. The pay is in the frontier and infrastructure layer — PyTorch over TensorFlow (the framework-rotation story of Chapter 4, priced), distributed compute (Spark), orchestration (Kubernetes), the LLM/RAG/MLOps production stack. The negative coefficients are the more interesting half, and the easiest to misread, so read them precisely: a −11.6% on Docker does not mean "learning Docker cuts your salary." It means that, once a posting already lists the modelled AI stack, also naming Docker, SQL or Python adds nothing — these are table-stakes that appear disproportionately in lower-paying generalist roles. Prompt engineering at −6.1% is the cleanest anti-hype data point in the report: the most-searched AI skill (Chapter 2) prices like a commodity, not a craft.

Honesty note. Correlation, not causation. R² = 0.10 — skills explain only a tenth of pay variance (seniority, employer and location carry the rest), so these are associations within disclosed US postings, not returns to learning a tool. The model controls for the top 25 skills only — there is no seniority/location/employer term — so a coefficient is net of the other modelled skills, not net of seniority; collinear skill bundles share credit. Standard errors are classical (homoskedastic), so the 95% intervals are, if anything, slightly optimistic.

3.3 The Pay-Transparency Tracker — flagship

Here the market splits in two. Of all AI postings whose text we could parse, 56.5% state pay — but that average is a transatlantic illusion. Disclosure runs at 77.3% in the United States and just 18.7% across the European markets we cover.

Inside Europe the variance is the headline, not the mean — disclosure is a national policy fingerprint, not a continental constant:

Austria already behaves like a disclosure regime (mandatory minimum-salary statements in job ads have been law for years); Germany, Spain and Switzerland sit near the floor. This is the live policy surface of the report. The EU Pay Transparency Directive (2023/970) — which gives candidates the right to pay information before hiring, bans questions about salary history, and phases in mandatory pay-gap reporting by employer size (250+ and 150–249 staff from 7 June 2027; 100–149 only from 7 June 2031; under-100 exempt) — had its national transposition deadline on 7 June 2026. The gap between Austria's 72% and Germany's 10% is precisely the gap the Directive is written to close, and our tracker is positioned to measure it country-by-country as each national law lands.

Honesty note. "Disclosure in the posting" is a lower bound on compliance — an employer can comply through a pay portal without printing a band in the ad. Small-n markets (Austria, Finland, Ireland) are shown for completeness but flagged; we never average them into a "Europe" claim. The pool mixes US-structured and EU-parsed salaries, which is exactly why we report the two scopes separately rather than as one number.

3.4 Salary range honesty

Stating a number is not the same as stating a useful one. Among AI postings that disclose pay, the typology is:

So disclosure, when it happens, is mostly an honest two-ended range — open-ended "from" and "up to" framings are rare (under 2% combined). The remaining question is how wide those ranges run, because a band can technically comply while telling a candidate almost nothing. We measure width as (max − min) / midpoint, a unit-free ratio that pools across currencies:

• Median range width: 40% of the midpoint (e.g. a $150K–$225K posting).
• Inter-quartile range: 25.6% to 59.7%.
• 47.4% of ranges are "wide" (>40% of midpoint); only 3.8% are tight (<10%).

A 40%-wide median band is the quiet caveat behind every "salaries are transparent now" headline: nearly half of all disclosed AI ranges are wide enough to defer the real negotiation, not resolve it. Transparency has a quality axis, and the market is clustered at the permissive end of it.

Honesty note. Min-only and "fake" one-sided ranges are excluded from the width calculation; width is a transparency proxy, not a pay level.

3.5 EU vs US compensation is built differently

Disclosure is not the only thing that differs across the Atlantic — the composition of the package does too. Reading the same incidence fields by region:

These are two different compensation cultures, visible in the postings themselves. US AI pay is variable and equity-loaded — nearly a third of US AI roles mention equity and almost half mention a bonus — the Silicon-Valley template of base + RSUs + bonus carried into the AI hiring wave. European AI pay is base-anchored: equity appears in under one in twelve postings, bonus in one in eight, and the 13th-month salary — structurally absent from US ads — surfaces only on the EU side. For a candidate comparing a US and an EU offer, the headline base is not the comparison that matters; the shape of the package is, and the shape is regional.

Honesty note. Incidence = mentioned in the posting, not contractually guaranteed or valued. Equity is structurally rarer in European private-company hiring, so part of this gap is labour-market structure, not disclosure behaviour.

3.6 Benefits beyond base pay (forthcoming — ILIKE batch)

Health coverage, paid time off, learning-and-development budgets and relocation support form a fourth layer of the package that lives in free text rather than structured fields. Quantifying it (health ~31% / PTO ~27% / L&D ~5.7% by country, per our analytics inventory) requires one text-mining pass over the description corpus; this section is scheduled for the next ledger batch and intentionally left as a placeholder rather than filled with un-frozen numbers.

Named indexes: AI Stack Index · Vendor Mindshare / Model-War Scoreboard (flagship) · AI Skill Demand–Supply Gap The gaps this chapter closes: adoption narrative vs revealed behaviour (Gap 5) and model vendors in the hiring stack (Gap 6).

This is the chapter where surveys can't follow. When a company writes "PyTorch, AWS, RAG, Claude" into a job description, it is not answering a questionnaire about its AI strategy — it is spending money to hire someone to build on that exact stack. A job posting is not usage telemetry; it is a budget signal, and budget signals are harder to fake than survey answers. Read at scale across our corpus, those signals reveal which foundation models employers are actually standardising on, which clouds, which frameworks are winning the rotation, and how absurdly long the average "required skills" list has become. We read the stack in four layers — foundation model, application tooling, production infrastructure, governance — and the most newsworthy layer is the first one.

4.1 The Model-War Scoreboard — foundation-model mindshare (flagship)

Job descriptions have quietly become the public scoreboard of the model wars. Among active AI postings, OpenAI/GPT appears in 14.7% and Claude (Anthropic) in 13.2% — a gap of just 1.5 points — with Gemini a distant third at 6.1%, then Llama (1.9%) and Mistral (1.2%).

Two things make this the report's most-shareable exhibit. First, the race is close — a year ago the gap was much wider; Claude has compressed it to a rounding error, and on present trajectory the hiring stack is heading for a dead heat at the top. Second, it inverts the demand side. In Chapter 2, candidate search interest put Anthropic at 1.88× OpenAI — yet employer job descriptions still name GPT marginally more often. Candidates are more curious about Anthropic than employers' stacks have yet caught up to. Those two instruments — what people search vs what employers write — disagree, and the disagreement is the story: mindshare leads stack-share, and we can see the gap closing in real time.

Honesty note. Presence of a vendor token in a description is mention, not a hard requirement (we do not parse negations or "nice to have"). It measures which models are entering the hiring conversation — the revealed-behaviour signal — not contracted spend. Pair with the demand-side mirror (ch2.employer_interest). Exhibit: chart_21 + vendor-share-over-time.

4.2 Cloud platform split & the multi-cloud reality

One layer down, the cloud split is a three-horse race tilting to the incumbent: AWS 33.4%, Azure 29.2%, GCP 23.9%, with Google's AI-specific surfaces (Vertex AI 5.5%, SageMaker 4.0%) trailing. The more interesting number is co-occurrence: 14,855 postings name two or more clouds. Multi-cloud is not an edge case in AI hiring — it is the default expectation, and it widens the real skill bar well beyond any single provider.

Honesty note. Skill/regex presence; multi-cloud = ≥2 cloud skills co-listed in one posting.

4.3 Framework rotation: PyTorch over TensorFlow

The framework layer shows the rotation everyone talks about, measured: PyTorch leads at 17.8%, TensorFlow at 13.8% — PyTorch is now the default research-to-production framework, TensorFlow the legacy one. This is not just a popularity contest; Chapter 3's hedonic model prices the rotation, with PyTorch carrying a +9.7% pay premium and TensorFlow a −8.2% conditional discount. The market is paying for the winning framework and discounting the fading one — the clearest possible signal of which way the stack is rotating.

Honesty note. Reported as current share level; the month-over-month rotation series needs the accumulating snapshot panel (🟡). Direction is corroborated independently by the pay coefficients.

4.4 Technique adoption: the application layer

Above the models sits the application toolkit, and it has moved fast from buzzword to baseline requirement: RAG appears in 22.7% of active AI postings, prompt engineering 14.7%, LangChain 10.8%, vector databases 9.1%, fine-tuning 8.7%. The signal worth flagging for 2026 is Model Context Protocol (MCP) at 6.2% and LangGraph at 5.9% — agent-infrastructure tooling that barely existed a year ago, already on one in sixteen AI postings.

RAG crossing into nearly a quarter of postings is the moment a technique stops being a differentiator and becomes table stakes; MCP at 6.2% is the next one entering that curve. This is the agentic shift of Chapter 2's demand data (Agentic AI Engineer +360%) showing up on the requirements side.

Honesty note. Presence in JD; "when a technique crossed from hype to requirement," not contracted use.

4.5 What clusters into a real stack — and the skill-bloat problem

Skills do not appear at random; they cluster into recognisable archetype stacks. The most co-occurring pairs trace the shape of the market:

But cluster size also exposes a pathology. The median AI posting now lists 8 distinct skills; the 99th percentile lists 33; the single most demanding lists 55. A 55-skill "requirements" list is not a specification — it is a wish-list, and the long tail of skill-bloat is one reason candidates and employers talk past each other (Chapter 2's misallocation has a supply-side cause here).

Honesty note. Skill = regex presence; the "unicorn" tail reflects JD wish-lists, not literal must-haves.

4.6 Skill demand vs supply — job-identities vs ingredients (forthcoming)

The final cut pairs what employers require (corpus skill share) against what candidates search (skill-level DataForSEO demand) — the same demand×supply logic as Chapter 2, but at the skill level. The expected finding: candidates search for professions ("Data Science," "OpenAI") while employers require ingredients (LLM, PyTorch, RAG, vector DB) that almost no one searches for by name. This section is scheduled for the next ledger batch (skill_demand_* join) rather than filled with un-frozen numbers.

Named index: AI Role Profiles The gaps this chapter closes: demand vs supply (Gap 2) and hype vs hiring (Gap 3), role by role.

The previous chapters measured the AI labour market in aggregate. This one makes it personal: a card per role, answering the seven questions a candidate actually asks — what is it, is it growing, is it over- or under-chased, what skills, what pay, where, and is it a real opportunity or a trap? We built the cards by collapsing 58,150 messy job titles into a 26-family taxonomy (95.6% classified), then joining each family to its demand curve, salary band, seniority mix and skill stack. The headline verdict for each role comes straight from the Opportunity-vs-Hype-Trap matrix of Chapter 2.

Reading note — two different supply measures. Live postings is the taxonomy family count (broad — it absorbs title variants and, for some families, adjacent roles). Searches/posting is Chapter 2's demand×supply ratio, computed on a narrower title-matched live supply — so the two columns use different supply definitions and are not directly divisible (Chapter 2 has the matched counts). The gap is small for most roles but large where a family is much broader than its title match: for AI Agent Engineer and LLM Engineer especially, read the ratio as Ch.2's title-matched leverage, not as searches ÷ the family count. Pay is the median of disclosed USD postings (US-skewed; European pay is lower and less often disclosed — Chapter 3); figures resting on fewer than 30 disclosed postings are marked † and are directional only. Seniority is the entry / mid / senior+ split among levelled postings.

The roster at a glance

¹ Agentic-AI demand momentum (the search term "agentic AI engineer"); see note on consulting contamination below.

† Prompt Engineer pay and seniority rest on a thin disclosed sample (USD-median n=16; levelled n=28 — both below our n-floor of 30) and a supply that is ~76% German, so the level is directional only. The hype-trap verdict does not depend on it: it rests on the demand×supply ratio (316.6 searches per posting) and falling momentum (−19% YoY).

Flagship cards

AI Engineer — durable core

What it is: the default build role of the AI era — ship LLM-powered features end to end. Demand/supply: balanced (10 searches per posting) and growing (+52% YoY) on the largest supply base in the market (6,761 live). Pay: $175K median (disclosed USD). Seniority: 58% senior+, but 14% entry — one of the more open senior-tilted doors. Top skills: Python, Machine Learning, LLMs, RAG, AWS. Where: US 47%, Germany 10%, UK 8%. Verdict: the safest bet in AI hiring — high volume, real demand, not over-chased.

AI Agent Engineer — opportunity (rising)

What it is: build autonomous, tool-using LLM agents. Demand/supply: the fastest-rising demand in the dataset (agentic search +360% YoY) against a still-modest 2,754 live postings. Pay: $191K. Top skills: LLMs, Agentic AI, Python, RAG. Where: US 64%, Germany 10%. Verdict: the clearest rising opportunity — but read the caveat.

⚠️ Honesty flag. ~26% of titles our taxonomy assigns here are consulting roles ("agentic commerce manager/consultant" at Deloitte/Accenture), not engineering. The de-duplicated engineering core is smaller and is split out before publication; the consulting share is itself a finding (Chapter 7).

MLOps Engineer — durable, employer-favourable

What it is: the production plumbing — deploy, monitor, scale models. Demand/supply: only 5.8 searches per posting (employer-favourable: postings outnumber searchers) and rising (+20%). Pay: $139K median — lower than the build roles, the one durable role that is not top-paid. Top skills: Kubernetes, CI/CD, cloud. Verdict: durable and under-chased — a strong, unglamorous bet.

AI Product Manager — crowded

What it is: own the what/why of AI products. Demand/supply: crowded (40.9 searches per posting) but growing (+40%). Pay: $188K, 73% senior+ — a senior-only door. Verdict: real and well-paid, but heavily chased and rarely open to juniors.

AI Safety / Governance & Model Evaluation — specialists

What they are: the responsible-AI function made concrete — evaluate, red-team, govern. Supply: small (676 governance + 213 evaluation live). Pay: Model Evaluation Engineer is the top-paid role in the entire roster at $207K — scarcity pricing. Verdict: specialist and early (Chapter 8: 3.8% of postings are dedicated responsible-AI roles), but the highest pay-per-posting in AI.

Prompt Engineer — hype trap

What it is: the role that defined the hype. Demand/supply: 316 searches per live posting — wildly over-chased — and demand is falling (−19% YoY). Pay: ~$117K† — but on a thin, US-only disclosed sample (n=16) against a supply that is ~76% German, so treat the level as directional. Verdict: the textbook hype trap — peak attention, fading reality, and a thin posted supply. The verdict rests on the demand×supply overhang and falling momentum, not on the noisy pay figure. The skills are better deployed inside an AI Engineer or LLM Engineer role.

The floor and the frontier (two contrasts)

• AI Trainer / Data Annotator — the floor: $73K median, 74% non-senior. The human-in-the-loop tier that powers model training, paid a third less than the engineering roles. A real on-ramp, honestly labelled.
• Forward-Deployed AI Engineer — the frontier: $200K, 79% US, the lowest remote share (12%) — a vendor-embedded, on-site, US-centric archetype (Palantir/AWS/Snowflake lineage) that barely exists in Europe.

Named indexes: AI Mobility Openness Index (flagship) · Language Barrier Map · Visa Sponsorship Scarcity · Remote Reality Index · Entry Access Index The gap this chapter closes: US vs Europe (Gap 4) — reframed from volume to access.

Every other chapter measures how big the AI market is. This one measures whether you can get in. AI is global, but hiring is intensely local: the same role is a senior-only market in one country and an entry door in another; some markets demand a C1 command of the local language before they will read your CV; visa sponsorship is the exception, not the rule; and "remote" usually means "remote, but inside this one country." For a candidate — and for any honest account of the European AI labour market — the access map matters as much as the demand map. We build it from the 45,142 postings we parsed field by field, and combine the four barriers into one AI Mobility Openness Index.

6.1 The seniority gradient — where the junior door is open

AI hiring is, almost everywhere, a senior market — but how senior varies sharply by country. Among postings with a parsed level, senior-and-above runs 76% in Ireland, 73% in the UK, 66% in the US and Germany — and collapses to 26% in Italy, where half of all AI postings (51%) are entry-level.

Italy is the clearest entry door in the dataset; Ireland and the UK are the steepest senior gates. For a junior candidate, which country is a bigger determinant of access than which role.

Honesty note. Computed among the ~79% of postings with a parsed seniority level (seniority_norm, model-agnostic); "entry" = junior + intern. This supersedes raw single-field seniority (which over-stated US senior share); the honest US figure is 66% senior+, not the 94% an unnormalised field suggests.

6.2 The Language Barrier Map

The single most underrated AI-hiring barrier in Europe is language. Among postings that state a language requirement, the share demanding a non-English language is 77% in Germany, 68% in the Netherlands, 67% in Austria, 58% in Switzerland and 57% in Belgium — versus the English-open markets of the UK and Ireland.

This single map explains a large share of why "AI is global" fails in practice. A German-resident AI role is, three times in four, also a German-language role — a gate invisible to anyone reading only salary and skills. The DACH region is the hardest language wall; GB/IE the most open.

Honesty note. "Required" non-English language from the parsed CEFR list; absence of a stated requirement is not proof a posting is English-friendly.

6.3 Visa Sponsorship Scarcity

Most postings say nothing about visas. Among those that do, sponsorship is the minority: just 23% offer it, while 77% explicitly state they do not (1,456 vs 4,804). Relocation help is rarer still.

For a candidate who needs sponsorship, the practical AI market is a fraction of the headline one — and the postings that say "no sponsorship" do the candidate a backhanded favour by being explicit. The silence of the rest is its own signal.

Honesty note. Among postings that explicitly address visa status only; the majority are silent, so this is the disclosed slice, not the whole market.

6.4 The Remote Reality Index

"Remote AI jobs" is mostly a myth of degree. Among postings with a parsed work mode, hybrid is the default at 60%, on-site 28%, and fully remote only 12%. And of the remote roles that specify a scope, the overwhelming majority are country-locked: 4,175 country-only versus just 820 that are genuinely cross-border (EU-wide or worldwide).

So the "remote revolution" in AI hiring resolves to: mostly hybrid, rarely fully remote, and when remote, almost never the kind that actually lets you live in a different country. True location-independent AI work is a rounding error.

Honesty note. work_mode/remote_scope parsed from text; "true remote" = explicitly EU-wide or worldwide scope.

6.5 The Entry Access Index + impossible requirements

The education floor is more open than its reputation: among postings stating a minimum, 74% accept a bachelor's, 19% require a master's, and only 4.6% require a PhD (a further 9% list PhD as "preferred"). The barrier is rarely the degree.

Where postings do over-reach is experience: a recurring pattern (quantified fully in the next ledger batch) is roles demanding "5+ years" of a technology that has existed for two or three — RAG, LangChain, agentic frameworks. The years-of-experience bar, cross-referenced against each technology's birth date (Chapter 2), is the report's sharpest "impossible requirement" exhibit and is scheduled for the ILIKE/years batch.

6.6 Country archetypes (forthcoming — needs role taxonomy applied per country)

Folding the role taxonomy (26 families, Chapter 5) into per-country role-mix gives each market a personality — Italy's ML lean, the US frontier-research tilt, Poland's platform-engineering signature. This cut is scheduled once the taxonomy de-duplication (Agent-Engineer consulting split) lands.

The flagship — AI Mobility Openness Index

Combining the four barriers — language openness + visa availability + remote scope + entry access — into one composite ranks each market on how open it is to entering AI, independent of how many jobs it has. The early shape: the UK and Ireland score high on language and seniority-adjusted openness but low on visa; the DACH markets are large but language-gated; Italy is small but the most entry-open. The composite is the chapter's citable headline and the spine of the per-country profiles.

Named index: AI Employer Concentration & Intensity Index The gap this chapter closes: part of adoption-narrative vs revealed behaviour (Gap 5) — who is really building.

Ask who is hiring for AI and the honest first answer is uncomfortable: a consulting firm. Deloitte is the single largest AI employer in our corpus, and consultancies as a class account for nearly one in five AI postings. Strip them out and the landscape changes shape entirely. This chapter has one ironclad rule — every employer ranking is shown twice, with and without consultancies — because any "top AI employers" list that doesn't is quietly misleading. Underneath the headline names, the market is a long tail: ten thousand companies, half of them posting a single AI role.

7.1 Concentration & the long tail

AI hiring is broad, not concentrated. The top 10 employers account for just 24.4% of active AI postings and the top 100 for 41.8% — a Herfindahl index of ~112, which is a competitive, unconcentrated market by any antitrust yardstick. The reason is the tail: 10,404 distinct companies post AI roles, and 48.5% of them post exactly one.

So the story is not "a few giants hoard AI talent." It is the opposite: AI hiring has diffused across the whole economy, with a very long tail of companies making their first AI hire. That diffusion is the bullish signal — and the single-posting half is where the next wave of demand is forming.

Honesty note. Entity resolution is incomplete (the same employer under name variants is undercounted), so concentration figures are a lower bound — the true market is at least this diffuse.

7.2 The consultancy distortion (must-read)

Here is the rule of the chapter, quantified. Consultancies account for 18.3% of active AI postings — Deloitte (4,326), Accenture (2,088), PwC (1,271), Capgemini and BCG behind them. Deloitte alone is roughly 8% of the entire AI market on its own.

These are not AI products being built in-house; they are bodies being staffed against client engagements. A consultancy posting "AI Engineer" 4,000 times is a labour-arbitrage signal, not an AI-adoption one. That is why every employer exhibit in this report carries an All employers bar and an Excluding consultancies bar side by side. Read the second one to see who is actually building AI for themselves.

Honesty note. Consultancy membership is name-pattern matched; borderline firms (e.g. IBM) are debatable and disclosed in methodology. The point is directional and large enough that omitting it distorts every ranking.

7.3 Industry intensity + global vs local hirers

By volume, Internet & Software leads AI hiring, followed by Consulting & Business Services, Computers & Electronics, Industrial Manufacturing and — notably — Education. (Industry is recorded on only ~12% of postings, so this is directional.)

The more striking cut is geographic reach: 405 companies post AI roles in three or more countries, and 21 in ten or more. A small set of genuinely borderless hirers — the global software and consulting firms — run AI teams that span the map, while the long tail hires within a single country. The borderless 21 are the companies for whom AI talent is a single global pool; everyone else competes locally.

Honesty note. Staffing reposters inflate multi-country counts; industry coverage ~12% → volume, not true intensity.

7.4 AI Team Blueprints (forthcoming)

The most distinctive employer cut — which roles companies hire together when they build an AI function (e.g. AI product + data engineering + MLOps, or LLM app + RAG + platform) — needs the role taxonomy (Chapter 5) folded into per-company co-posting. Combined with a maturity read (experimentation → application → production → governance → scaling), it shows not just who is hiring but what stage they are at. This AI Team Blueprints analysis, and the company-stage split, are scheduled for the next batch (role-taxonomy join + ILIKE stage signals).

Sidebar — AI-washing detector (lite)

A careful, non-accusatory read: compare a company's public "AI transformation" claims against how many AI roles it actually has open, of what kind (research / product / platform / governance / data-labelling), and at what velocity. We frame any gap as a revealed signal, never as "the company is lying." Scheduled alongside Team Blueprints.

Named index: Responsible-AI Hiring Density The gap this chapter closes: adoption-narrative vs revealed behaviour (Gap 5), in its sharpest form.

Everyone says they are doing responsible AI. Almost no one is hiring for it. That one sentence is the chapter. The question we ask — are companies hiring for responsible AI as fast as they claim to adopt AI? — has a measurable answer, because a posting either creates a dedicated safety, governance or evaluation role or it does not. Responsible AI is mentioned in 17% of AI postings but is the actual job in under 4%. The rest is language. We grade the difference on four levels so the headline can't be gamed, and the gap between talking and hiring is the most important number in the chapter.

8.1 How much responsible-AI is actually required

To avoid counting marketing copy as governance, we grade every posting on four levels: (1) mention (the words appear), (2) responsibility ("responsible for the AI model…" — not ethics), (3) hard requirement (governance/safety as an actual requirement), (4) dedicated role family (the job is safety / evaluation / governance). The headline is levels 3–4.

• Mention (level 1): 17.4% of AI postings reference responsible AI, ethics, governance, safety, alignment or evaluation.
• Dedicated role family (level 4): 3.8% — postings that are genuinely AI governance, safety or security roles.

The spread — 17% talk, under 4% hire — is the revealed-behaviour finding. Responsible AI has saturated the language of job postings while remaining a niche as an actual function. The gap is not hypocrisy so much as immaturity: governance is still being written into the culture faster than into the org chart.

Honesty note. Levels 1–4 deliberately separated; "mention" is an ILIKE lower-bar signal, "role family" comes from our category model. We headline the role-family figure precisely so the 17% can't be mis-sold as "17% of AI hiring is responsible-AI."

8.2 Compliance & the AI Act hiring signal

Regulatory compliance surfaces in AI postings at a similar low level, concentrated in regulated sectors: GDPR is named in 3.0% of AI postings and security/assurance frameworks (SOC 2, ISO 27001, HIPAA) in 3.2%.

These are small but rising shares, and they are the leading edge of the EU AI Act hiring wave: as the Act's obligations phase in, the compliance vocabulary that today sits in 3% of postings is the series to watch. A report that tracks it now owns the baseline against which the AI-Act hiring response will be measured.

Honesty note. Regex presence, not a verified hard requirement; tied to sector (finance, health) more than to AI maturity per se.

8.3 Rhetoric vs reality

Putting the two together: companies adopt the vocabulary of responsible AI far faster than the roles. 17.4% of AI postings invoke responsible AI; 3.8% create a job to do it; ~3% mention the compliance frameworks that would operationalise it. The marketing of trustworthy AI is everywhere; the budgeted, staffed function is rare — exactly the kind of gap that job postings reveal and surveys miss, because a survey records intent while a posting records spend.

This is not a cynical finding. It is an early-market one — and it is directional: the categories are small today but among the fastest-rising (model evaluation, AI safety), which is why this is a standalone chapter and not a footnote.

Not an appendix for show — a credibility engine. Open data, transparent definitions and named caveats are what make a single-corpus report citable and resistant to "it's just one source." Every headline number in this report is frozen in a single Facts Ledger (research/data/report_ledger/ledger.json, v2026.06) with {value, n, as_of, source, scope, caveat}, so prose, charts and web pages cannot disagree.

M1 — Data sources & provenance

Five linked assets, hubbed on job_posting.id and skill.key:

1. Posting corpus — 336,518 postings / 58,150 AI roles, 18 countries — our own first-party corpus, with Apply links resolving to 18,093 unique employer & ATS sites. AI = ai_relevance ∈ {core, probable}; probable measures ~83% true-AI on LLM re-classification.
2. LLM text extractions — 45,142 AI postings parsed by gpt-5-mini into 32 strict fields (salary, seniority, languages, visa, remote, education, comp components).
3. Role-demand panel — DataForSEO Google Ads search demand, 155 keywords × 18 markets × 47 months (demand_*), normalised to June 2022 = 100.
4. Skill-demand panel — DataForSEO dual-intent, 136 skills × 18 markets × 48 months (skill_demand_*).
5. Daily supply snapshot — market_snapshot, the accumulating series behind pay premium and intensity.

(Optional 6th, not in this edition: an independent ATS/source layer — 140K postings / ~10K AI / 126 countries — held as a separate data product.)

M2 — Cross-source concordance & salary provenance

Where two captures describe the same role, source agreement can be measured on shared dedup keys. Critically, salary is reported in three states, never blended: stated (disclosed in the posting), inferred (third-party benchmark estimate), and none. Disclosed ≠ estimated; the disclosure rate (Chapter 3) is a separate metric from the pay level, and its missingness is not random (it tracks national law and employer policy).

M3 — Demand–supply joinability

There is no shared key between search keywords and posting titles, so the demand×supply gap (Chapter 2) uses a manual keyword↔role crosswalk with word-boundary title matching. Demand latency is ~1–2 months (latest panel month 2026-04).

M4 — The honesty caveats (consolidated)

Stated plainly, because self-criticism is the armour of a single-corpus report. Several supersede earlier internal estimates now that the corpus has grown:

• Seeded corpus → we never claim "AI = X% of all jobs"; growth and composition only, share triangulated externally.
• Seniority — seniority_norm (model-agnostic) covers ~79% of extractions and includes a mid tier; the honest US senior-and-above share is 66%, not the ~94% an un-normalised field implies.
• Disclosure scope — computed by region (EU 18.7% / US 77.3%), because the extraction pool now mixes US-structured and EU-parsed salaries; a single blended "~32%" would be wrong.
• Hedonic skill pricing — associations, not causation; R²≈0.10 (skills explain a tenth of pay); USD-only to remove currency confound; collinear skill bundles share credit.
• Skills = regex presence, not LLM-verified requirements; the "unicorn" tail (up to 55 skills/posting) reflects JD wish-lists.
• Model/vendor/technique shares = token presence in descriptions, not hard requirements (no negation handling).
• Consultancies = 18.3% of AI postings → every employer ranking is shown All vs Excluding-consultancies.
• Responsible-AI graded on 4 levels (mention 17.4% vs role-family 3.8%); we headline the role-family figure.
• Company entity resolution incomplete → concentration figures are a lower bound.
• Industry coverage ~12% → industry cuts are directional volume, not true intensity.
• Trend by date_posted is ramp-up-distorted → we report shares, not absolute counts, over recent windows.
• City = search seed, company_rating/reviews ~empty; a subset of captures is research-only (no direct apply link; excluded from public listings).
• Role taxonomy is a first-pass regex mapping (26 families, 95.6% coverage); AI Agent Engineer is ~26% consulting and is de-duplicated before Role Cards.
• Small-n countries are suppressed below an n-floor (default 30; per-country floors higher) — we never call 3–5 markets "Europe."
• Search-demand ambiguity — Google merges close head-variants; brand keywords and homonyms (e.g. AI trainer, zodiac/animal collisions in raw skill demand) handled via job-intent terms and a confidence-tiered keyword appendix.

M5 — How to cite / changelog / data access

Named-index registry with stable ids and Cite this chart / Cite this report / Dataset version (v2026.06); downloadable CSV tables + chart registry; DOI placeholder (Zenodo). A "coming editions" roadmap covers the deferred cuts (benefits atlas, impossible-requirements years, country archetypes, team blueprints, skill demand×supply).

M6 — Definitions & normalization (disclosure checklist)

Data sources · AI classification (native / enabled / infra / governance) · role taxonomy (26 families, manual) · skill taxonomy (pre-classified skill.kind) · demand-panel methodology · salary normalization → stated / inferred / none (disclosure is a separate metric; missingness non-random) · disclosure definition · country coverage · duplicate handling (dedup key + source hierarchy) · minimum-n thresholds (suppress, confidence labels) · known biases · keyword appendix (confidence tiers, manual exclusions) · changelog / dataset version · public CSV / chart registry · cite-this.

M7 — Risks & mitigations