Reading the AI Hiring Gap: Which Roles Are Open

What you need to know

Most people look for an AI job the same way: open a job board, type "AI engineer", and start applying to whatever comes back. That method has a specific failure built into it. The results page mixes together at least three completely different kinds of job, hired through different processes and judged on different evidence, and it gives you no way to tell them apart. You end up applying broadly into the largest category, which is almost never the category you were aiming for, and concluding from the silence that the market is brutal.

Two sets of published figures, one for each of our markets, say the same thing in different accents. In the United Kingdom, the PwC AI Jobs Barometer 2026, published on 23 June 2026, reports that AI-related job postings recovered to roughly 180,000 in 2025 after falling from 182,000 in 2022 to 119,000 in 2023 and 112,000 in 2024. Specialist AI postings surged 61 per cent in 2025, adding 68,000 roles. But the split inside that surge is the whole story: AI user roles — people using AI inside a job they already had — rose 65.8 per cent, roughly 66,000 positions, while AI developer roles rose about 21 per cent, roughly 2,600 positions. That is a ratio of about twenty-five to one.

In India, the direction is the same and the scale is larger. Projections put AI job postings at roughly 3.82 lakh (382,000) in 2026, about 32 per cent up on roughly 290,000 in 2025. The Quess Corp India AI Workforce Analysis 2026, which examined around 350,000 postings, sizes the working population at about 920,000 AI professionals — of whom roughly 257,000 hold core AI roles and roughly 663,000 hold AI-embedded roles. NASSCOM-BCG puts year-on-year growth in AI engineer roles specifically at 67 per cent, and NASSCOM's own estimates say India needs close to a million AI-skilled professionals by 2027 against a trained pool of roughly 500,000 to 650,000.

Those numbers will be out of date within a year. The method for reading them will not, and the method is what this guide is about. We reported the two markets side by side as a news story in our coverage of the India and UK AI postings gap; this piece is the practitioner's companion to it. You will learn to sort any advert into one of three classes in under a minute, to build your own market map from thirty to fifty live adverts in your own city, to read what the resulting composition means for how you should position, and to choose a target role from evidence rather than from the vibe of a job board. Then you re-run it in six months, because the numbers move and the method does not.

The one distinction that does most of the work

Before anything else, learn to separate three classes of job that all advertise under the same vocabulary.

AI-user roles. An existing job in which a model has become one of the tools. A marketing executive drafting copy with an assistant, a civil servant triaging casework with a classifier, a paralegal summarising bundles, a support agent working a suggested-reply queue. The work is the original work. The AI capability being hired for is fluent, sceptical use — knowing what the tool is good at, checking its output, and not pasting confidential material into it. This is by a wide margin the largest and fastest-growing class in both markets.

AI-embedded roles. A genuinely technical job in which model behaviour is one component among several. A backend engineer wiring a provider API into a product surface, a data engineer building the retrieval pipeline that feeds it, a mobile developer shipping the interface around it, a platform engineer running the queue that carries the requests. This is real engineering and often excellent work. What distinguishes it is that the model is a dependency rather than the object of the job: someone else usually chose it, and quality is somebody else's number.

Core AI roles. The job exists because the model exists. The holder owns the model choice, builds and maintains the evaluation set, owns the serving path, owns the cost and latency budget, and is accountable when output quality moves in either direction. This is the smallest class in both markets by a large margin, and it is the class most readers of this article believe they are applying to.

The three classes are not a hierarchy of worth. They are a hierarchy of scarcity, and they are hired through different funnels — user roles increasingly through internal capability programmes and volume recruitment, embedded roles through ordinary engineering pipelines, core roles disproportionately through search, referral and public evidence. Applying to one with the evidence appropriate to another is the single most common wasted effort in AI job hunting.

Classifying an advert in under a minute

You do not need to read a job description carefully to classify it. You need to answer four questions, and the answers are almost always sitting in the responsibilities section.

Who owns the model choice? If the advert names a specific model or provider as a fixed part of the environment, someone else chose it. If it asks the holder to "evaluate and select" models, or talks about routing between them, that decision sits with the role. Is there an evaluation or latency requirement? Any mention of an evaluation set, a golden set, offline scoring, regression testing of outputs, p95 latency or cost per request is a strong core signal, because those are the things only an owner is asked to hold. Does it name a serving stack? Inference servers, GPU scheduling, vector stores, quantisation, batching, model registries. Is the deliverable a system or a report? A job whose outputs are decks, analyses, recommendations and adoption metrics is a user role, however much AI vocabulary surrounds it.

When an advert genuinely straddles two classes — and perhaps one in five does — classify it by the non-functional numbers. An advert that asks anybody to hold a latency or cost-per-request target is treating that person as an owner, whatever the title says. If you want to know how those ownership signals are then assessed once you are in the room, our breakdown of what AI hiring managers actually evaluate maps the interview stages onto the same distinction.

Build your own market map

Published national figures tell you the shape of a country. They do not tell you the shape of the market you will actually apply into, which is one city, one title family and one seniority band. That map you have to build yourself, and it takes about ninety minutes. Here is the exercise.

Step one: fix the scope before you look at anything. One geography — Bengaluru, or Manchester, or "remote, UK-contract". One title family, meaning the two or three titles you would genuinely accept, not everything with "AI" in it. One seniority band. One time window, normally the last thirty days. Write the four constraints down before you open a job board, because the temptation to widen them the moment results look thin is what destroys the exercise.

Step two: collect thirty to fifty live adverts. Do not filter for quality, appeal or whether you think you would get it. You are measuring the market, not shortlisting. Include the ones that look too senior and the ones that look beneath you; their presence in the tally is information.

Step three: classify each one using the four questions above, and write down the single sentence from the advert that decided it. That sentence is the part you will reuse later.

Step four: record the employer type for each advert — global capability centre, product startup, consultancy or services firm, enterprise in-house team, public sector, agency or recruiter listing. Employer type predicts class far more strongly than sector does, and it is the axis most job seekers never record.

Step five: record the three most repeated concrete requirements. Not "strong communication skills". Concrete: a named framework, a named cloud, a specific domain, a security clearance, a specific model family, a stated experience threshold. Three per advert, tallied across the set.

Step six: total the columns and read the shape. That is the map.

The tally template

Copy this into a spreadsheet or a plain text file. It is deliberately small; the discipline is in filling every column for every advert, not in the sophistication of the sheet.

MARKET MAP — [city] · [title family] · [seniority] · [window]
Run date: [DD MMM YYYY]         Sample size: [N] adverts

 #  | Employer type | Class | Deciding sentence (short) | Top requirement
----+---------------+-------+---------------------------+----------------
 01 | GCC           | EMB   | "integrate the model API" | AWS
 02 | Startup       | CORE  | "own the eval harness"    | evals
 03 | Consultancy   | USER  | "drive AI adoption"       | stakeholder mgmt
 ...

TALLY
  Class split      USER [__]   EMBEDDED [__]   CORE [__]
  Employer mix     GCC [__] Startup [__] Consultancy [__]
                   Enterprise [__] Public [__] Agency [__]
  Core roles by employer type: ____________________________
  Three most repeated requirements:
    1. ______________________  (appeared in __ of __ adverts)
    2. ______________________  (appeared in __ of __ adverts)
    3. ______________________  (appeared in __ of __ adverts)

READ-OUT
  My realistic target class this cycle: ____________________
  Employer types that actually post it: ____________________
  The one requirement I cannot currently evidence: _________

Three patterns show up in almost every map, and each has a different response. If core roles are concentrated in one employer type — very often product startups in the UK, and a mix of startups and the more mature global capability centres in India — then your search is not a job-board search at all, it is a list of perhaps fifteen companies to be known by. If core roles are spread thinly across many employer types, the market is early and the winning move is domain specificity rather than general AI capability. And if your tally returns three or fewer core roles in thirty adverts, that is not a bad month; that is the actual base rate, and it means the target for this cycle should probably be a strong AI-embedded role at a company where core work is visible from the inside.

What each market's shape implies for how you position

The United Kingdom: a small builder market inside a shrinking pool

The UK picture as of September 2026 is unusually legible, because PwC's series is consistent and the Barometer separates user roles from developer roles explicitly. Specialist AI jobs now make up 2.2 per cent of the overall UK job market, up from 1.3 per cent the year before, and Indeed Hiring Lab found that UK postings referencing AI stood 127 per cent above pre-pandemic levels by February 2026. Sector share concentrates: technology, media and telecommunications accounts for around 10 per cent of AI postings and health for around 8 per cent.

Two facts change what you should do with all that. First, the growth is overwhelmingly in user roles — roughly 66,000 net new against roughly 2,600 developer roles in 2025. Second, this is happening inside a contracting labour market: overall UK vacancies across the whole economy fell 6.6 per cent year on year. So the builder market is small in absolute terms and small relative to the number of people applying into it, and the general market is not absorbing the overflow.

The strategic implication is uncomfortable but simple. Volume applying is a losing strategy in a market that added roughly 2,600 advertised developer roles in a year, and where employers have a strong incentive to search rather than sift. Being findable and being specific wins. The pricing supports this: the UK AI-skills wage premium reached 34.2 per cent, up from 11 per cent in 2024, and a premium of that size is not paid for familiarity with a tool. It is paid for provable capability, which means the burden is on you to make the proof public and legible. If your route in is from conventional software engineering, the sequencing in our backend-to-AI-engineer transition roadmap is built around exactly this constraint, and the premium is also what makes the two-tier picture in our guide to AI engineer salary negotiation worth reading before any offer conversation.

India: demand is not the constraint, verified supply is

India's numbers describe a different problem. Postings are projected at 382,000 for 2026, up about 32 per cent on around 290,000 in 2025, and NASSCOM-BCG puts growth in AI engineer roles specifically at 67 per cent year on year. Against that, NASSCOM's estimate of close to a million AI-skilled professionals needed by 2027 sits opposite a trained pool of roughly 500,000 to 650,000. Hiring concentrates in Bengaluru, Hyderabad, Pune and Delhi NCR, with tier-2 cities growing, and the leading sectors are banking and financial services, e-commerce and retail, IT services and global capability centres, and healthcare and pharmaceuticals. You can follow the sector-level debate in NASSCOM's community research, which is where much of the Indian workforce analysis is discussed openly.

The constraint in India, in other words, is core-role supply rather than demand. That sounds like unambiguous good news for a candidate, and it is diluted by one fact from the Quess Corp analysis: 663,000 of the 920,000 people counted as AI professionals sit in AI-embedded roles, against 257,000 in core roles. Because the embedded population is so much larger, an "AI" job title in the Indian market is weak evidence of core capability, and experienced hirers read it that way. They discount the title and look for something else.

What they look for instead is specific and entirely learnable: a shipped system, the constraint that made it hard, the evaluation set you built, the measured result against a stated baseline, and the cost per task with a date attached. That evidence substitutes for the title, and it travels across both markets. Reported salary bands, aggregated across market surveys rather than drawn from any single authoritative source, run roughly from 6 to 9 lakh rupees a year for freshers in AI and machine learning roles, 12 to 30 lakh at three to six years, and 30 to 60 lakh at seven to ten years and beyond, with generative-AI specialists commonly cited around 25 to 50 lakh. Treat those as orientation, not as an anchor: the spread inside each band is wide, and it is driven far more by demonstrable ownership than by years served.

The public sector, described honestly

One finding in the PwC data deserves separate treatment because it is routinely misread. UK government accounts for nearly 5 per cent of AI postings as of 2026, up from under 1 per cent in 2022 — a genuinely striking rise, and it makes the public sector a real and growing employer in this space. But 97 per cent of those government AI roles are AI-user positions.

That is not a criticism of the roles. Public sector AI-user work can be substantial, stable and unusually consequential: casework triage, document handling, service redesign at national scale. It is a criticism of how the headline gets used. If you are a builder reading "government AI hiring up fivefold" as a signal to redirect your search, the composition figure says otherwise. Go in with accurate expectations: apply for public sector roles because the work and the conditions appeal, not because you expect to own a model there in the near term.

Choosing your target role from the map

A map is only useful if it changes your next application. The honest version of that decision is not "which role do I want?" but "which class of role does my current evidence actually support, and what is the one thing that moves me up a class?" The table below is the shortest version of that conversation I know how to write. Find your starting point, and read across.

Two notes on using this table honestly. The last column is deliberately singular: one proof, not a programme of study. A single well-documented system with a measurement attached does more than a stack of certificates, because it is the only evidence type that cannot be produced without doing the work. And the middle column is not a demotion. Taking a strong AI-embedded role at a company where core work happens in the next room is, in a market adding roughly 2,600 advertised builder roles a year, usually a faster route to core than a year of applying directly. If your map shows the core roles concentrated in platform and infrastructure work, our guide to getting hired as an agent platform engineer covers what that specialisation looks like in practice, and the project selection in our guide to portfolio projects that get you hired maps neatly onto the proofs in the final column.

Building evidence a hirer can verify

Everything above converges on the same requirement. In a market where the UK adds only around 2,600 advertised developer roles in a year, and where an AI title in India is a weak signal by base rate, the scarce thing is not capability. It is capability that a stranger can verify without meeting you.

Verifiable evidence has a consistent shape: a named constraint that made the work hard, an evaluation set you built and can describe, a measured result against a stated baseline, at least two named failure modes with what you did about them, and a cost or latency number with a date attached. That is five sentences per project. Written once, it works in a repository README, in an application, in a referral message and in the first ten minutes of an interview — and it is precisely the material that makes a referral possible, which is the mechanism behind our guide to getting referred into an AI team without a network.

The last piece is distribution, and it is the one most people skip. Evidence that lives only in a private repository does none of the work it is capable of doing. When the opening count is small and hirers search rather than sift, the question is not whether your work is good enough; it is whether the person looking can find it at all.

Five ways people misread the hiring gap

Chasing the headline growth number. "AI postings up 61 per cent" and "AI engineer roles up 67 per cent" are both true and both mostly describe classes of work other than the one being aimed at. Any growth figure that is not split by class is a number you cannot act on. Ask what proportion of it is user roles before you let it change your plan.

Treating postings as hires. This is the most consequential error in the list. A postings series counts adverts, and adverts overstate net headcount in three separate ways: a role that is re-advertised after a failed search is counted twice, agencies and job aggregators duplicate the same underlying vacancy across boards, and some listings exist to build a pipeline rather than to fill a seat. Meanwhile postings undercount in one important way — roles filled internally or from a network are never advertised at all, and core roles are disproportionately filled that way. So the true number of open builder jobs is smaller than the postings number in one direction and larger in another, and neither correction is knowable from the outside.

Optimising for a job title rather than a class of work. Titles are inconsistent across employers and across markets in a way that classes are not. Two adverts both titled "AI Engineer" can be a core role at a startup and an integration role at a services firm; two adverts titled "Software Engineer II" and "Machine Learning Scientist" can be the same job. Sort by the four questions, never by the title.

Assuming India and the UK reward the same signals. They reward the same underlying evidence with different emphasis. Indian reviewers, particularly at capability centres and services firms, tend to weight process discipline and behaviour at volume, because the system will be run by a rota of people for years. UK reviewers at startups tend to weight breadth and shipping speed, because one person will own a feature end to end. Same five sentences, different first sentence.

Reading one month as a trend. The UK series fell from 182,000 in 2022 to 112,000 in 2024 before recovering to around 180,000 in 2025. Anyone who read the 2024 trough as a structural verdict on AI hiring drew exactly the wrong conclusion. Composition moves slowly; volume moves fast and noisily.

Re-running the map

Set a reminder for six months. That interval is chosen deliberately: short enough that the composition of your local market will not have moved far, long enough that you are not reacting to a single employer's posting batch. Re-run the same scope — same city, same title family, same seniority band, same thirty-day window — because a comparison is only meaningful if the frame is identical, and keep every old tally sheet.

Three signals justify running it earlier. A visible funding or investment shift in your target sector, because funding cycles change the employer mix at the top of the tally within a quarter. A change you notice anecdotally in who is posting — a wave of new capability centres, or a consultancy that has suddenly started advertising core roles. And three months of applications producing a materially different response rate from what your map predicted, which usually means the map was drawn on the wrong scope rather than that the market changed.

What you are looking for on the second run is not the new totals. It is the movement between them. Did the core share of your local market rise or fall? Did core roles concentrate further into one employer type, or spread out? Did the three most repeated requirements change, and did the one you could not evidence last time appear more often or less? Those four comparisons are worth more than either map on its own, because the first run tells you what the market looks like and the second tells you which way it is going. Direction is what you position against.

None of this makes the market smaller or larger than it is. As of September 2026, the honest summary is that both India and the UK are adding AI roles quickly, that the great majority of the new ones are roles in which people use AI rather than build it, that the builder market is genuinely small in the UK and genuinely under-supplied in India, and that in both cases the thing that separates candidates is verifiable evidence of ownership rather than a title or a headline number. That is a workable position. It just requires reading the market rather than scrolling it.