What 4,000 AI job adverts say about hiring in the UK and US

Job adverts give away more than most surveys do. Before the headcount reports and salary studies catch up, employers have already written down what they will pay for, and you can just read it back to them. So in early July 2026 we did. We pulled every AI and machine learning role we could find across the UK and US, cleaned out the noise, and counted.

What came back was 4,012 separate listings, taken from LinkedIn, Reed and Totaljobs over thirty days and cut down from more than 6,800 raw records once the duplicates were gone. None of it comes from an estimate or an outside report. It is all straight from the adverts, with the method at the bottom if you want to check our working.

How to read this. Every percentage here is the share of our own scraped adverts that mention a term, drawn from 4,012 distinct listings (2,258 in the US, 1,754 in the UK). We are not citing an outside survey; the numbers are counts we made ourselves, and the sample size behind each one is shown as you go. A single advert only proves a requirement is real in the wild, so the counts do the work and the links are there to let you see the raw listings for yourself. Example listings from the sample, captured on 7 July 2026 and subject to the usual churn as roles get filled: Data Scientist, Research at Google and Senior Machine Learning Software Engineer at Qualcomm in the US; Associate Director, Data Scientist at Fitch Solutions and Lead Software Engineer at Kainos in the UK. The full listing-level dataset is available on request.

Python still sets the floor

No prizes for the top skill. Python shows up in 78.3 per cent of listings, miles ahead of anything else. The surprise is the tier just below it. Communication and research both outrank most of the named technologies, and large language model experience has already climbed to third.

The soft skills are not padding. Communication appears in nearly two thirds of adverts, research in almost half. That fits a job where much of the work is framing the problem and checking the output, not typing out every line. Scala at 48.4 per cent tells a different story. A good chunk of this hiring is really data engineering with an AI label on it, because the roles near the big pipelines still live on the JVM.

Large language models are now the baseline

Two years ago, large language model experience was a specialist line on a CV. Now it is in 56 per cent of listings (2,247 of 4,012). More than half of roles simply take it for granted that you have worked with these models. It has stopped being a nice extra and turned into the price of entry, and some employers now put it right in the job title, like this AI and LLM engineer role.

Under that headline is a whole layer of tooling that barely existed in job specs eighteen months ago. Retrieval-augmented generation, which grounds a model in your own documents, now turns up in nearly one advert in four (907 of 4,012), including this AI transformation advisory role. Vector databases and orchestration tools like LangChain are not far behind, and adapter-based fine-tuning has quietly become a standard line item.

The speed of it caught us off guard. A whole category has appeared almost from nothing. Vector databases, meaning named products like Pinecone, Weaviate, Chroma and Qdrant, are in 15.6 per cent of listings. The OpenAI API, one company's specific interface, is named as a hard requirement in one advert in eight. You rarely see a single supplier written straight into job specs like that, and it shows how concentrated the tooling has become.

The framework question has settled

Teams spent years arguing over which deep learning framework to standardise on. The adverts say that fight is more or less over. PyTorch appears in 24.1 per cent of listings to TensorFlow's 16.7, and the lead holds on both sides of the Atlantic.

The Atlantic gap: the US wants the newer stack, and pays for it

Split the data by country and the US pulls ahead on almost every model-related skill. Generative AI demand there leads the UK by eleven percentage points, and the same gap shows up in retrieval, vector databases and orchestration tools. It is not the fundamentals that separate them. Python demand is close on both sides. The difference is the newer layer sitting on top of the models.

The pay gap is wider again. Where a salary was disclosed, the median advertised base was 140,000 dollars in the US and 76,000 pounds in the UK. Convert roughly and you are still tens of thousands apart, and the real gap could be bigger given how rarely UK employers show a number at all.

That disclosure rate is a finding in itself. Half of US adverts name a salary. Only a quarter of UK ones do. The UK sample also leans towards employers willing to publish pay, so the true British median probably sits below the figure we can see. If you are benchmarking a UK offer against these numbers, treat them as a floor, not a midpoint.

Working patterns diverge

The two markets also set up work differently. Hybrid is the UK norm at 38.5 per cent of listings, while US roles lean on-site or simply do not say. Fully remote is a minority on both sides, and, against the usual story, it is a touch more common in the US than the UK.

Who is hiring

The employer mix explains much of the regional gap. At the top of the US table you get big tech and AI-native labs sitting next to defence and consulting names. In the UK it is financial services, recruitment aggregators and consultancies, with a thinner band of AI-native firms.

The staffing and data-labelling firms high on the UK list are worth a second look. A lot of that volume is aggregators reposting the same roles rather than hiring directly. That is exactly why the raw counts had to be deduplicated before any of these numbers could be trusted.

What it means

Taken together, the adverts describe a market that has grown up faster than the headlines let on. The base skill has not moved. Python is still the common language. What has changed is everything stacked above it: large language models, retrieval and the plumbing around them have gone from research curiosities to standing requirements in a large share of roles. The US is about a year ahead of the UK in asking for the newest parts of that stack.

If you are building a team, the practical read is fairly simple. Hire and train for LLM and retrieval experience as a baseline, not as a rare specialism you pay over the odds for. Assume the tooling around models, the vector stores, the orchestration, the fine-tuning, will keep landing in job specs, because that is where the day-to-day work now is. And if you are hiring in the UK, remember the US market is publishing pay your candidates can see, which sets a benchmark whether you like it or not.

Method

The snapshot was taken on 7 July 2026. We collected listings from LinkedIn in both regions, and from Reed and Totaljobs in the United Kingdom, restricted to the previous thirty days. Starting from 6,823 raw records, we filtered to AI and machine learning roles and then deduplicated on an employer-and-title fingerprint within each region, stripping seniority prefixes so that a senior and a lead version of the same posting did not count twice. That left 4,012 distinct listings: 2,258 in the United States and 1,754 in the United Kingdom.

Skill counts come from matching a fixed vocabulary against each advert's text, so a percentage means the share of listings that mention a term, not a formal requirement weighting. Salary figures use the advertised minimum where one was disclosed, which is why coverage varies by region and board. These are counts from live adverts on three job boards over one month, not a census of the whole labour market, and they should be read as a directional snapshot rather than a definitive measure.