The short answer
To hire an AI engineer, work out which single requirement on your spec is genuinely scarce. Most AI engineers have shipped RAG, vector search and document ingestion. Far fewer have shipped natural language to SQL in production. That one rare requirement, not your salary band, sets your timeline.
I ran seven detailed conversations against one AI engineer brief last week. France, Norway, Portugal, Bulgaria, Greece, the Netherlands.
By the fourth call I could have written the CVs myself.
That is not a criticism of the people. Every one of them was strong. It is a description of the market right now, and it changes how you should write the spec.
Why Does Every AI Engineer CV Look The Same?
Because the tooling settled.
All seven had built retrieval augmented generation in production, worked with embeddings and vector search, ingested messy documents and PDFs and scanned files needing OCR, and built agentic workflows and tool calling. Most had built a model gateway so they could route between providers on cost, latency and quality.
Three years ago that list was a differentiator. Now it is the job, in the same way that a backend engineer writes APIs.
So when a founder sends me a spec with those six things on it and says “we need someone who has done all this,” the honest answer is that most of the market has. The spec is not selecting anyone. It is describing the category.
There is a second reason the paper starts to blur, and it is less comfortable.
Some of the people on the other end of the call are not who the CV says they are. I have had a video call where the person in front of me was plainly not the person on the profile. I have had others where answers arrived with an odd delay, in a register that did not match how the person spoke for the rest of the conversation, and where a follow up question that should have been trivial landed badly. I cannot tell you for certain what was happening on those calls. I can tell you I ended them early.
I am not raising this to make anyone paranoid, and it has nothing to do with where people are based. It is an argument for talking to people properly and early. A CV cannot be verified. A thirty minute conversation about a system someone claims to have built can be, because the second and third follow up questions are very hard to fake.
So the paper is doing even less work than it looks. That is still a good problem, because it means the pool is bigger than you fear. It just means the spec and the screen both need one more pass.
Which Requirement On Your Spec Is Actually Rare?
Here is what separated the seven.
Natural language to SQL in production. Three of the seven had genuinely shipped it to users. One with small open source models running locally, because the customers were in the DACH region and nothing could leave the building. One had built the GraphQL equivalent alongside it. One had done it inside an air gapped enterprise environment.
The other four had not. Two had tested it in a sandbox. All four were confident they could learn it. But they had not done it, and they said so straight away.
That is the shape of almost every AI spec I read at the moment. Five or six commodity requirements and one that roughly a third of the market has actually done. Writing the brief properly is mostly the work of finding which one it is, and everything downstream gets easier once you know.
Why Does The Rare Requirement Set Your Timeline?
Because the people who have it are not sitting still.
Of the three who had shipped natural language to SQL, one was immediately available, already at final stage with two other businesses, and expecting to sign something within a fortnight. Another was actively interviewing with a month’s notice. The third had no other processes running at all, was not applying anywhere, and only took the call because someone reached out directly.
Read that again, because it is the whole point. The people with the scarce skill are either already in a race you have not entered, or they are not looking and will never see your advert. Neither group is reachable by posting a role and waiting.
If your rare requirement is held by a third of a small pool, and half of that third is mid process somewhere else, your window is measured in weeks. A two month interview loop does not lose to a better offer. It loses to an earlier one. That is also why we work the way we do on senior searches, going direct rather than waiting for applications.
Does Your Offer Structure Cost You More Than Your Salary Band?
This surprised me and it is worth passing on.
Among the seven who reached a call, compensation expectations clustered tightly, roughly 85,000 to 110,000 euros on a business to business basis across six countries. But that tight band is an artefact of the screen, not a picture of the market. Earlier in the same search I spoke to people on the same brief who wanted up to 150,000 euros on a business to business basis.
Worth saying those are all contractor numbers. An employed salary through an entity in someone’s own country sits lower for the same take home, because the tax and social contributions move from them to you. If you hold a B2B rate up against a salaried budget line, you will frighten yourself for no reason.
That gap is worth sitting with. Your range does not filter your shortlist. It filters who ever reaches you, weeks before you see a CV. If you set a band and hear very little back, the market has not rejected you. The sorting happened upstream, and you never met the half it removed.
Among the people who did come through, though, the band was not the problem. The structure was.
One candidate’s only real hesitation was fully remote working. He had done two days in the office for years and wanted to think properly about whether he would be happy without that. Money never came up. Another asked how a Greek company invoices a UK business, and whether payment would arrive monthly and in euros. A third asked whether VAT sat on top of his figure before he quoted. A fourth needed to check whether the micro company he runs for teaching work could legally be used.
These are not awkward questions. They are people working out whether they can actually say yes. Every one of them came before the technical conversation.
If you are hiring across Europe on a contractor basis, the mechanics are part of the offer. Knowing how you pay, in which currency, on what day, and what happens about equipment, removes more friction than another five thousand on the number. Engineers weigh the whole picture, not just the headline.
How Do You Write An AI Engineer Spec That Works?
Four things.
Separate the list. Commodity requirements in one group, stated as expected. The scarce one on its own, named as the thing you care most about. Candidates self select honestly against it, which saves everybody a fortnight.
Decide whether the scarce thing is a must or a strong want. If four in seven strong engineers have not built it and are confident they can, you may be looking at a four week ramp rather than a hard filter. That decision changes your pool enormously, and it is yours to make.
Set the process to the speed of the scarce group, not your calendar. Three stages in ten days beats three stages in six weeks, and it is the same three stages.
Write the commercial mechanics down before the first call. Currency, invoicing, payment date, equipment, holiday, equity. It takes an hour and answers the questions people are holding back.
Questions this search answers
What skills should an AI engineer have in 2026?
Retrieval augmented generation, embeddings and vector search, document and PDF ingestion including OCR, agentic workflows and tool calling, model evaluation and guardrails, and multi provider model routing. Treat these as the baseline rather than the differentiator. In seven detailed conversations last week, every candidate had shipped all of them in production.
How long does it take to hire an AI engineer?
Faster than most founders plan for. The engineers holding genuinely scarce skills are frequently at final stage with other businesses and working to a two week decision, not a two month one. Build a process of roughly ten days across three stages, and be ready to move when the right person appears rather than batching candidates.
How much does an AI engineer cost in Europe?
Expectations across seven conversations in six countries clustered between roughly 85,000 and 110,000 euros on a business to business basis, with earlier screening reaching 150,000. Those are contractor figures, often quoted before VAT. An employed salary through an entity in the engineer’s own country sits lower for the same take home, because tax and social contributions shift to you.
How do I know an AI engineer is who they say they are?
Talk to them early and ask follow up questions about systems they claim to have built. I have had a call where the person on video was not the person on the profile, and others where answers arrived with an odd delay and a simple follow up landed badly. A CV cannot be verified. A detailed conversation about a design decision is very difficult to fake.
Why can I not find AI engineers with the right experience?
Usually the spec mixes commodity requirements with one genuinely scarce one, and the scarce one is doing all the filtering invisibly. Identify which single requirement roughly a third of the market has actually shipped, then decide whether it is a must have or a four week ramp. That decision changes your pool more than your salary band does.
Most AI hiring problems are not talent shortages. They are one line on a spec that nobody has separated out yet, and if you can name that line you can plan around it.
