What hirers actually check before booking an AI specialist
6 minute read. Updated 2026-08-08.
Hirers spend less time on your CV than you think and more on 3 things: what a referee says when asked what went wrong, whether you ask about data ownership and failure modes before proposing an architecture, and whether your day rate is defensible against the advertised range for your role. Certifications and polished decks get you shortlisted. Specific references and sharp early questions get you booked.
The CV is a filter, not the decision
Most candidates spend their polishing time on the document least likely to get them booked: a CV with certification logos, a GitHub profile of green squares, a summary written in the third person. That gets you past the first filter. It doesn't answer the question the hirer is actually trying to answer, which is whether you'll make good judgement calls once you're inside their systems and they can't check your working in real time.
A CTO or head of data at a company somewhere between 50 and 2,000 people has usually been burned once already, often by someone who looked exactly right on paper: strong logos, fluent in the vocabulary, confident in the pitch. What they check for next time is different from what they checked for the first time.
What they actually ask referees
The reference call a serious hirer runs isn't 'would you work with them again', which almost everyone answers yes to out of politeness. It's specific: what did this person get wrong, and what did they do when it went wrong. A referee who can only produce generic praise counts as a weak signal, whichever way the praise points.
If your reference list is 3 people who'll say you're great and nothing else, it won't survive that call. Build a shorter list of people who watched you handle something going badly, and brief them that you'd rather they were candid about a real decision than warm about nothing in particular.
The scoping conversation is the real interview
Before a hirer books anyone, they watch how that person handles the first conversation about the problem. Do you ask about the data before estimating anything. Do you ask who owns the label definitions, who signs off on the model going live, what happens if it's wrong for one customer rather than on average. Candidates who go straight to a proposed architecture and a day count get marked down even when the architecture is sound.
This is where a written scoping habit earns its keep, and it's worth building deliberately rather than improvising it each time on a call; the case for turning it into a paid diagnostic phase is really the same discipline made saleable rather than given away for free.
A hirer weighing 2 similarly qualified candidates will usually book the one who asked 3 sharp questions about their data over the one who produced a slicker slide deck about method.
Rate as a signal, not just a number
Hirers benchmark. Advertised UK contract rates over the six months to August 2026 put the median for Machine Learning Engineer at £575 a day, Data Scientist at £600, Generative AI work at £550. A quote wildly above the 90th percentile for the role, £738 for Machine Learning Engineer, £788 for Data Scientist, invites a question about why. A quote near the 10th percentile, £500 and £425 respectively, invites a different one: is this person available because nobody else wants them, or because they don't understand what the work is worth.
Neither figure is what anyone actually gets paid; these are advertised figures, and the gap between advertised and agreed tends to widen the higher you go. A hirer who has done this before knows that, and reads your rate as one data point among several, not as the whole case. Being unable to explain your own number, though, is a worse signal than the number itself.
Whether you know what you're being hired to be
Nothing is published in the UK advertised market for AI governance, AI risk and compliance, or AI safety and evaluation as job titles. That absence isn't a gap in the data; it means this work is routinely bought under other titles, usually Data Scientist or Machine Learning Engineer, by hirers who need someone who can also speak to risk and regulation and haven't found a clean way to advertise for it.
A hirer in that position is checking whether you notice the mismatch between the title on the brief and the work inside it, and whether you can talk about the EU AI Act's staged application or UK GDPR obligations without reaching for a specific figure you don't actually hold. Candidates who bluff a threshold or a penalty amount get caught out fast by anyone who's spoken to their own legal team recently.
The candidates who get booked for this kind of work are the ones who say plainly that the title undersells what's being asked of them, and price and scope accordingly, rather than pretending the mismatch isn't there.
What to do about it
- Spend less polishing time on the CV and more on a reference list that will say something specific.
- Brief referees to talk about a real decision, including one that went wrong, not just to praise you.
- Lead every first conversation with questions about data ownership and failure modes, not architecture.
- Be ready to explain your day rate against the advertised range for your role, not just state it.
- If the title on the brief is governance or risk work wearing a data science label, say so and price accordingly.
- Never invent a specific legal figure or threshold to sound more credible; being wrong is worse than being unsure.
Questions people also ask
Does a strong portfolio still matter if the reference call is what really counts?
Yes, but it does a different job. A portfolio gets you shortlisted; a reference call gets you booked. Treat the portfolio as proof you can execute and the reference as proof of how you behave when execution gets hard, and don't expect one to substitute for the other. There's a separate case for building portfolio evidence deliberately when your production work can't be shown.
Should I quote at the median for my role or push higher?
Quote where your actual risk and scope justify it, and be ready to say why in one sentence. Pushing toward the 90th percentile without a specific reason, named integration risk or scarce domain knowledge, reads as confidence a hirer can't verify. Pushing below the 10th percentile reads as availability a hirer should be suspicious of. The number matters less than whether you can defend it.
What if I genuinely don't have a referee who'd talk about a failure?
That's worth fixing before your next search, not glossing over. Ask a past client or colleague directly whether they'd be willing to speak candidly about a decision that didn't land, and reassure them it reflects well on you to have handled it visibly. If you truly have no such story yet, that itself is worth being honest about rather than inventing one.
How do I bring up that a brief is underspecified without sounding like I'm criticising the hirer?
Frame it as a description of the market, not a criticism of their brief: titles like AI governance or AI risk aren't published roles yet, so most companies write the brief they know how to write and expect the specialist to fill the gap. Say what you'd actually be doing, given what you can see of the problem, and let them adjust the title or scope from there.
Where the figures come from
Every rate and salary quoted in this article is a median or percentile of figures advertised in UK job postings over the six months to 8 August 2026. They are not rates paid, and the gap widens at the top of a range.
- IT Jobs Watch, UK contract rates, 6 months to 8 August 2026, read 2026-08-08.
- IT Jobs Watch, UK permanent salaries, 6 months to 8 August 2026, read 2026-08-08.