Eleven employers in a crawl of 1.1 million British careers pages put “applied AI” into a job title. That is a rounding error, and the most informative number here: almost nobody advertises this role, so most people holding it acquired the description after being hired.
The word that matters is applied, and it is usually pasted on without thought — a seniority flourish, or a way of making a data science vacancy sound current. Taken seriously it is a complete specification, and it binds the employer more than the candidate.
The word doing all the work in the title
An applied AI scientist is judged on whether a result reaches a product. Not on whether it is novel, not on whether the method is elegant, not on whether a write-up exists. Reaching a product means a user experiences the difference, and it survives real traffic.
Most of the distance between an interesting result and a shipped one is not scientific work at all. It is data access, latency, cost per call, monitoring, and a negotiation with whoever owns the surface the model changes. A scientist can do some of that. None does all of it alone.
Before writing the brief, answer two things on paper: which production system is this person’s work supposed to change, and who walks it there with them? If either answer is vague, you are not hiring an applied AI scientist. You are hiring a researcher and hoping.
Two disciplines that share a CV and not a purpose
A research scientist and an applied scientist look identical on paper — same degrees, same libraries, often the same supervisor. What separates them is what each treats as a finished piece of work.
The research scientist finishes when a claim is established and defensible, and the right answer to a soft result is more careful work. The applied scientist finishes when the result is in front of users, which means calling a model four points behind the published state of the art done, because those points cost six weeks nobody has.
Neither disposition is superior. But a team with a product roadmap will be miserable with the first, and a real research group exasperated by the second. Decide which you are before interviewing, because both answer the same advert.
If you want somebody to take published methods and make them run reliably at your scale, the AI research engineer is a closer and cheaper fit. If the problems are tabular or forecasting rather than frontier model behaviour, the data scientist guide describes what plenty of these vacancies should have advertised.
Ask about the result that never reached anyone
One question sorts this pool better than any technical exercise. Tell me about a piece of research you did that never shipped, and why.
People who have worked at the boundary between a model and a product answer at once and without defensiveness. You get specifics: inference cost was three times what the feature could carry; the gain vanished on newer data; the team owning the endpoint had a freeze, and by the time it lifted the requirement had changed. Those are not confessions but the texture of the work.
Candidates measured only on output find the question awkward. Some redirect to something that shipped; some put the failure in another department and stop. That is not a character defect — it reads as never having owned the last mile, the thing being bought.
The follow-up that matters more than the answer
Ask what they would do differently, and listen for whether the change is scientific or organisational. Strong answers are almost always organisational: I would have had the cost ceiling and the data agreement before building anything, and found the owning engineer in week one rather than week nine. An answer that stays inside the modelling belongs to somebody who watched the failure rather than absorbed it.
If there is nobody downstream, do not make this hire
The precondition is structural rather than technical. An applied AI scientist needs a named engineer, or a small team, whose allocated job includes turning their output into a service. Not goodwill. Not a colleague with a roadmap of their own.
Without that, the pattern is quick and predictable. Good work accumulates in a repository. The second quarter goes on building deployment scaffolding, badly. Somewhere in the third they conclude the organisation is not serious and leave, usually for a lab. The post-mortem blames the hire. The hire was fine; there was nowhere for the work to go.
If no such engineer exists, hire that person first. A machine learning engineer in post before the scientist arrives is the better sequence, and a cheaper mistake to correct.
Papers prove capability and say nothing about usefulness
Publications are the one externally verifiable thing here, and they get weighed badly in both directions.
Over-valuing means reading a strong venue record as a proxy for impact. It is evidence that somebody can frame a question, run an honest experiment, survive hostile review and write clearly — rare abilities, none of which is the ability to get something live, and one or two of which cut against it.
Dismissing means declaring academic work irrelevant to a commercial team. Also wrong: experimental discipline cannot be installed in a quarter, and its absence is why so many internal AI projects report gains that evaporate on contact with next month’s data.
Read the record as capability, then test usefulness separately — the shipping question above, plus a live problem of your own where the right answer is a small model, a rule, or nothing. What you want is both: something citable and something released.
Money, and the two markets you are bidding against
Indicatively £80,000 to £130,000 permanent for a doctorate plus two or three years of industrial shipping, higher where the work sits close to revenue. Contract engagements are unusual and priced as such: indicatively £650 to £900 a day.
The awkward part is bidding in two markets at once. Industry labs pay above that range and add compute, peers and permission to publish. Universities pay well below it and offer autonomy and the freedom to follow one question for years. You will lose candidates to each.
What you can offer is what neither does well: a short, unobstructed route from an idea to something a user touches, and a named person to walk it with. Say that in the advert, with the product named. It is the pitch that moves people out of labs — and only true if you did the work in the first section.