Hiring AI profiles in 2026: what to look for beyond the CV

You open the role, post the ad and within a few days you have a pile of CVs that all say roughly the same thing: LLMs, RAG, agents, prompt engineering. Everyone seems to know AI. Then you start interviewing and find that behind those words are profiles with almost nothing in common, plus a few people who have done a couple of tutorials.

It is one of the conversations we are having most often this year with technology companies. The request usually arrives as “we need someone for AI”. Our first question is always the same. To do what?

“Someone for AI” is six different jobs

The most sought-after right now is the AI engineer, the person who builds products on top of existing models. They integrate APIs, build retrieval (RAG) systems or agents, evaluate them and get them into production. Their background is nearly always software engineering, and that matters a great deal, as we will see.

A machine learning engineer, by contrast, trains, fine-tunes and deploys the company’s own models, and is at home with data pipelines, training, metrics and model serving. Alongside them you often find an MLOps or ML platform engineer, who looks after the infrastructure: deployment, monitoring, data and model versioning, cost. A data scientist analyses data, designs experiments and builds models to answer business questions. A research scientist works at the frontier, often publishing, and honestly makes sense in very few companies. Finally there is the AI product manager, who decides which problems are worth solving with AI, how success is measured and what to do about the model’s limitations from the user’s point of view.

More than once we have seen a company convinced it needed a research profile when what it was missing was a strong AI engineer. Almost as often, a company looking for an AI engineer first needed someone to sort out its data. Getting this straight before you go to market saves you months.

What really sets a good AI engineer apart

It is not knowing this month’s favourite tool. An LLM system is still software, and it needs tests, error handling, observability and a proper deployment. Without that foundation, prototypes stay as demos and never make it to production.

In our experience, what separates the good ones most is an evaluation mindset. Ask them how they know whether a change to the prompt or the model makes the system better or worse. Someone who has done this seriously will talk about the evaluation sets they built. Someone who has not will tell you they tried a few examples and it “seemed better”.

Judgement on cost and latency counts too: choosing the model, caching, cutting calls, or realising that a small model is enough. Those choices show up directly on the bottom line. Add an awareness of risks such as prompt injection, data leakage and made-up answers, and how to mitigate them. And one thing we value highly is knowing when not to use AI. A good engineer will sometimes suggest a simple rule or a classic search instead, and be right.

ML and product roles

With machine learning engineers and data scientists, start with the basics. Overfitting, leakage between training and test data, metrics that make sense for the problem at hand. Then look for experience with real, messy data rather than the usual sample datasets, and the ability to get a model into production and keep it healthy by watching for drift, retraining and versioning. Do not underestimate communication either. Being able to explain a result to someone non-technical, uncertainty included, is what gets the work used.

For an AI product manager we take the usual product skills as a given and look for more on top. Do they understand models well enough to know what is feasible? Can they define quality metrics that go beyond “users like it”? Have they thought about the experience for when the model gets it wrong, with human review, warnings or fallbacks? And do they know the EU AI Act well enough to recognise when to bring in the legal team?

Questions you cannot answer with vocabulary

These are some of the ones that work best for us in interviews:

• “Tell me about an AI system you took into production. What went wrong in the first few weeks?”
• “How did you measure answer quality? What did you do when the metric and user feedback disagreed?”
• “When have you decided not to use a model, or to switch to a smaller one?”
• “What did each request cost, and how did you bring that down?”

Pay less attention to what they say than to how they say it. Answers that stay generic or turn into a list of tools make us wary. Good answers come with rough numbers, specific problems and decisions along with their trade-offs.

If you want to add a practical exercise, keep it short and close to the real work. For an AI engineer we like this format: give them a small set of documents and some sample questions, ask them to build a first version of a system that answers them and, above all, to propose how they would evaluate it. The follow-up conversation about what they would improve and how they would measure it tells you far more than the output. Say upfront whether they can use AI assistants during the exercise. For most roles the realistic choice is to allow it and watch how they use them.

The signals that make us slow down

Some patterns come up again and again. CVs with very long lists of tools and no examples of impact. Impressive demos with no way of measuring whether they actually work. A thin engineering background for roles that are all about shipping systems to production. And difficulty talking about the limitations of models, which usually means the person has never had to wrestle with them in a real setting.

None of these rules anyone out on its own, but when several turn up together it is worth digging deeper.

Before the job ad goes live

If you keep only one idea from this, make it this one: write down, in a single sentence, what this person will have built six months from now. If you cannot, you do not yet know which profile you are looking for, and no interview will fix that.

That exercise is exactly where we start when a company asks us for help with an AI role. If that is where you are right now, tell us what you are working on.