An “AI engineer” can mean someone who builds AI-powered product features, someone who owns training pipelines, or someone who optimises inference. Those are different jobs. Before searching for candidates, write down the work your team needs to complete.
1. Define the first meaningful outcome
Replace “build our AI capability” with a concrete outcome. For example: deliver a document-search feature that can be evaluated against representative customer questions, with an agreed latency budget and a way to investigate poor answers. This gives you something to assess beyond enthusiasm for AI.
List the constraints alongside the outcome: data access, deployment environment, quality requirements, operating cost and the people available to support the hire. A role can look very different when the new engineer is joining a platform team rather than becoming the first technical employee.
2. Choose the role around the bottleneck
| Main challenge | Experience to investigate |
|---|---|
| Shipping an AI product feature | Product engineering, evaluation and integration |
| Running ML reliably | Data pipelines, deployment, monitoring and operations |
| Improving inference performance | Profiling, serving systems and hardware constraints |
| Developing new modelling approaches | Experiment design, research judgement and implementation |
These are starting points, not rigid job titles. One person may span several areas, but asking for depth in all of them can make the brief unrealistic. Decide which capability is essential on day one.
3. Use a production discussion
Ask the candidate to walk through a system they helped build. What was the objective? What did they personally decide? How did they measure improvement? What failed, and what changed as a result? Allow anonymised examples so they do not need to disclose confidential information.
Then introduce a change: traffic grows, the data distribution shifts, or the quality target becomes stricter. Look for a reasoned investigation plan rather than a memorised stack. A strong answer should expose assumptions and explain how the candidate would gather evidence.
4. Make the assessment small and relevant
Use a bounded task based on representative work, with a shared rubric. For an applied role, that might be reviewing an evaluation plan and explaining missing failure cases. For infrastructure, it could be interpreting a simplified performance profile. Avoid asking candidates to complete unpaid production work.
Tell candidates what is being assessed, how much time to spend and which tools they may use. Offer an accessible alternative when appropriate. Score the reasoning and evidence consistently, rather than rewarding presentation polish alone.
5. Align before you make an offer
Confirm ownership, reporting line, working arrangements and compensation expectations early. Be candid about research uncertainty, technical debt and the support the engineer will have. A precise role is easier for both sides to evaluate.
Use our technical interview scorecard to turn the brief into a consistent assessment, or explore our AI and ML recruitment focus.