Start with the outcome, not the AI label
Define what should improve: response time, information retrieval, request classification or draft preparation. Without a metric, every demo can look intelligent.
A good freelancer also recognizes when rules, traditional search or a better interface are enough. They do not force generative models into every problem.
Ask how quality will be evaluated
The decisive question is how you will know it works. You need representative examples, acceptance criteria and a way to observe errors and regressions.
Non-deterministic output needs more than three happy-path prompts. Candidates should discuss evaluation sets, human review and known limitations.
Check data, costs and dependencies
Ask what data reaches external providers, how long it is retained and how secrets and personal information are protected.
Request a cost-per-operation estimate and a plan for changing provider or model. A cheap prototype becomes fragile when it depends on a single uncontrolled API.
Assess the ability to deliver the whole system
An AI project includes interface, backend, authorization, integrations, monitoring and exception handling. The model is only one component.
Portfolio and GitHub show technical depth, but interviews reveal communication, assumptions and willingness to say no. Trust grows from clarity about risk.