Almost anyone can get a language model to produce a convincing answer in a demo. Far fewer people can make that answer reliable inside your product, on your data, at a cost you can predict. These questions help you tell the difference before you sign anything.
1. Can you show me something running?
Slides are cheap. Ask for a working example, even a small one, and ask what is real and what is scripted. A good engineer will tell you plainly which parts are prototype, because they know the difference.
2. How will we know it works?
AI features fail quietly. The answer you want involves an evaluation set: a fixed list of real inputs with known good outputs that is re-run whenever the prompt, model or data changes. If the plan is 'we'll try it and see', expect regressions you discover from customers.
3. What happens when the model is wrong?
It will be wrong sometimes. Ask about fallbacks, confidence checks, human review for high-stakes actions, and what the user sees when the system declines to answer. Designing the failure path is most of the work.
4. Where does our data go?
Ask which providers see your data, whether it is retained or used for training, where keys are stored, and what the minimum data needed actually is. For anything sensitive, ask for read-only access and an allow-list of what the system can touch.
5. What will it cost to run each month?
Model calls are a variable cost. A reasonable answer includes an estimate per request, spend caps, caching where possible, and a plan for what happens if usage spikes.
6. Who owns the code, prompts and data?
You should. Confirm in writing that code, prompts, configuration and any fine-tuned assets are handed over, along with documentation your team can follow without the freelancer.
7. How do you scope and price the work?
Prefer fixed milestones agreed in writing, with something you can review at the end of each. Be cautious of open-ended hourly work on a problem nobody has scoped yet, and consider a small paid pilot first.
8. What would you say no to?
Someone who takes every project is guessing. A good answer names things they would refer elsewhere, and explains why. It is a sign they will also tell you when your idea needs changing.
If you would like a second pair of eyes on a plan or a proposal, that is exactly what our free audit is for.