Guides
Questions to Ask Before Hiring an AI Developer
Questions to ask before hiring an AI developer: scope, data, ownership, approvals, running costs, testing, exit and contract, with good answers.
IO Creative10 min read
The questions to ask before hiring an AI developer come down to eight topics: what the job is and how you will measure it, what data the AI sees, who owns the code and accounts, what needs your approval, what it costs to keep running, how it is tested, how you leave, and what the contract says. A good developer answers each one in plain words and in writing. A vague answer to any of them is your warning.
This guide gives the questions in that order. Each one says why it matters and what a good answer sounds like. Copy it into an email and send it to every developer you are considering.
1. Scope and success: what job is this, and how will we know it worked?
Ask: "What exactly will the AI do, for whom, and what does done look like?"
Why it matters: "Add AI to the business" is not a job. Without one task and one measure, you cannot tell a working project from an expensive demo.
A good answer: One task in one sentence, such as "draft replies to booking requests" or "answer the ten most common customer questions from our pages." Then a measure you can check: time saved per week, how many drafts a person had to correct, how often it said "I don't know." They ask what you do today, so there is a before to compare with. They are also willing to start with one small feature and prove it before adding a second. Our guide to adding AI to the tools you already use shows what a small first step looks like.
Red flag: A long feature list, no measure, or "it will learn your business."
2. Data: what does the AI see, where does it go, and who can delete it?
Ask: "Which of my files, pages, inboxes and customer details will the AI read, and which will it never see?"
Why it matters: The moment your content reaches an AI model, it is processed by another company. Your customers' details may be in it.
A good answer: A named list of what it reads (these folders, these pages, this inbox) and what is excluded. Then three more answers:
- Where it is stored. Which service holds the prompts, the answers and any copies, and in which account.
- Who can delete it. You can, on request, and the developer says how and how fast. Ask the same of every service in the chain.
- Whether it is used for training. This depends on the model provider and on the kind of account, so ask which provider and which plan will be used, then check the provider's own terms. As examples, here is what three providers say today:
- OpenAI says data sent to its API is not used to train or improve its models unless you explicitly opt in. It also says abuse monitoring logs are kept for up to 30 days by default, and that eligible customers can apply for controls that exclude content from those logs (OpenAI, Data controls in the OpenAI platform).
- Anthropic says that by default it does not use inputs or outputs from its commercial products, including the Anthropic API, to train its models. Feedback you choose to submit, or data you otherwise allow it to use, is an exception (Anthropic Privacy Center).
- Google says that for the Gemini API's Paid Services it does not use your prompts or responses to improve its products. For Unpaid Services, such as the free quota, it does use submitted content to improve and develop its products, and it tells you not to submit sensitive, confidential or personal information to them (Gemini API Additional Terms of Service).
The point is not that one provider is best. The point is that the answer depends on the plan, so a developer who tests your real customer data on a free tier has made a choice you did not approve.
Red flag: "Don't worry about the data," or a developer who cannot name the provider.
3. Ownership: who owns the code, the prompts and the accounts?
Ask: "When the project ends, what do I own, and whose name is on every account?"
Why it matters: In most places, whoever writes the code owns it until a signed document says otherwise. Paying an invoice does not move ownership. The same trap applies to the AI parts: the prompts and instructions, the connections to your tools, and the accounts that run it all.
A good answer:
- The finished code and prompts are assigned to you in writing, or licensed to you with no end date and no fee to leave.
- The model provider account, the hosting account and any API keys are in your name, with the developer added as a user. If you cannot log in to it, you do not control it.
- Anything the developer keeps, such as their own tools or templates, is listed, and you can run and move the project without it.
- Another developer could take the project over from what you are handed.
We cover the legal side in detail in who owns my website, and what to check before you sign. The same checks apply to AI work.
Red flag: Everything runs under the developer's own API key and account, and you are told you will "get access later."
4. Human approval: what can the AI do on its own?
Ask: "Which actions does the AI take alone, and which wait for a person to say yes?"
Why it matters: Reading and drafting are low risk. Sending an email, publishing, paying or deleting is not, because a wrong answer reaches a customer or cannot be undone.
A good answer: Sending, publishing, paying and deleting wait for your yes. The request shows you the exact content, where it came from, and lets you approve, edit or reject. Nothing goes out if you do not answer. Every action is logged, you can pause the whole thing at any time, and any step you later let run alone is your decision. Our guide to AI agent approval lists what to gate and how to keep approvals quick.
Red flag: "It runs fully automatically, you won't have to look at it."
5. Recurring costs: what does it cost every month after launch?
Ask: "What will I pay each month to keep this running, who is billed, and what happens if usage doubles?"
Why it matters: The build price is one cost. Most AI features also have running costs, and they are easy to leave out of a quote.
A good answer: A list of the recurring items, each with who pays:
- Model or API usage. Ask how the provider bills. If the charge follows usage, cost grows with volume. The developer should estimate it from your real numbers, such as messages per week, and say how to set a spending cap.
- Hosting. Where the feature runs, what that costs, and in whose account.
- Other services. Any search, storage, email or automation tool it depends on.
- Upkeep. Who fixes it when a model, a provider or one of your tools changes, and whether that is included or quoted separately.
Usage should be billed directly to you by the provider, on your account, so you see the real number. For the site side, see what a website maintenance plan should include.
Red flag: A single price with no mention of what happens next month, or usage costs marked up inside a bundled fee with no detail.
6. Testing: how is it checked, and how are errors caught?
Ask: "How will you test it, on what, and what happens when it gets something wrong after launch?"
Why it matters: AI can sound sure and be wrong. A test on invented examples proves little.
A good answer:
- It is tested on real questions or documents from your business, including ones it should not be able to answer, and the results are shared with you.
- It answers from your own content, shows the source of each answer, and says "I don't know" instead of guessing.
- It starts with a small group, and you watch the corrections for a few weeks before widening it.
- After launch, every action is logged in plain words, so when something goes wrong you can see what it did and why, and the cause gets fixed rather than the symptom.
- You have an easy way to report a bad answer, and someone who owns fixing it.
Red flag: "We tested it internally," with nothing to show you.
7. Rollback and exit: how do I undo it, or leave?
Ask: "If a release goes wrong, how fast can you restore the last working version? If I end the contract, how do I leave?"
Why it matters: Every system breaks at some point. What matters is whether you can undo a bad change in minutes, and whether you are stuck with one person.
A good answer: Before each release, the working version is recorded, so going back is one action. We describe the habit in how to roll back a website deploy. For exit, they commit to handing over the code, the prompts, the documentation and the accounts, built on common tools so another developer can take over. They tell you what it costs, ideally nothing, and how long it takes.
Red flag: Proprietary tools only the developer can run, or a handover fee that is not written down.
8. The contract: what should be in writing?
Ask: "Can I see the agreement before any work starts?"
Why it matters: Everything above is only a promise until it is written. If it is not in the contract, you are relying on goodwill.
A good answer: A written agreement, signed before work starts, that covers at least:
- Scope. The job from question 1, the one feature or task, and what is not included.
- Price and payment schedule. Fixed per project, or an estimate with a stated way to approve changes.
- Success measure. The test from question 1 that decides whether it is finished.
- Data. What the AI may access, which provider and plan it uses, and who can delete what.
- Ownership. Assignment or license of code and prompts, and accounts in your name.
- Approvals and logging. What needs your yes, and that actions are logged.
- Running costs. The recurring items from question 5 and who pays them.
- Testing and acceptance. What is tested, on what, and who signs off.
- Support after launch. Who fixes problems, how fast, and at what cost.
- Exit. What you receive if you end the contract, and when.
This is not legal advice. If the contract involves personal data, money at scale or regulated work, have a lawyer read it.
One-page checklist: questions to ask before hiring an AI developer
- The job is one sentence and has a measure I can check.
- I know what the AI reads and what it never sees.
- I know the provider and plan, and have read its data terms myself.
- I know where my data is stored and who can delete it.
- Code, prompts and accounts are mine, in writing.
- Sending, publishing, paying and deleting wait for my yes.
- I have a monthly cost estimate, and I know who is billed for usage and hosting.
- It was tested on my real questions, and every action is logged.
- A bad release can be rolled back, and I can leave with everything.
- Scope, price, payment schedule and exit are in a signed agreement.
How IO Creative approaches it
Our AI development service is built around the same points. It is small AI features inside the tools you already use, answering from your own content, tested on your real questions first. Accounts are in your name, and usage costs are paid directly by you. Pricing is per project, with a written agreement that sets the scope, price and payment schedule before work starts. Where an AI does a repeat task for you, our agentic AI service keeps you approving the steps that matter, with a log of what it did.
Your next step
Send the eight questions to the developers on your shortlist and compare the answers side by side. If you would like ours, start on our AI development page, then book a call or write to us.
