You have a first call booked with an AI consultant. What should you ask, so you come away knowing whether they understand your problem or are simply good at selling? This guide gives you the questions, grouped by what each one reveals, and describes what a strong answer and a weak answer sound like. There is a printable checklist at the end.
Key takeaways
- The best questions are about your problem, their evidence, their failures and what happens after handover.
- A strong answer is specific and in your terms: hours, cost per job, who checks what.
- The most revealing question is about the last time something went wrong.
- Ask data questions plainly, then take the answers to a qualified adviser.
- Keep detailed contract questions for the proposal stage.
What should I ask in the first call, and what can wait?
The first call is for finding out whether they understand your problem and have done similar work. Keep it to the questions in the first three groups below. Detailed questions about price, contract terms, ownership and exit belong at the proposal stage, when there is something concrete to question. Asking them too early gets you general answers.
Have they understood my problem, or are they pitching a tool?
"What would you need to know about our business before recommending anything?" A strong answer asks which task or decision the work should improve, what it costs you today and who will have to work differently. A weak answer names a product or platform before it has heard your problem.
"Why would something simpler not work?" A good adviser can explain why a template, a rule in your existing software or plain automation would not do the job, or will tell you it would. A weak answer treats AI as the starting point rather than one option.
"Can you put your value in my measures?" A strong answer talks about hours saved each week, cost per invoice or fewer quotes sent with mistakes. A weak answer stays with accuracy scores, features and the number of questions a chatbot can answer. Our guide on how your consultant will measure success goes further.
What evidence of past work is reasonable to ask for?
"Can I speak to a customer you did similar work for?" It is reasonable to ask for one live customer, on work like yours, whom you can call. A strong answer offers one readily. A weak answer offers only case studies they wrote themselves. When you speak to that customer, ask whether the results held up after the trial ended, not just whether the launch went well.
"Who will actually do the work?" Ask to meet them. The person who sells is not always the person who builds.
How to check what you are told is covered in how to check an AI consultant's credentials.
The surprising one: ask about failure
"Tell me about the last time the AI got something wrong. How was it caught?" This is the most revealing question you can ask. Someone who really delivers AI work has a concrete, slightly unglamorous story: what went wrong, who spotted it and what changed afterwards. A weak answer says it does not happen, or that their system needs no checking. No honest provider can promise flawless output.
"What happens when the inputs are messy or incomplete?" Your real emails, invoices and records are messier than any demo. A strong answer describes how errors are spotted, flagged and undone. A weak answer says the tool handles everything.
"Will your savings figure include the time my staff spend checking the output?" A strong answer says yes, and explains how it will be measured. Gross savings that ignore checking time make any tool look better than it is.
How will they use, store and hand back my business data?
Ask these plainly, and write down the answers:
- Where will our data be stored, and who will be able to see it?
- Will it be used to train or improve anything outside our project?
- Can we have a full copy back at any time, and can it be deleted when we finish?
A strong answer is clear and willing to put the details in writing. A vague or impatient one is a warning sign. We cannot tell you what the right answers are for your business: what the law asks of you depends on your data and how it is used. Before you share customer or staff records, take the consultant's answers to a data protection adviser.
Who owns what they build, and what happens when the project ends?
"Which changes will my team be able to make without you after handover?" A strong answer lists them: updating instructions, adding examples, adjusting settings, switching it off. A weak answer is "just call us".
"What will we have at the end?" Look for your data, the settings and instructions that make it work, plain-English notes and training for the people who will run it.
"How do you handle training and getting people to use it?" A tool nobody uses delivers nothing. A good adviser plans for this from the start.
Who owns the work in legal terms is a question for your contract, and for a solicitor if the project is large. Our guide to avoiding AI lock-in covers the practical side.
What does a good answer sound like compared with a vague one?
| Question | A strong answer sounds like | A vague answer sounds like |
|---|---|---|
| What do you need to know first? | "Walk me through how the task is done today." | "Our platform can do that." |
| Why not something simpler? | "A rule would cover most of it; AI helps with the rest." | "AI is the future." |
| What is your value? | "Fewer hours a week on invoice entry, measured before and after." | "Better accuracy and efficiency." |
| Evidence? | "Here is a customer you can ring." | "We cannot share client names." (and nothing else) |
| Last failure? | A specific story of an error and how it was caught | "Our system does not make mistakes." |
| After handover? | A list of changes your team can make | "We will look after it for you." |
A single vague answer is not a reason to walk away; everyone has a bad moment in a call. A pattern of them is.
Common questions
Is it rude to ask about failures?
No. Good consultants expect it, and many welcome it, because it lets them show real experience. Ask it politely and listen to the detail.
What if they will not name any customers?
Some work is confidential. A reasonable compromise is a conversation with a past customer who agrees to speak to you, even if their name never appears on a website. If there is no one at all, weigh that carefully.
Your first-call checklist
- What would you need to know about our business before recommending anything?
- Why would something simpler not work?
- Can you put your value in our measures?
- Can I speak to a customer you did similar work for?
- Who will actually do the work, and can I meet them?
- Tell me about the last time the AI got something wrong. How was it caught?
- What happens when our inputs are messy or incomplete?
- Will your savings include the time our staff spend checking?
- Where will our data be stored, who can see it, and can we have it back?
- Which changes will we be able to make without you after handover?
- How will you handle training and getting people to use it?
Download: first-call questions (Word)
Next in the series: the behaviours that should make you think twice, in red flags when hiring an AI consultant. Or go back to how to choose an AI consultant.