Questions to Ask Us Before Starting an AI Project

Why we publish our own interrogation list

Buyers of AI projects are often at a disadvantage. The supplier knows which questions reveal weak spots, and the buyer does not. That imbalance is how businesses end up with impressive pilots that cost more to run than they save, or systems they cannot move to another supplier.

So here are the questions we think you should ask us, with the answers we give. Use them on anyone else too. A general list for software suppliers is in questions to ask a software developer; this one is specific to AI.

Questions about whether AI is the right answer

• What would you build if AI were not allowed? If the answer is a simpler system that does most of the job, you should know that before paying for the AI version.
• What happens when the AI is wrong? Every AI system will be wrong sometimes. The design question is whether a mistake is caught before it matters.
• How often does this task really happen? A task that happens a few times a month rarely justifies a custom AI build.
• Is there an off-the-shelf product that does this? Sometimes there is, and it is cheaper.

Our answer to the first is often the most useful part of a discovery call. Plenty of briefs we see are mostly workflow automation with a small AI component, and pricing them as AI projects would not be honest.

Questions about data and privacy

Ask where your data goes, which providers process it and under which terms, whether it is used to train anyone's models, and where it is stored. Ask what personal information the system will handle and whether it can be redacted before reaching a third-party model.

Our answers: we use enterprise API terms that exclude training on your data, redact personal details where the task allows, offer self-hosted models when data cannot leave your infrastructure, and document the data flow for your data protection lead. If a supplier cannot draw you a simple diagram of where your data travels, that is worth pausing on.

Questions about quality and how it is measured

1. How will we know the system works, and on what cases will you test it?
2. What is our current baseline, and who measures it?
3. What result at the proof-of-concept stage would make you recommend stopping?
4. How will quality be tracked after launch?
5. When you change a prompt or model, how do you check nothing got worse?

The answer you want involves an evaluation set built from a few hundred of your real cases, agreed before building. Any supplier quoting an accuracy percentage before seeing your data is guessing.

The most revealing question is the third one. A supplier who cannot name a result that would make them stop has already decided the project will continue.

Questions about cost, including running cost

AI projects have two price tags: the build and the monthly running cost. Ask for both, and ask how the running cost scales with volume.

At SpiderHunts the build is scoped and fixed-price before work starts, and we estimate running cost per task during the proof of concept using your real inputs, which is when that estimate becomes trustworthy.

Questions about ownership and leaving

Ask who owns the code, the prompts, the evaluation data, the trained models and the cloud and API accounts. Ask what another team would need to take over. With us, the answers are that you own all of it from the first commit, the accounts sit in your name, and the handover includes documentation written for someone who was not involved.

Ask it anyway. Some suppliers answer 'you own the code' and keep the prompts, which in an AI system can be the part that matters most.

Questions about people and life after launch

• Who exactly will work on the project, and will they be the people on this call?
• Where is the team based, and how much working-day overlap will we have?
• What does the warranty cover, and for how long?
• What happens when the model provider changes or retires the model we use?
• Who do we call when something looks wrong on a Monday morning?

For the record: our offices are in London and Lahore, the engineers you meet during discovery work on the project, and a 90-day warranty covers defects after launch, with retainers available after that. More on the first conversation is in what a discovery call with SpiderHunts covers, and our AI integration page describes the delivery stages.

Answers that should make you cautious

From any supplier, including us: guaranteed accuracy figures before seeing data, a running-cost estimate of 'very little' with no calculation, reluctance to name the model provider, an evaluation plan that consists of 'user acceptance testing', and ownership terms that exclude prompts or configuration. None of these proves bad intent. All of them are worth a follow-up question.

Frequently asked questions