AI strategy checklist: 20 questions before you spend a euro
Twenty questions that turn AI enthusiasm into a plan: what problem, whose data, who owns it, what it costs and what could go wrong.
The short answer
Twenty questions turn AI enthusiasm into something fundable. They cover six areas: the problem, the data, the people, the money, the risk and the measurement. Answered honestly with the right four people in the room, they take an afternoon, and they almost always change what gets built first and make it smaller. The questions most often skipped are the baseline, the owner and the exception path, and skipping those three explains most of the AI projects that are quietly abandoned after six months. Where an answer does not exist, it names the work to do first, measuring, naming, deciding, which takes days rather than months. Any use that cannot answer the risk and measurement questions should not be funded yet, however impressive the demonstration was.
The twenty questions
| # | Area | Question | Good answer |
|---|---|---|---|
| 1 | Problem | What specific problem does this solve, in one sentence? | A process, not a capability |
| 2 | Problem | How is it handled today, and how many hours does that take? | A measured number |
| 3 | Problem | How often does it happen and by whom? | Frequent, by named people |
| 4 | Problem | What does it cost in errors, delays or lost work? | An estimate with reasoning |
| 5 | Problem | Why now rather than next year? | A reason beyond availability |
| 6 | Problem | What are we not doing instead? | An explicit trade-off |
| 7 | Data | What data does it need and where does it live? | Known systems, named owners |
| 8 | Data | Is that data complete, current and consistent enough? | Checked, not assumed |
| 9 | Data | Does it include personal or sensitive data, and under what basis? | Reviewed with your adviser |
| 10 | Data | Which provider and terms will process it? | Business terms, training excluded, region known |
| 11 | People | Who owns this after launch? | A named, willing person |
| 12 | People | Whose work changes, and have they been involved? | Involved from the start |
| 13 | People | Who handles what it cannot? | Named, with time allocated |
| 14 | Money | What is the build cost and what is the yearly running cost? | Both, including usage and maintenance |
| 15 | Money | What is the cost per task compared with today? | Calculated |
| 16 | Money | What happens to the cost if volume triples? | Modelled, with caps |
| 17 | Risk | What happens when it is wrong, and how would we know? | Visible, reversible, monitored |
| 18 | Risk | Does it decide anything about a person? | No, or with human decision and explainability |
| 19 | Measure | What number shows it worked, and what is it today? | Defined and measured now |
| 20 | Measure | When do we review, and what would make us stop? | A date and a stopping condition |
Using the checklist
- Get the four people together: decision-maker, prospective owner, someone who does the work, data protection adviser.
- Answer questions one to six before discussing any tool.
- Check the data questions against reality, by looking at the actual data rather than describing it.
- Name the owner and the exception handler before going further.
- Do the money arithmetic, including running costs and a tripled-volume scenario.
- Start the baseline measurement the same week, because it takes two weeks.
- Write the review date and the stopping condition into the decision.
- Fund the build only once the answers exist.
What honest answers usually change
Businesses that answer these questions properly tend to build something smaller and duller than the idea they started with, sooner, with a clearer measure. They also frequently discover a different first project: the hours are somewhere else, or the data for the exciting idea is not ready while the data for a neighbouring process is. That rerouting is the checklist’s main value, and it happens in an afternoon rather than after six months and an invoice.
What this means for you
Before spending on AI, answer twenty questions about the problem, the data, the people, the money, the risk and the measurement, with the decision-maker, the owner, someone who does the work and your data protection adviser. Where answers are missing, do that work first. Fund only what can name an owner, a baseline, a measure, a review date and a stopping condition.
Frequently asked questions
Can we answer these before knowing what we want to build?
That is the best time. The first six questions are about the problem and the hours, not the solution, and their answers usually point at a different and better first project than the one that prompted the conversation. Businesses that answer these before choosing a tool build something smaller, sooner and more useful.
What if we cannot answer several?
That is a result, not a failure. Unanswerable questions name the work to do first: measure the baseline, name an owner, sort out the data terms, define the exception path. Those take days, not months, and they are what makes the project succeed. Funding a build while several answers are missing is how the money gets wasted.
Who should answer them?
The decision-maker, the person who would own the result, someone who does the work today, and whoever can speak to data protection obligations. Answers from one person describe an intention; answers from those four describe a project.