An AI feature feasibility checklist tests, before anything is built, whether the idea holds up in eight areas: a genuine need for AI, data and right of use, quality evaluation, error tolerance, human review, cost and latency, privacy and monitoring. Each answer scores from 0 to 3, and blocking rules stop a high score from hiding a serious risk.
How to use
- Describe in one sentence what the AI will do and for whom.
- Answer the eight questions by choosing the option closest to today’s situation, not the planned one.
- Read the verdict, the readiness score and the reasons: blockers come first.
- Follow the next steps, which start with the weakest areas.
- Export as PNG, PDF or DOCX and run the checklist again once the gaps are dealt with.
What each field asks for
| Field | What to enter |
|---|---|
| AI feature | What the AI will do and for whom, in one sentence. |
| Does the task really need AI? | Compare it with a solution based on rules, search or a form. |
| Data and right of use | Real examples of the use case, to build and to evaluate. |
| How will quality be evaluated? | What counts as a good answer and how to measure it before launch. |
| Error tolerance | What happens when the AI gets it wrong, and whether anyone notices. |
| Human in the loop | Who reviews, approves or undoes what the AI produces. |
| Cost and latency | Cost per use and response time against margin and experience. |
| Privacy and compliance | Personal data, legal basis, provider contract and sector rules. |
| Monitoring after launch | How you will know that quality has dropped in production. |
Example
A fictional support team wants an automatic summary of conversations for the agent. Rules have already failed (3 points), there are real examples with right of use (2), the agent approves the summary before using it (3), errors are annoying but easy to fix (2), and cost and privacy are mapped (2 each). Missing are quality evaluation, done by looking at a few cases (1), and monitoring beyond usage and technical errors (1). Result: 16 of 24 points, 67% readiness and a "prototype first" verdict, with the evaluation set as the first step.
When to use
- When someone proposes "adding AI" to a feature and the team must decide whether it is worth investing.
- Before asking for budget or forming the team, to know what a prototype has to answer.
- In the pre-launch review, to confirm that evaluation, human review and monitoring are in place.
- To compare two AI ideas on the same scale.
Frequently asked questions
How is the verdict calculated?
Each area scores from 0 to 3, for a total of 24. "Go" requires every area at 2 or more and at least 18 points. "Not yet" appears below 10 points or with a blocker. Everything else is "prototype first".
What is a blocker?
It is an answer that stops progress even with a high score: rules or search already solve the problem, there is no data from the real case, personal data is involved without a legal basis, or an error can cause serious harm with no human review before it takes effect.
Why ask whether the task really needs AI?
Because rules, search or a good form are usually cheaper, more predictable and easier to explain. AI is justified when the task handles varied content, such as free text, images or voice, and rules have already failed.
What is an evaluation set?
It is a collection of real cases with the expected answer and pass criteria. It measures quality before launch and with every change of model or instructions. Without it, quality becomes a matter of opinion.
When is human review mandatory?
When an error can cause serious harm that is hard to notice, such as with money, health or rights. In those cases the person must approve or edit before the effect. Undoing afterwards is not enough.
Does the checklist replace legal review?
No. It shows whether privacy and compliance have been addressed, but the legal basis, the provider contract and sector rules need review by whoever is accountable for them in the organisation.