Every AI idea looks good in a demo. The question that matters is whether it will still be saving time, money or mistakes a year from now, on your data, inside your systems, with your team relying on it. These seven questions are the ones we work through in discovery. You can answer most of them yourself in an afternoon.
1. Is there enough volume?
A task done twice a month rarely justifies a system, however tedious it is. A task done 50 times a day almost always deserves a look. Count how often the work happens, and how long each one takes, including the waiting and the chasing. Volume times time per item is the ceiling on what the project can ever save.
2. What does a mistake cost?
This cuts both ways. If errors are expensive, such as a missed coverage shortfall, a wrong invoice paid or a contract clause overlooked, then catching them is valuable, and AI that flags problems for a person can pay for itself on the mistakes alone. But expensive errors also mean the system needs careful design: every value traced to its source and a person approving anything that matters. If a mistake is cheap and easy to undo, you can let more run on its own, sooner.
3. Is the information there, and can it be reached?
AI works from information. Ask where the documents, emails and records the task depends on actually live, and whether they can be reached. Messy is fine: scanned PDFs, inconsistent spreadsheets and long email threads are normal, and handling them is most of the job. Missing is a different problem. If the facts a person uses to make the call are only in someone's head, the first project is writing them down.
4. Can the rules be written down, and who approves?
If two experienced people would make the same call on the same item, there are rules, even if nobody has written them down yet. Those rules are what the system gets checked against. Then decide who approves what. Some steps can run on their own once they've proven reliable; anything that spends money or is out of the ordinary should wait for a named person. If nobody can say who owns the decision today, the project will stall on that question later.
5. How will you know it works?
This is the question most projects skip, and the one that matters most. Before anything is built, collect a set of real examples with known right answers: 50 to 200 past items is usually enough, including the awkward ones. That set becomes the test. Every version of the system is scored against it before anyone relies on it, and again after every change.
6. Does it fit the systems you already run?
A tool that saves ten minutes but adds a login, a copy-and-paste step and a new place to check won't be used for long. Ask where the result needs to end up: your accounting, CRM, loan system or shared drive. Most modern systems offer a way in, such as an API or a scheduled export, and some older ones need more work. Knowing which before you start is the difference between a tool that fits into the day and one that sits beside it. Our integrations work exists for exactly this.
7. What will it cost to run, not just to build?
The build is only the start. A fair estimate also includes AI service usage, which grows with volume; hosting, storage and monitoring; the time your team spends reviewing what the system flags; and keeping it current as your rules, documents and systems change. Put those next to what the work costs today, from questions 1 and 2. If running it costs more than the work it replaces, it isn't worth building yet, and that's a useful answer too.
Score it
| Question | 0 | 1 | 2 |
|---|---|---|---|
| Volume | A few times a month | Weekly | Daily, or many times a day |
| Cost of a mistake | Unknown | Known, but rare | Known, frequent and measurable |
| Information | Mostly in people's heads | Exists, hard to reach | Exists and can be reached |
| Rules and approvals | Every case is a judgment | Rules exist, owner unclear | Rules can be written, owner named |
| Testing | No past examples | Some examples, answers unclear | Real examples with known answers |
| Systems | No way in | Possible with work | API or export available |
| Running cost | Not estimated | Rough idea | Estimated against today's cost |
- 11 to 14: a strong candidate. Take it to discovery and plan the first release.
- 7 to 10: promising, with gaps. Discovery should close them before anything is built.
- 0 to 6: not yet. Start with the gaps: write down the rules, gather examples, or pick a busier workflow.
A low score isn't a verdict on AI for your business. It usually means this particular workflow isn't the right first one. Our guide to where to start with AI covers how to pick one that is.
Where discovery comes in
Scoring on paper gets you most of the way. Discovery finishes the job on your own files: we watch the work, test what's possible on real documents, confirm how your systems can be reached and estimate what it would cost to build and to run. Our discovery is free: a 40-minute call, then one to three weeks if there's a fit, ending in a written recommendation of what we would build, why and what it costs. Book a free discovery call (opens in a new tab) and bring your scored checklist.