When we run AI discovery inside a business, the hard part isn't finding ideas. Every team has a list. The hard part is killing the wrong ones early, before they eat a quarter's budget. These are the five questions we put to every candidate project.
1. What does this cost you today?
If nobody can put a number on the current pain — hours, rands, missed deals — the project has no baseline, and "success" will be a matter of opinion. We help teams estimate it roughly rather than skip it: even "two people, one day a week" is enough to rank against other options.
2. What happens when the AI is wrong?
Every model is wrong some of the time. The question is whether a mistake costs you an awkward edit or a client. Draft-for-review workflows tolerate errors beautifully. Fully automated customer-facing decisions don't. This one question usually reshapes the project more than any other.
3. Does the data actually exist?
Half the AI projects we've been asked to rescue failed here. The process everyone wants to automate turns out to live in email threads, WhatsApp messages, and one veteran employee's head. That's not a reason to stop — it's a reason to sequence: capture the knowledge first, automate second.
4. Who owns it after go-live?
AI systems drift. Prompts need tuning, edge cases accumulate, models get upgraded. If there's no named person whose job includes caring about this in month six, the system will quietly rot. Ownership is a budget line, not a hope.
5. Would a simpler tool do?
Sometimes the honest answer to an "AI project" is a spreadsheet formula, a form, or a well-configured workflow tool. We'd rather tell you that in week one than bill you for discovering it in month four. The businesses that trust us with the big projects are the ones we've talked out of small ones.
The scorecard
A project that clears all five questions is rare — and worth moving on quickly. A project that fails one isn't dead; it has a known weakness to design around. A project that fails three is a demo waiting to happen. Spend the money elsewhere.