Every organisation now has an AI list. Somebody keeps it in a document: quoting could be automated, support could be triaged, there is a demo somebody saw at a conference that looked like magic. The list grows faster than anything on it gets done, and that is normal. The hard question is not whether you can use AI. It is which item deserves to go first, and which items you should refuse to do at all.
This is a decision problem, not a technology problem. Treat it like one and most of the difficulty disappears.
The costly mistake is confidence, not caution
The expensive AI failures we see rarely come from moving too slowly. They come from moving confidently on the wrong project: six months on a flagship chatbot nobody asked for, while a quoting process that eats twelve hours a week sits untouched because it sounded boring.
Boring is a compliment in automation. The best first AI projects are almost always the dull ones: high volume, clear rules, an obvious definition of success, and a human already doing the work who can tell you exactly what good looks like.
Score returns, not excitement
Put every candidate through the same four questions before anyone opens a laptop:
- What is the return? Hours saved multiplied by how often the task happens, plus the cost of the errors it would prevent. If you cannot estimate this within an order of magnitude, the project is not ready to be ranked.
- Is the process stable? Automating a process that changes every quarter means rebuilding every quarter. Automate what has settled.
- What happens when it is wrong? Every AI system is sometimes wrong. If a mistake is cheap and a human can catch it, proceed. If a mistake is expensive, silent, or public, the project needs a review step designed in from the start, and that cost belongs in the score.
- What does it touch? A tool that reads one inbox is a week of work. A tool that must talk to your CRM, your accounts package and a twenty-year-old database is a programme. Integration effort sinks more AI projects than model quality ever does.
Score each answer roughly. Precision is not the point; honesty is. The ranking that falls out of this exercise is usually surprising, and it is usually right.
Build, buy, or wait
Every scored opportunity lands in one of three columns.
Buy when the problem is generic. Transcription, document extraction, meeting notes, customer-service deflection: vendors have spent years and millions on these. Buying is not a defeat. It is what winning looks like for problems that are not yours alone.
Build when the process is genuinely yours. If the way you quote, schedule, or assess risk is part of why customers choose you, an off-the-shelf tool will flatten exactly the thing worth keeping. Build small, on top of proven models, and own the result.
Wait when the return does not clear the effort. In code, the rule fits on one line: if roi < effort: wait. Waiting is a verdict, not a shrug. It comes with a condition attached: revisit when the volume doubles, when the process stabilises, or when the tooling matures. A refusal with a review date is worth more than a pilot with no purpose.
A worked example
A services firm brought us a list of four ideas. Ranked on returns, it looked like this:
| Opportunity | Return | Verdict |
|---|---|---|
| Quote automation | 3.2× | Build |
| Document triage | 2.4× | Buy |
| Enquiry routing | 1.3× | Buy, later |
| Computer-vision demo | below 1× | Wait |
The vision demo was the project everyone was excited about. It was also the only one with no measurable return. Quote automation, the least glamorous item on the list, paid for the entire exercise inside a quarter. That inversion, hype at the bottom and returns at the top, is the usual shape of an honest ranking.
What this means for you
You do not need an AI strategy document. You need your list, the four questions, and the discipline to act on the answer even when it points at the boring project.
That ranking exercise is exactly what our AI consultancy engagement produces: an evidence-based verdict on every candidate, including the ones we tell you not to do. If you want a lighter starting point, the Clarity Report will assess your website in about a minute, free.