AI Opportunity Sprint
A fixed-price assessment that tells mid-cap boards which AI investments are real — before the seven-figure commitment. Two weeks to an evidenced decision map, then a working prototype on the one or two processes worth proving.
The call is free, it isn't a pitch, and if none of your candidates are real I'll say so.
The nine that didn't are the return. Each one carries a written reason, which is what makes the map defensible in a board paper.
The problem
Usually the model was fine. In the cases I've seen, one of three things happened — and all three are findable in a fortnight, before the money goes out.
Good output, and no downstream system consumed it. Someone retyped the result into the system of record. Net saving: zero. That's an integration problem discovered in month seven, at implementation prices.
It existed, but not in a form anyone could get to inside the pilot window. The team hand-assembled a sample, proved something on it, and couldn't repeat it at volume.
It worked, and the people doing the job didn't trust it, because no one had asked them what trustworthy would look like. An accurate system at 12% utilisation.
Most stalled pilots were never intelligence problems at all. They were data, integration or process problems with an AI budget attached.
The method
Every candidate process goes through the same lens, and the reasoning is recorded either way. Most candidates fail question one. That is the point of question one.
The lens runs across the operating model, not just the AI list — because the work that qualifies is usually the work a capable person does in two seconds, several hundred times a day, and can't explain.
Or is it data, integration or process wearing an AI costume? Judgement applied repeatedly to unstructured input is the shape that qualifies.
Fails when: the SOP covers everything and the exception rate is near zero.Not "does it exist" — can it be reached, in a usable form, inside the window, with an owner who can say yes.
Fails when: the only route to it is a six-week access negotiation nobody has started.Rules engines scale fine and cost less. If the task is deterministic, a model is an expensive way to be less predictable.
Fails when: the ERP already does this and nobody switched it on.Where does the output land, who owns that system, and what does the handover cost. This is the question that turns a demo into a programme.
Fails when: integration cost exceeds the saving and no one has added it up.What does that person lose if the output is wrong, and who carries it? Can they see why the system reached its answer, in language they already use? Can they override it without raising a ticket? Adoption isn't a phase to be handled after build. It's a design constraint, and it should be answered before you spend anything.
A note on evidence
ERP implementations encode an idealised process. Reality arrives, people build workarounds, and the workarounds become the actual job. Nobody updates the SOP, because the SOP was never how the work got done.
So the process gets reconstructed from the event trail: what arrived, what got touched, in what order, how often the sequence broke, and what people did when it did. The gap between the two documents is where the exceptions live — and the exceptions are the part worth putting a model on rather than a rule.
The engagement
Capability by capability across the operating model. What each process actually does, where the volume and the exceptions sit, who carries the judgement, what systems touch it, where the data lives and who owns it.
Every candidate through the five questions. Plotted on value against feasibility, with a written reason recorded for each process that didn't make the cut.
Stopping at this point is a legitimate outcome and a cheap one. Nothing further is committed until you choose.
A working prototype on synthetic or sanitised data, embedded with your team, with the integration and governance requirements written down. The number that decides it isn't raw accuracy — it's what proportion the system can handle on its own, and whether it correctly refuses the rest. If the checker still has to check everything, you've added a step and removed nothing.
Numbers derived from a thing that ran, on data that looks like yours, with the exception profile and the integration cost written down rather than assumed.
Nine processes ruled out, each with a defensible reason. Some of those were going to get funded and stall in month eight. I've seen this outcome save more money than the first one.
The candidates that qualify are often not the ones on the list. They're the queues that were too small and too boring to justify a slide — which is also why nobody has ever costed them.
Proof
The process on paper: submissions prepared from ERP extracts on a monthly cycle. The process in reality: inbound senders in the dozens, data in the email body rather than attached, formats differing by jurisdiction.
Until this is real, the honest line on the page is that the method comes from 25 years of doing the underlying work — not from a portfolio of AI assessments.
Who runs it
Twenty-five years designing the systems this assessment interrogates: regulatory reporting, master data, and group integration across multi-country operations. Currently also building a voice-first AI product, which is where the practical view of what models do and don't do reliably comes from.
You get the person who did the work on the call, in the workshops, and on the page. Delivery capacity is deliberately limited to two concurrent engagements.
Objections
Then you have a documented decision not to spend seven figures, with a written reason for every candidate. That's a defensible board paper and it costs a fraction of finding out the same thing in month eight of a programme.
Sensible. The difference is incentive: there's no implementation practice behind this, so I'm free to tell you which candidates aren't worth doing. If that's useful as a second opinion on a shortlist, it's a cheaper conversation than the one you're already having.
Correct, and you shouldn't. Synthetic or sanitised samples will tell you whether the process is tractable, where the model breaks, what the exception profile looks like and what integration you'd need. What they won't give you is a production accuracy figure, and I wouldn't claim one at this stage.
The two-week assessment and the prototype stage are quoted separately and priced by whether it's one process or two. Total is typically around £30,000, fixed and dated up front. The breakdown goes out before you commit to anything.
Next step
Bring three or four of those, plus whatever is already on the AI list. Thirty minutes. I'll tell you honestly which look real and which look like plumbing. If two of them are real we can talk about proving it. If none are, that's a useful thing to know for free.
Take the five questions and run them across your own list first.
Get the AI Readiness Lens