What AI will not fix in your business

Consultancies are not usually incentivised to publish this list. But the fastest way to lose a client's trust is to build something well that should not have been built at all, and the second fastest is to be visibly surprised when it does not help.

A process nobody can explain

If three people describe the same process three different ways, the problem is not that the process is slow. It is that there is no process — there are three, running in parallel, held together by informal correction.

Automating one of the three versions does not resolve this. It makes one version faster and louder, and pushes the disagreement downstream. The work here is to agree what the process is. That is an operational conversation, not a technical one, and it usually takes a fortnight and no software.

Data that does not exist yet

A surprising number of AI proposals depend on information the business does not currently capture. Not captured badly — not captured. Predicting which clients are likely to churn requires a history of clients who churned and why, recorded consistently. If that history lives in the sales director's recollection, no model will recover it.

This is not a reason to abandon the objective. It is a reason to sequence honestly: start capturing the data now, and revisit the model when there is something to build on. What is not acceptable is a project that quietly assumes the data problem away and discovers it in month four.

A capacity problem that is really a demand problem

Teams frequently ask for automation because they are overwhelmed. Sometimes the honest diagnosis is that the business has taken on more work than it is structured to deliver, or has never declined anything.

Releasing capacity into that situation does not produce relief. It produces slightly more throughput and the same overwhelm, because the constraint was never the speed of the work.

Automation makes a system faster. It does not make a badly shaped system the right shape.

A quality problem with no definition of quality

'Our reports are inconsistent' is a reasonable complaint. But if the business cannot articulate what a good report looks like — beyond recognising one when it appears — then there is nothing to build towards and nothing to test against.

The useful move is to have your best practitioner mark twenty examples and explain each mark. That exercise produces a standard. Sometimes the standard alone fixes the problem, at which point there is nothing left to automate.

Anything the business is not willing to change

Meaningful AI implementation changes how work is done. If the requirement is that nothing about the current process may change, what remains is a thin layer over an unchanged workflow. That can be worth doing, but it should be recognised as a small intervention with a small return, and priced and scoped accordingly.

Why this list matters commercially

Each of these can be dressed as an AI project, and each will produce a system that technically functions and commercially disappoints. The disappointment then hardens into a conviction that AI does not work here, which delays the projects that would have worked by a year or more.

The most valuable thing an adviser can say is often that the answer is not the thing they sell.