Opinion
Six situations where a scheduled job, a lookup table or a two-hour conversation beats anything we could build.
We turn down roughly a third of what we are asked to build. Not out of principle — we would like the work. It is that in these six situations the honest answer is cheaper, faster and more reliable than anything with a model in it, and pretending otherwise ends with a client who does not call back.
"Flag orders over RM 5,000 from new accounts shipping to a different state." That is an if statement. It costs an afternoon, runs in a millisecond, never drifts, needs no evaluation set, and can be read by anyone in the company.
The tell is whether you can write the rule down in a sentence without using the word "usually". If you can, write the rule. A model asked to apply a rule it could have been given is strictly worse: slower, more expensive, and occasionally creative in ways nobody wanted.
A task that takes twenty minutes and happens four times a month costs about sixteen hours a year. Any real deployment — build, evaluation, integration, monitoring — costs more than that to maintain, before it has done anything.
Our rough threshold is twenty hours of human work per week on one repeatable task. Below it we say so on the first call. This is the single most common reason we decline, and it is usually the first time anyone has done the arithmetic out loud.
If the company is mid-ERP-migration, or the ops director started six weeks ago with a mandate to redesign the flow, automating today's process locks in something everyone has already decided to abandon. Worse, the automation becomes an argument for keeping a bad process because it is now expensive to move.
Wait for the new process. Then automate it — with the enormous advantage that we can design the flow knowing part of it will be machine-handled, which usually produces a simpler flow than either party would have drawn alone.
Some tasks are genuinely judgement calls where two experienced people disagree 30% of the time. You can automate those, but you cannot evaluate them, and a system you cannot evaluate is a system you cannot safely trust or improve.
The diagnostic is cheap: take fifty real cases, have two experienced staff decide independently, and compare. If they agree less than about 85% of the time, the problem is not automation-shaped yet. What that organisation needs is a decision framework — a two-hour conversation and a one-page policy. Sometimes we run that conversation and bill nothing, because the resulting policy is what makes the later project possible.
Not "the data is messy" — messy is normal and we work with it. This is the case where the information genuinely was never recorded. If the reason an order was cancelled lives only in a phone call, no amount of retrieval will find it.
The fix is upstream: capture the field. That is a three-week change to a form and a habit, not an AI project. Come back in a quarter with six months of the field populated and the same project becomes straightforward.
A surprising number of automation requests are organisational problems wearing a technical costume. Sales promises dates that operations cannot meet, so someone wants an agent to reconcile the gap. Purchasing does not tell the warehouse about the promotion, so someone wants a system that predicts the surge.
You can build both. They will work, in the sense that they will paper over the gap and make it permanent. The cheaper intervention is a weekly fifteen-minute meeting, and we have recommended it more than once — including to a client who had budgeted six figures.
This is not an argument that AI is overhyped. The deployments we do run absorb six-figure task volumes and hold up. It is an argument that the diagnosis has to come before the prescription, and that a vendor whose only tool is a model will find that every problem is model-shaped.
The scoping week exists mostly to answer this question honestly. It is fixed-fee, it is one week, and the output is yours whether or not the answer is yes. A written "here are the three things not worth automating and why" has saved clients more money than several of our builds.
Recognise the process?
Bring one process that is eating your team's week. We will tell you on the call whether it is a good candidate — and we say no to roughly a third of what we are asked about.
