AI Consulting Starts With Process Design

The fastest way to spot an AI project that will fail is to listen to the first question. If it's "which model should we use?", the project is already off course. The model is one of the last decisions, not the first. The first decision is process design: who is this for, what do they actually decide, what is that decision grounded in, and how does the whole thing get better over time?

I've started a lot of AI engagements, and the ones that work all begin the same way — not with technology, but by mapping a loop. AI is a component inside a business loop. It is never the loop itself.

Map the loop before the model

Every useful AI system sits inside a cycle with four parts. Get the cycle right and the technology choice becomes almost boring — which is exactly what you want. Get it wrong and the smartest model in the world has nothing to attach to.

The four parts, in order

The user — who acts on the output

Name the person and the exact moment they use the result. If you can't, you're building a toy. "The model summarises support tickets" is not a use case; "the duty manager reads a three-line summary before each escalation call so they walk in informed" is. The user defines what good output even looks like.

The data — what the answer is grounded in

Where does truth live, how fresh is it, and who owns it? An AI system is only ever as good as what it stands on. Most projects I'm asked to rescue don't have a model problem; they have a data problem wearing a model costume — the inputs are scattered, stale, or contradictory, and no amount of model cleverness fixes that.

The decision — what actually changes

The output has to change an action. If nothing downstream changes because of it, you've built a dashboard nobody opens. Trace the line from the output to a different behaviour: an email sent, a case routed, a price adjusted, a call prioritised. No changed action, no value.

The feedback — how it learns it's working

How will you know the system is right, catch it when it's wrong, and improve it? This is the part teams skip and the part that decides whether the thing survives contact with reality. Without a feedback path you're not running an AI system — you're flying blind and hoping.

Wrong order vs. right order

Almost every troubled project I see has the sequence backwards. The fix is rarely a better model; it's a better starting point.

Sometimes the honest answer is "not AI"

Mapping the loop occasionally reveals that the real problem is a missing form, an un-joined dataset, or a single automation rule — no model required. Saying so is part of the job, and it's why my advice is worth something: I'm an economist and a builder, not a vendor with one product to sell. The goal is the business outcome, not the AI.

AI is a component inside a business loop, not the loop itself. Design the loop first, and the model choice becomes almost boring — which is exactly what you want.