Most of the agent projects we're asked to quote shouldn't be agent projects. Not because the technology can't do it — often it can — but because the process underneath the agent is the actual problem, and an agent built on a broken process just produces wrong answers faster.
Start with the workflow
Before we talk about models, we map the workflow. Where does the work come from, what has to happen to it, and where does a human genuinely exercise judgement? Most workflows have a surprising amount of mechanical judgement — decisions that feel like thinking but are really lookups. That is where an agent earns its place.
The rest — the parts that need accountability, context, or a relationship — should stay with a person. An agent that quietly makes those calls is a liability dressed as a productivity gain.
Four conditions
We'll build an agent when four things are true. The workflow is understood and repeatable. The judgement we're automating is mechanical rather than political. There is real data to test against. And a human can still sign off where the stakes require it.
When one of those is missing, the honest answer is to fix that first. We say so in the discovery sprint, in writing, even when the answer is not yet.
An agent built on a broken process just produces wrong answers faster.
Measure, or don't ship
The part that gets skipped is evaluation. We build the evaluation harness before the agent — a labelled set of real cases the agent is checked against on every change. If we can't measure whether the agent is right, we don't ship it. A wrong answer delivered confidently is worse than no answer at all.
None of this is exotic. It's the same discipline you'd apply to any system that makes decisions. The model is new; the engineering isn't.