Most AI projects I hear about start with a sentence like this: “We need an AI agent.” Sometimes that's right. But it's a conclusion, and it arrives before anyone has done the work that would justify it.
The existence of a powerful technology doesn't make it the right starting point for every problem. So I start somewhere else.
Start with what you're trying to cause
Before I ask which model to use, I ask what we're actually trying to cause. More qualified conversations? Fewer leads that sit untouched? Faster answers that are still correct? Each of those is an outcome. A model is one possible way to produce it.
Then I ask three more questions, in this order:
- What mechanism should cause that outcome?
- What evidence would tell us the mechanism is working?
- Only then: what technology should implement it?
That order matters. If the technology comes first, the questions after it get bent to fit it. If it comes last, it has to earn its place.
Understand the system it lives in
An outcome is produced by a system: people, process, technology, data, information, timing and ownership, all depending on each other. Before I change any part of it, I want to know how the outcome happens today. Where does information move? Where does ownership change hands? Where does time matter? Where does it fail, and how would anyone find out?
Most of the problems worth solving live in those handoffs, not inside any one tool.
Find the constraint
Once the system is visible, the next question is what's actually limiting the outcome. That's rarely the step that looks most automatable.
Here's the kind of case I mean. A dealership wants an AI assistant to answer internet leads faster. But suppose the leads that go unanswered are assigned to a salesperson who's off that day, or who no longer works there. The constraint isn't response speed. It's ownership. A faster assistant would answer the leads that were already being answered, and the orphaned ones would still sit.
An impressive AI system solving a problem that isn't the constraint is still a poor investment.
Then choose the intervention
Sometimes the right intervention is an AI agent. Often it's something plainer: a database query, a deterministic rule, an integration, a notification at the right moment, a change to the process, or removing a step nobody needs. Those are frequently cheaper, safer and more reliable, and they're easier to verify.
When AI is the right answer, it's usually because the step involves language, ambiguity, interpretation or synthesis that rules can't capture well. Even then, I want to know which parts around it should stay deterministic and where a person should still decide.
AI is a tool available to the system. It isn't the starting assumption.
Good implementation isn't measured by how much AI gets used. It's measured by whether the system works.