Why "just add AI" is the wrong starting point for consultants

Most agent projects fail not because of the model, but because the workflow around it was never mapped out first. Teams get excited about what's possible, wire up a demo in an afternoon, and then discover three weeks later that nobody trusts the output because nobody agreed on what "done" looks like.

The three-question test

Before we write a single line of automation for a client, we make them answer three questions:

• Who currently does this task, and how do they know when it's done correctly? If there's no clear definition of correct, an agent can't be held to one either.
• What data does this task actually depend on? Not what data exists — what the person doing the task actually looks at.
• What's the cost of a wrong answer? A drafting agent that occasionally needs a edit is fine. A compliance agent that occasionally hallucinates a citation is not.

Answering these honestly usually cuts the scope of the first project in half — and that's a good thing. A narrow agent that reliably saves two hours a week beats a broad one nobody trusts enough to use unsupervised.

Start where the paper trail already exists

The workflows with the clearest paper trail — engagement letters, weekly status reports, intake forms — are almost always the best first candidates. They already have a definition of correct built in, because a human has been producing that exact output for years.