You don't have an AI problem. You have a process problem.

The request always arrives in the same shape. We want to add AI to our support inbox. We want an AI agent for sales. Can AI handle our onboarding?

So I ask the boring question: what happens right now, today, when a message comes in?

And there's a pause. Then three people give three different answers, none of which match the document someone wrote in 2023.

AI is an amplifier, not a mop

Nobody wants to hear this, but a model will not clean up after you. Point one at a process with unclear ownership, inconsistent inputs and no definition of "done", and you get the same mess — faster, in higher volume, with more confidence in its tone.

Automating an unclear process doesn't make it clearer. It makes it unclear at scale, in writing, to your customers.

The tell

You can spot a process problem in about ten minutes. Ask three people who do the same job to describe it independently. If the three descriptions differ on anything that matters — which cases get escalated, what counts as urgent, who signs off — you've found the actual project.

It's not a glamorous finding. There's no demo. But it's the reason the last two automation attempts quietly stopped being used.

What good looks like

A process worth automating can be stated as a handful of unambiguous sentences: this event starts it, these are the branches, this is the human check, this is how you know it worked, this is who gets told when it doesn't.

Get there and the AI part becomes almost anticlimactic. You're not asking a model to exercise judgement on your behalf — you're asking it to do a specific, checkable job inside a process that already knows what it wants.

That's when automation gets genuinely good: classification with a defined set of labels, drafting inside a format you've agreed on, extraction against a schema you actually use. Narrow, verifiable, boring. Boring is the goal.

Start smaller than feels worth it

Don't automate the inbox. Automate one message type in the inbox — the one that arrives most often and has the least judgement in it. Run it alongside the humans for two weeks and compare.

You'll learn more from that one narrow workflow than from six months of planning the agent that was going to do everything. And if it works, you've got a template, a credibility win, and a much better argument for the next one.

AI is not the hard part any more. Deciding exactly what you want to happen still is.