"AI automation" is the phrase clients use when they want me to remove a thing from their team's plate. Most of the time, the thing is paperwork. Sometimes it's coordination. Rarely is it judgement, even though that's the part everyone wants automated first.
The most useful frame I've found is this: AI doesn't replace a process. It changes the shape of one or two specific moments inside a process. The moment a quote gets drafted. The moment an invoice gets categorised. The moment an email gets routed. Everything around those moments — approvals, exceptions, audits — still belongs to people.
The first place I look in any business is the inbox. Not because email is interesting, but because email is where decisions hide. Every "can someone look at this" message is a moment that could be triaged automatically. Every threaded back-and-forth is a coordination cost that compounds. Most of the AI features I ship live in or near the inbox.
The second place is the document pile. Invoices, contracts, statements, forms, transcripts, claims. Every business has a corner where a person types data from one document into another system. AI is genuinely good at this now, and the engagement is short — usually a four-week build for the first workflow, less for the next.
The third is the report. Most weekly reports are people spending Friday afternoons assembling charts that nobody reads carefully. The right AI feature here isn't another dashboard; it's a paragraph at the top that says what changed, in plain English, with caveats. Three sentences instead of nine charts.
Where I've watched it go wrong: trying to automate decisions that aren't routine. If a human would think for more than a minute before saying yes or no, the model probably shouldn't say it either. Use the model to draft, summarise, classify, and surface — let people decide.
The trap I'd flag for any operations team: "AI agents" that act on multiple steps without supervision. They're tempting because they sound like magic. In practice, every error compounds, and the explanations afterward are exhausting. Single-step assistance that gets reviewed is dramatically more reliable than multi-step automation that doesn't.
The right way to start: pick one workflow, time how long it takes today, ship a four-week build that addresses the one moment with the most cost, then measure again. If the time saved is real, do another. If it's not, you've learned something cheap. Most of my clients are five or six workflows in before they consider a bigger investment.