AI automation is the phrase clients use when they want something taken off their team's plate. Most of the time the thing is paperwork. Sometimes it's coordination. It's rarely judgement, even though judgement is what everyone wants automated first.
The way I've come to think about it is that AI doesn't take over a process. It changes the shape of one or two specific moments inside one. The moment a quote gets drafted. The moment an invoice gets categorised. The moment an email gets routed to the right person. Everything around those moments, the approvals and the exceptions and the 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 message that says can someone look at this is a moment that could have been sorted automatically. Every long back-and-forth thread is coordination cost adding up. A lot of the AI features I've shipped live in or near the inbox.
The second place is the document pile. Invoices, contracts, statements, forms, transcripts, claims. Nearly every business has a corner where someone is typing information out of one document and into another system. AI is genuinely good at this now and the build is usually short, around four weeks for the first workflow and less for the ones after it.
The third is reporting. Most weekly reports are someone spending Friday afternoon assembling charts nobody reads carefully. What usually helps isn't another dashboard, it's a short paragraph at the top explaining what changed and what to be careful about. Three sentences instead of nine charts.
Where I've watched this go wrong is trying to automate decisions that aren't routine. My rough rule is that if a person would think for more than a minute before saying yes or no, the model probably shouldn't be saying it either. Use it to draft and summarise and classify and surface, then let people decide.
The thing I'd flag for any operations team is AI agents that take several steps without anyone checking. They're tempting because they sound impressive. In practice each error compounds into the next step and working out what went wrong afterwards is exhausting. Single-step help that someone reviews is far more reliable than multi-step automation that nobody does.
The way I'd start is to pick one workflow, time how long it takes today, build something in about four weeks that addresses the single most expensive moment in it, then measure it again. If the time saved is real, do another one. If it isn't, you've learned something for not much money. Most of my clients are five or six workflows in before they think about anything bigger.