Watch any AI demo day and you will see the same highlight reel: agents booking calendars, ordering groceries, planning trips. Fun to watch. Almost nobody signs a budget for it.
The software that gets bought looks boring by comparison. It reads PDFs. It sorts spreadsheets. It chases approvals. It is document work, and it is where the hours go.
I keep meeting teams who run on documents and do not think of themselves as having an AI use case. A consultancy where every engagement starts with forty CVs in a shared drive. A training provider where assessors retype the same observations into three systems. A small lender where underwriting means reading bank statements line by line. None of this looks like a keynote. All of it costs real money every week, and you can count the hours.
That countability is the whole point. When someone asks me where to start with AI, I tell them to start where the hours are. Not where the hype is.
Good looks like this, and none of it is exotic. Pull the facts out of the documents instead of asking people to retype them. Sort what you pulled: this invoice matches, that one does not, this application is complete, that one is missing two pages. Let people search across everything instead of asking who has the latest version. Show it on a dashboard. And keep a human on the exceptions, because the exceptions are where judgment lives and always will.
I have deliberately described commodity parts. Extraction, classification, search, a dashboard, an approval queue. You have seen each of these before. The value is not in any single piece. It is in the assembly: the boring version that works every morning beats the brilliant version that works in a demo.
There is a second effect that only shows up after you have done this a few times. The second client in the same niche goes faster than the first, because you kept what you learned. The document types repeat. The exceptions repeat. The dashboard is eighty percent the same. We treat that residue as an asset: every engagement grows a small library of reusable pieces, and owned products grow out of the patterns that keep recurring. Client work pays for the learning; the learning compounds.
None of this requires believing Standard Claims about AI. It does not need a model breakthrough or a bigger GPU. It needs someone to look at where the hours go and start there.
So here is my suggestion, and it is the same one I give in sales calls. If your team runs on documents and approvals, that pile is your first AI project. Not the calendar agent. The pile.
We run a short Discovery Sprint for exactly this: we map the workflow, prototype the boring version, and give you scope, architecture, and an estimate. Then you decide with evidence instead of slides.