Start with the inbox

Asked what their first AI project should be, most businesses reach for something impressive. The right answer is usually sitting in the shared inbox, arriving as you read this.

"What should our first AI project be?" produces a lot of whiteboard sessions it doesn't need. Walk into almost any office — an insurance agency, a conveyancing firm, a wholesaler, a property manager — and the answer is already there, filling up by itself: the shared inbox. Documents arrive all day. Someone reads each one, works out what it is and who it belongs to, files it, updates a record, sends a reply that says little more than "received, thanks" — and repeats that until five o'clock. It is the least glamorous workflow in the building, which is precisely the recommendation.

Why the inbox keeps winning

In our guide to choosing what to automate first, we score candidate tasks on three questions: how often it happens, how rule-shaped it is, and what a mistake costs. The inbox is the rare job that scores perfectly on all three. Volume: it never stops arriving — weekends included — and volume is what pays for automation. Rules: "read it, work out which client it concerns, file the attachment, update the record, acknowledge it" is an instruction a capable temp could follow on their first morning, which is the test for whether an agent can. Cheap mistakes: a misfiled email isn't a disaster, it's a re-file — especially when every action is logged and reversible. Most businesses also discover, when they finally count, that the inbox eats more hours than anything else on the candidate list. It just doesn't look like it, because the time goes in ninety-second sips all day rather than one visible lump.

What an inbox agent actually does

Take the insurance agency, since it's the example we know best. Post arrives electronically all day: policy documents, claims correspondence, statements, queries. The agent reads each email and its attachments as they arrive, works out which client and which matter they belong to, files the documents where the firm's own conventions say they go, updates the record, and sends the acknowledgement — at two in the afternoon or two in the morning. Anything it can't confidently place — the ambiguous, the unusual, the possibly-urgent — it doesn't guess at. It flags the item to a person, with its best reading attached, and moves on. All of it happens under the discipline we bang on about everywhere on this site: the agent has its own login with minimum access, every action is recorded, and anything that matters more than filing waits for a human.

The week-one feeling

This project has a particular quality the others don't: the result walks in the door on day one. The morning backlog — the forty unopened items a person used to clear before they could start their actual job — isn't there. Clients who emailed at 9pm have an acknowledgement timestamped 9:01. The customer-visible number, the time from a thing arriving to a thing handled, collapses from "sometime Thursday" to minutes. The team's first reaction is usually suspicion — and the second, about a fortnight later, is that you'd have a mutiny if you switched it off. For a first project, that early, visible, felt result matters almost as much as the hours: it's what buys the goodwill for project number two.

The honest trouble spots

Three, in practice. The attachments are not all tidy PDFs. Photographs of paper taken at an angle, scans of scans, spreadsheets saved as images — modern systems read these far better than most people expect, but "far better" is not "always", so the design assumption is that some percentage gets flagged to a person rather than filed with false confidence. Senders bundle. One email, three clients, five attachments and a question — the agent needs to be allowed to split its handling, and the unusual compound cases go to the flag pile. And your exception list is longer than you think — every business believes its post is 95% routine and discovers it's more like 80%, so expect the human-flagged share to start noticeably higher than anyone estimated and to fall as the rules absorb what the first month teaches. None of these are reasons to hesitate. They're the reasons the flag-to-a-person path is designed in from the start rather than bolted on after the first embarrassment.

How to start

The same way as every project on this site: count first. A fortnight of honest numbers — items per day, minutes per item, how often something is misfiled or missed and what that cost — which is exactly what our free workflow review produces, and which you can equally run yourself with a tally sheet and discipline. Then start narrower than feels necessary: filing and record-updating live, outgoing acknowledgements drafted but held for a person's approval until the wording has earned its freedom. Loosen the reins the way our ninety-days guide describes — routine cases to spot-checks, exceptions to people, permanently. Six months in, you'll have the before-and-after that makes the next decision easy — and a team that has quietly forgotten the inbox was ever anyone's morning job.