When a leadership team decides to automate, the first candidate is almost always the process that is most visible, most political or most intellectually interesting. It is rarely the process that pays back fastest. The fast payback is usually hiding somewhere nobody has looked in years, precisely because it is dull.
The shape of a quick win
Ninety-day automation has a recognisable shape. Four properties tend to be present together, and when one is missing the timeline roughly doubles.
- High volume. The task happens hundreds or thousands of times a month, so even a modest saving per instance compounds into something a finance director can see.
- Low variance. Eighty per cent of instances look alike. The remaining twenty can be routed to a person without anyone minding.
- A clear rule. Someone can already explain how the decision is made, even if it has never been written down. If the answer is 'it depends', you are looking at a judgment task, not an automation task.
- A human currently does it by hand. Not because it requires a human, but because nobody built the alternative.
Where to look first
In almost every company we have mapped, the same five places turn up. They are unglamorous, which is why they are still manual.
- Inbox triage. Shared mailboxes where a person reads, classifies and forwards. The classification is the automation; the forwarding is the integration.
- Document intake. Invoices, forms, orders and contracts arriving as PDFs and being retyped into a system. Extraction plus validation, with a human confirming the low-confidence cases.
- Reconciliation. Two systems that should agree and do not, checked by a person every week. The check is trivially automatable; the resolution often is not, and that is fine.
- Status reporting. Someone assembling the same update from the same sources every Monday. The report can write its own first draft.
- Onboarding steps. Account creation, access grants and welcome sequences that follow a checklist nobody enjoys.
Where not to look first
Customer-facing judgment calls, anything where a wrong answer is expensive and hard to reverse, and any process whose real rules live in one person's head and change weekly. These are legitimate targets — later, once the organisation has learned how to run an automated process and what the failure modes feel like.
The first automation is not there to save money. It is there to teach the company how automation fails, cheaply.
How to know it worked
Decide the measure before the build: hours returned, cycle time, error rate, or a cost line. Measure it for two weeks before anything changes, so the baseline is not a guess. Then measure the same thing after. If the number did not move, the automation did not work — however elegant it was — and the honest response is to stop, learn and pick the next candidate.