Two years into the enterprise AI wave, a pattern is visible: the deployments still running are rarely the impressive ones. The chatbot that promised to answer anything is gone; the document extractor that reads supplier invoices into the ERP, every night, is quietly saving four hours a day.
The survivors share three properties. First, they are grounded: they answer from the organisation's own documents and records, with access control intact, rather than from a model's general knowledge. Second, they keep a human in the loop wherever money moves or commitments are made — AI drafts and extracts; people approve. Third, they were measured against a baseline from day one, so their value is a number, not a feeling.
That discipline points to where to start: document-heavy, judgement-light workflows. Invoice and purchase-order capture. First-draft reconciliation matching. Summarising a month's exceptions for review. Answering 'what is our policy on X' from the policy library. None of these makes a keynote; all of them return hours every week, and none of them frightens an auditor.
The controls conversation is where most programmes stall, and it needs to happen first, not after the pilot. Where does the data go? Which model sees it, and is it retained? Who approved the output before it posted? An AI integration designed around those questions passes review; one retrofitted to them usually does not.
Start boring. Measure honestly. Expand from evidence. The organisations doing this are building a compounding advantage — one unglamorous workflow at a time.