Start with the repetitive work, not the demo
Useful AI sits inside a job people already do: qualifying an enquiry, drafting a reply, extracting fields from a document, routing a ticket. If you cannot name the job, you are buying a demo.
Measure hours or revenue, not novelty
The question after launch is whether a person got time back or a customer got a faster answer. If you cannot measure either, the project will stall at a slide.
- Which workflow, and who owns it today
- What data the model is allowed to see
- Where a human must still approve the output
- What 'better' looks like in a week, not a quarter
Keep a person in the loop on purpose
Automation that cannot be supervised becomes a liability. Design the handoff: the model drafts, the person sends. Widen autonomy only after the error rate is boring.