Every team we meet has a favourite AI use case. It is almost never the one that pays first.
The favourite is the one that demos well: a chat assistant on the website, a campaign generator, something with a visible wow. The one that pays is the handoff nobody wants to own — the account brief a rep rewrites every morning, the quote that waits two days for an approval, the CRM fields that only get filled during a pipeline review.
Find the step people do twice
Duplicated work is the clearest signal. If two people assemble the same context from the same sources, that context should be produced once, by an agent, and reviewed by a human. It is unglamorous and it compounds.
The model is the cheap part. The handoff is where the money is.
Ship it with an owner
An agent without a named owner degrades quietly. Pick the person who feels the pain, give them the evaluation set, and make the weekly review five minutes long.
Then earn the interesting one
Once the boring agent is trusted, the team has the muscle — data hygiene, review habits, guardrails — to take on the use case they actually wanted. That order matters more than the tooling.