Every owner I talk to has been pitched AI. Few have been asked what they would actually use it for. That question does more work than any tool comparison.
The pattern that works
AI earns its cost on tasks that are high volume, low stakes per instance, and reviewable. Drafting first versions of routine communication. Summarizing long documents into something a person then checks. Sorting and tagging inbound inquiries. Turning messy notes into structured records.
Notice what those share: a human still owns the outcome, and being wrong once is inconvenient rather than damaging.
The pattern that fails
It goes badly on judgment calls with real consequences and no review step. Pricing decisions, anything client facing sent without a person reading it, compliance sensitive work, and any process where nobody can tell whether the output was correct.
It also fails when the underlying process is broken. Automating a workflow nobody understands produces faster confusion. Which brings up the more common problem.
Most AI projects are actually data projects
The businesses that get value have their information somewhere a system can reach. The ones that struggle have it spread across a spreadsheet, someone's inbox, a notebook by the register, and institutional memory.
If your customer records live in four places that disagree, no tool fixes that. It reflects it back faster. The unglamorous work of getting your data into one reliable place is what makes everything after it possible, and it is usually where I spend the early part of an engagement.
Four questions before you buy
What specific task, and how often? If it happens twice a month, automation will not repay the setup.
Who reviews the output? If nobody, either add review or do not deploy it there.
What happens when it is wrong? Wrong tone in a draft is fine. Wrong number on an invoice is not.
Does this replace a tool you already pay for? Frequently yes, and that changes the math from new cost to substitution.
Start narrow
The successful pattern is one task, one team, measured for a month against how long it took before. Small enough to abandon without regret, real enough to prove something. What you learn from that first task tells you far more than any vendor demonstration about whether the second one is worth doing.