The smallest useful AI integration

Most failed AI projects didn’t fail because the model was bad. They failed because the scope was too big.

Companies we talk to about AI usually want one of two things. They want an “AI strategy”, a six-month engagement that produces a slide deck nobody implements. Or they want to “transform” something, replace an entire workflow with an AI version, with no idea whether the new version will be better.

Neither one is a project. Both are noise.

What a useful first pilot looks like

The shape we keep coming back to: a doc-aware assistant for a single internal team, scoped to one well-defined question type, grounded in one well-defined data source, deployed to one well-defined audience.

Concrete example: support team has a 200-page knowledge base. Every new ticket starts with the agent searching it for relevant policies. We build them a chatbot that takes a customer question and returns the matching policy snippets, with sources. That’s it.

Why this works:

• Bounded scope. One team, one data source, one task.
• Easy to evaluate. You can sit a support agent in front of it for an hour and immediately tell whether it’s useful.
• Low downside. If the AI gets it wrong, the agent ignores the suggestion and answers the ticket the way they would have anyway.
• Real upside. Even modest accuracy (~80%) saves measurable minutes per ticket. At any meaningful ticket volume, that pays back the build cost in months.

What we don’t do first

Multi-agent autonomous workflows. Replace-an-entire-team systems. AI strategy decks. “Transformation” frameworks. Anything that requires three teams to integrate before it ships.

Not because those things can’t work, sometimes they do. But they require an order of magnitude more investment, with much higher risk of producing nothing usable. We start small for the same reason a chef starts with a sauce: prove the technique works before scaling it up.

The honest version

The smallest useful AI integration is probably worth $15-30k of build cost. It probably saves the team it’s deployed to more than that in the first year. And it gives you something concrete, a working system, with measurements, to base the next AI decision on.

That’s a much better foundation than a strategy doc.