Where AI actually pays off in a business, and where it doesn’t

Every week a new demo makes AI look magical. But a demo runs once, in perfect conditions, for an audience that wants to be impressed. A business runs the same task ten thousand times, in messy conditions, for people who just want it done. The gap between those two is where most “AI projects” quietly die.

Start from a metric, not a model

The teams that get real value from AI never start with “let’s use AI.” They start with a number that hurts, hours lost to manual data entry, slow first-response times, a support queue that never empties, and then ask whether intelligence can move it. If you can’t name the metric, you’re buying a demo, not an outcome.

Three shapes of problem AI is good at

• Reading and routing, classifying, tagging, and triaging a flood of unstructured text (tickets, emails, documents).
• Answering over your own data, retrieval-based assistants that cite real sources instead of guessing.
• Removing repetitive judgement, the small decisions a human makes hundreds of times a day that follow a learnable pattern.

If a task is high-volume, pattern-heavy, and tolerant of a human check at the end, it’s a strong AI candidate. If it’s rare, high-stakes, and unforgiving of error, it usually isn’t, yet.

Where it usually disappoints

AI struggles where the cost of being confidently wrong is high and there’s no cheap way to verify the answer. It also disappoints when the underlying data is a mess, no model rescues you from that. Clean data and clear guardrails do more for an AI feature than a bigger model ever will.

The honest test

Before you build, answer three questions: What metric moves? How will we know it’s wrong? What happens when it is? If you have crisp answers, you have a project. If you don’t, you have a demo, and demos don’t pay salaries.