Skip to content
← Writing

Agents, not demos

Why most AI pilots stall — and the three questions that decide whether an agent ever reaches production.

11 Oct 2026·1 min read·AI-assisted

Every business owner I meet has seen an impressive AI demo. Very few have an AI agent quietly doing real work in their company. The gap between the two isn't the model. It's everything around it.

The demo trap

A demo answers one question: can the model do this? Production asks harder ones:

  • Who owns it? An agent without an owner is a science project.
  • What is it allowed to touch? Every tool an agent can call is a permission you are granting.
  • How will we know it's working? If you can't measure hours saved or revenue moved, you can't defend the budget.

Treat an agent like a new hire

The most useful mental model I know: write the job description first.

  1. Role — one sentence on what this agent is for.
  2. Tools — the systems it can read and write: CRM, inbox, calendar, WhatsApp.
  3. Limits — what it must never do without a human.
  4. Manager — the person who reviews its work every week.

If you can't write the job description, you're not ready to build the agent.

Start small, ship fast

Pick one workflow that is frequent, boring and measurable — lead qualification, meeting follow-ups, invoice chasing. Ship it to a small group in two weeks. Measure for a month. Then decide whether to scale, change or switch it off.

That's how AI stops being a deck and starts being a line in your P&L.