Why most SMB AI projects fail in week three

---

title: "Why most SMB AI projects fail in week three" description: "The first two weeks are champagne. Week three is when the system meets the inbox." date: "2026-04-25" readingMinutes: 6

Most AI projects in Australian SMBs do not fail because the model was wrong. They fail in week three, when the demo meets the actual inbox.

The pattern

Week one is exciting. The owner sees the demo, signs the contract, briefs the team. Week two is configuration: connecting Gmail, importing the CRM, drafting the first prompts. The system responds well to test cases. Everyone is optimistic.

Then week three arrives. A real customer sends a real email at 6pm on a Friday. The agent drafts a reply. Nobody knows whether to send it. The owner is at dinner. The admin is gone for the weekend. The draft sits.

By Monday, the customer has phoned a competitor.

The pattern beneath the pattern

The technology worked. The integration worked. What failed was the operating model around the technology — who approves, when, on what device, with what authority, against what risk threshold. None of that was specified in week one because the demo did not surface it.

This is why we sell governance before we sell agents. The approval matrix is not paperwork. It is the single thing that determines whether an AI workflow survives contact with a real business week.

What we do differently

Every AAO pilot starts with a one-page approval matrix:

• Who can approve this action?
• On what device?
• Within what timeframe?
• What happens if nobody approves?

The matrix is signed by the owner before any agent goes live. It is the contract between the business and the system. It is also the document that makes week three survivable.

The takeaway

If you are evaluating an AI vendor and they cannot show you the approval matrix in the first meeting, you are buying technology without an operating model. Buy the operating model. The technology is the cheap part.