Automate Transactions, Augment Judgment: Where AI Stops in a Small Business

One principle runs through every chapter of the book: Automate transactions. Augment judgment. Escalate exceptions. Preserve human relationships.

It sounds like a slogan. In practice it is a design rule, and it answers the question most AI rollouts skip: what is this system allowed to do on its own?

Drafting is not sending

Calling every tool an "agent" hides that question. Drafting a reply is different from sending it. Suggesting a payment is different from releasing it. Decisions about people, money, contracts, safety, and customer relationships need an accountable human, even when AI helps prepare them.

Two questions, four answers

The Human / AI Decision Matrix places work on two scales:

1. Consequence if wrong. What happens if an error goes uncaught?
2. Judgment required. Could two competent people reasonably differ?

Where a piece of work lands sets the most AI may do with it:

• Automate (low consequence, low judgment): AI or rules do the work, with monitoring and a way to undo mistakes.
• Draft (low consequence, high judgment): AI writes a first version; a person edits and owns it.
• Approve (high consequence, low judgment): AI prepares; a person approves every item before anything happens.
• Human decides (high consequence, high judgment): AI may gather facts but not rank or shortlist. The person sees the sources, not just a summary.

The quadrant is a ceiling, not a target. And relationship work goes no further than Draft, whatever else it scores.

Place first, control second

It is tempting to say, "It's high-impact, but we'll check it, so it's fine." That reverses the logic. The matrix decides which controls you need, so the controls cannot decide where the work sits. Place the work by its consequence before you think about controls.

The review that is not a review

A human checkpoint only counts if it is real. The signs of a rubber stamp: nearly everything is approved, complex items clear in seconds, and reviewers cannot see the source the AI used.

The corrections are unglamorous: put the evidence beside every proposal, keep the volume readable, re-check a sample every month, and record what reviewers change.

Test the boundary against last month

Pull five real cases from last month - an odd refund, an unusual complaint. Walk each one through the boundary you just drew. Wherever your table does not say who decides and what AI may do, rewrite the row before any tool depends on it.

AI is very good at transactions. Your people are still where judgment, accountability, and relationships live. The best AI design keeps it that way on purpose.