Start narrow. Widen on proof.

The fastest way to kill this kind of project is to try everything at once.

Five half-built projects each get a fifth of the attention. None of them produces a result anyone can point to. The team decides "this doesn't work here", and that story kills the next attempt.

So we start with one use case, prove it, and only then add the next.

Picking the first one

In the Review, we score each candidate from 1 to 10 on four questions:

• Can we build it with the tools and data you already have?
• What's it worth, in a number you'd put in a board update?
• How likely is it to work, given your data and your team today?
• How little effort does it take, including getting your team to change habits? (10 means very little.)

Then we multiply the four scores, rather than adding them, so one weak spot can't hide behind three strong ones.

An illustration, with three candidates:

The big idea loses because it probably won't work, and in a small industry it could do harm. The easy one loses because it isn't worth much on its own. The winner is likely to work and worth showing your board.

Scores are illustrative. Yours come from the Review.

Agree on success before building

"More referrals" isn't a goal. This is:

"Every customer who reaches first value gets an introduction ask from the person who knows them, within a week. We count the introductions made, from our starting point to the target we signed."

Three parts, written down:

• The starting point. Measured, not guessed.
• The target. A number, set from your own starting point.
• The scope. What's in, and what isn't.

In the Sprint, all three are signed by day 45. Without them, "did it work?" comes down to opinions. With them, the answer is a number.

How trust grows

An agent earns room to act the same way a new hire does.

1. It drafts. A person sends. Every action is checked.
2. It handles the routine ones. When people have sent its drafts unchanged for a while, it may send the routine ones itself. Anything unusual still goes to a person.
3. Its limits widen as its record proves it. Each step is based on what the log shows, not on a feeling.

Some things never move to an agent. Pricing, contracts and an unhappy customer always go to a person.

Every exception teaches the system

When a person handles something the agent couldn't, that case is logged: what happened, what the person did, how it turned out.

The next time something similar comes up, the agent suggests what worked last time. A person approves it. After enough of those, it becomes routine, and the agent's limit moves up.

Each case handled makes the next one easier. That's how the loop gets more reliable over time, not just more clever.

Reliable beats clever

An agent that's brilliant most of the time and badly wrong some of the time is a risk once it runs every day.

The arithmetic: an agent that drafts 40 introduction asks a month and gets one in twenty wrong puts two wrong messages a month in front of your best customers. In an industry this small, people talk about both.

So we build for steady and predictable first. Clever comes second.

A person stays accountable

An agent can do the work. A named person on your team answers for the result. That never moves to the agent.

Be in a hurry to start. Be patient about how much you let agents do on their own.