The Dimmer, Not the Switch.

The accounting station was drafting AP follow-ups for about a week before it asked me a question.

It said. “I noticed you CC Hailey on every payable over $5,000, but only sometimes on smaller ones. Is the threshold the dollar amount, or is it specific vendors? Want me to make $5,000 the standard, or should I keep flagging the edge cases for you?”

That’s the moment the rollout actually worked. Not when we deployed it. Not when it drafted its first email. When it asked the question.

A year ago I would have stood that station up differently. Written the prompt, given it the examples, set it loose, and waited for the inevitable Slack message from Hailey saying “hey, why didn’t I get CC’d on the Acme invoice?” Three weeks later I would be back in the spec rewriting the rules I should have asked about on day three.

That’s the switch. Deploy, hope, fix what breaks. We don’t run that play anymore.

What we run instead

Five stages. Each one earns the next. The agent doesn’t move up the ladder until the previous rung is solid. I call it the dimmer because that’s what it is. You turn the dial up gradually as trust builds, and you can dial it back down the moment something looks off.

This is what the Four D’s calls Dialing, by the way. Standing at the pass, refining the recipe, while the station is cooking. Most AI deployments fail at this stance because the Chef retreats to the line and rewrites the whole prompt instead of staying at the pass and letting the agent do the work of finding what’s missing.

Here’s the same accounting station, climbing the dimmer.

Stage 1. Pull and analyze. Before the agent drafts anything, it reads. Last 90 days of AP follow-ups Hailey sent. The vendors. The amounts. The CC patterns. The tone shifts between long-term partners and one-off invoices. The agent isn’t writing yet. It’s mapping how the work actually gets done. Most rollouts skip this and start at stage 2. That’s why the first drafts come back generic. The agent never saw what good looks like before being asked to produce it.

Stage 2. Shadow drafts. Now the agent drafts, but nothing goes out. Hailey writes her version. The agent writes its version. They sit side by side in a doc. We compare. Not “is the AI version good”, the wrong question. Where is it different from Hailey’s, and why? That gap is the spec we’re missing. Maybe Hailey’s tone with vendors over five years is warmer than the agent caught. Maybe she opens with a thank-you that the agent skipped. The shadow stage is where we find the silent rules. The things Hailey does without thinking that the agent has no way to know.

Stage 3. Clarify, daily. This is the stage everyone skips. And it’s where the deployment lives or dies.

The agent surfaces a few questions a day. Not “should I send this.” Pattern questions.

“I see you change tone for two of the long-term vendors, both over three years, both warmer than the rest. Should I treat any vendor over three years that way, or is it specific to those two?”

“You CC’d Hailey on three of last week’s drafts but not the others. The three were all over $5,000. Is that the rule, or is it about the vendors? Want me to standardize?”

“You changed ‘past due’ to ‘overdue’ twice last week. Preference, or context-specific? Want me to make that the standard?”

Three questions a day. Five minutes. Each one closes a gap in the spec the agent couldn’t have closed alone. By week two the agent isn’t asking anymore. It’s confirming. “Drafting the Acme follow-up now, $7,200, CC’ing Hailey per the standard. Anything different here?” By week three the questions stop because the spec is sharp. The agent has literally learned your judgment by demonstrating it’s been paying attention.

Here’s the thing. Most operators retreat to the line at this stage and rewrite the whole prompt themselves. They feel the friction of the agent asking three questions a day and they go, “forget it, I’ll just write it.” That’s the move that kills the rollout. The dimmer says, let the agent do the work of finding what’s missing. It’s standing at the pass next to you. Use it.

Stage 4. Draft and review. Agent drafts the full batch. You read before send. Not rewrite. Read. If you’re rewriting at stage 4, you’re not done with stage 3. Go back, surface the pattern that’s still off, update the spec, try again. Stage 4 is the test of stage 3’s work. When the drafts read clean and you’re hitting send without edits, you’ve earned stage 5.

Stage 5. Autonomous. Agent drafts and sends. You see the output in your sent folder. You spot-check weekly, not daily. The dimmer is at full bright.

The thing about the dimmer. You can turn it back down any time. If the agent’s outputs start drifting in week six, you don’t tear the deployment apart. You drop back to stage 4 for a few days, find the pattern that broke, fix the spec, climb back up. The dimmer is bidirectional. Right?

Why stage 3 is where the magic is

Stages 1, 2, 4, and 5 are pretty standard onboarding logic. Anyone who has deployed an agent has seen some version of them.

Stage 3 is the part that changes everything.

Most rollouts go shadow to autonomous. They skip the clarification dialogue because it feels like the agent is asking too many questions and it’s faster to just write the spec yourself and ship. That’s the shortcut that costs you. When you write the spec yourself you write what you think the rules are. You miss the rules you follow without naming them. The ones a new hire would catch in week one by asking “wait, why did you do it that way?”

Yeah. That’s the move. Stage 3 is the agent doing the new-hire move. It’s the agent saying “I noticed something. Help me understand.” And every clarification you give it sharpens the spec in a way you couldn’t have sharpened it alone, because you didn’t know it was unclear until the agent named the pattern.

Skip stage 3 and the agent never learns your judgment. It only learns your stated rules. Which is why most deployments stall at “good enough but not quite right” forever. Does that make sense?

The dimmer, not the switch

Old rollout. Flip the switch. Deploy and pray. Find out three weeks later you missed a rule, because the agent never had a way to surface what it didn’t know.

New rollout. Turn the dial. Coffee, then a meal, then real trust. The agent earns each stage by demonstrating it’s been paying attention. You earn each stage by feeding it the judgment calls only you can make.

Trust in an agent is a dimmer, not a switch. And the dimmer doesn’t move because time passed. It moves because the agent showed you it was learning, and you confirmed.

The Monday Move

Pick one agent you’ve deployed.

Honestly. What stage is it actually at? If you flipped it from shadow to autonomous without a clarification phase, that’s why it’s drifting. Drop it back to stage 3. Have it surface three pattern questions a day for a week. Watch what changes.

The dimmer goes both directions. Use it.

Original framework. Distilled from client work.

Framework spine: The Four D’s, the Dialing stance, where most deployments die. Read the full framework. See also: The Silent Critic (closing the feedback loop) and The Prompt Hoarder (one shared spec, not five private ones).