A sales manager at a small distributor sits at his desk on a Tuesday morning, rewriting an AI-drafted follow-up email for the fourth time this week. He is not a lazy operator. He is the person whose voice the customers know, and he is not sending anything that sounds off. The rewrite takes him twenty-two minutes. The customer opens the email at 3:14 that afternoon and replies at 3:31, exactly the way that customer replies to everything he sends.
He does not know it yet, but he just paid himself twenty-two minutes in confidence for a rewrite the customer could not see.
The customer cannot see the polish. Measure what the customer can see.
Here’s the thing. That sales manager is not wrong to guard voice. He is the last human to touch the work before it goes out the door. He knows the accounts by name. He knows which customer wants three sentences and which one wants ten. The AI draft looked fine. It read close. It read almost right. And almost right felt riskier to send than definitely him. So he rewrote it. And he had rewritten every AI-drafted email he had touched all month. And the rewrites felt like quality control the whole time.
His owner asked a question at the end of the month. “Are the rewrites moving anything?”
He did not have an answer.
The sent folder flipped the frame
We pulled three months of his own sent emails, filtered to the same kinds of follow-ups the AI was drafting. Sixty-eight of them. Real ones. His voice. His rhythm. His account-by-account calibration.
Then we pulled a set of AI drafts the AI had produced against the same customer situations. Side by side. Same length. Same three-sentence opener. Same one specific ask at the end. Same warm-but-brief sign-off he had been using since 2019.
He read three of them and said “I would have sent that.” He read three more and said “I did send that.” He was not being generous. The AI had learned from his sent folder. It was matching the pattern because the pattern was in the data.
Yeah. That was the whole first insight. He had been rewriting the AI to sound like himself, and it had already sounded like himself.
What the phantom actually is
The missing polish was not in the words. It was in his sense of authorship.
He needed the rewrite to feel like the email was his, because if it went out sounding like him without him actually touching it, some part of his job was moving somewhere he could not track. That is not a small feeling. That is real. That is a person doing quality control on his own value inside a business that is changing under him.
The trap is what the feeling costs. Twenty-two minutes per email. Two hundred and twenty minutes a week. A full workday every two weeks. And the customer cannot see any of it. Not one of those twenty-two-minute rewrites moved a reply rate, a booked call, or a next step.
He was paying himself in confidence by typing words that did not move the customer.
Our Professional Recipe has seven ingredients that make an AI-native workflow hold. The one that names the fix here is the seventh. It is called Feedback, and inside it is the half he was skipping. The instrumentation half. The one where you decide what metric maps to value, and you count it.
Plain translation: send it two ways and count what came back.
Feeling is not evidence
The sales manager had a feeling. Feeling is not evidence. Feeling is a signal that something is worth measuring. It is not the measurement.
What turns the feeling into a test is a small side-by-side. Send three AI drafts as-is. Send three manual rewrites. Same customer type. Same situation. Same week. Then look at the customer response and count what came back.
We ran that test the next week. Six emails. Three AI, three rewritten. The response rate was identical. The manual rewrites did not beat the AI drafts on reply rate, on booked next step, or on time-to-response from the customer. Six is not a giant sample. It was enough. Enough that when he ran a second round of eight the following week and got the same result, he stopped rewriting most of them.
He is not doing zero edits now. He caught two AI drafts in the second round that would have gone to the wrong customer segment with the wrong ask. Those edits were real. The other twelve edits he had been doing every week? Gone. Not because we told him to stop. Because he counted, and the count did not support the effort.
Examples did quiet work upstream
One more callback, then the Monday Move. Do not skip it. The reason the AI drafts sounded that close in the first place is that the sent folder had already loaded the fourth ingredient. Examples. Sixty-eight of his own emails, ranked implicitly by which ones had actually earned a reply. The AI was writing off a pattern of good, because the pattern of good was already sitting in the data.
Examples got the draft close. Feedback proved whether close was enough.
That is the recipe move. If Examples had been thin, the AI drafts would have been generic and the rewrites would have been earning their keep. Because Examples was rich, the rewrites were mostly touch. The recipe is what tells you which ingredient is doing the work and which one is being paid to do work that was already done.
Where phantom polish shows up next to you
The distributor is the story we ran the test in. It is not the only place the pattern lives. Look at your own team this week.
The customer service rep who rewrites the AI’s estimate reminder because it does not sound warm enough. Warm compared to what? The version she was going to send on her own? Or the version the customer needed, which was the next step in one clean sentence?
The account manager who edits an AI-drafted renewal nudge because it feels too generic. Generic compared to what? A rewrite that changed the offer? Or a rewrite that just changed the rhythm?
The intake coordinator who rewrites the AI’s first-touch reply because a sentence “does not flow.” Flow compared to what? A version the customer would have preferred? Or a version that made her feel like she still writes the emails?
Every one of those rewrites is real work with a real time cost. Some of them are catching real misses. Some of them are phantom polish. The recipe does not tell you in advance which is which. The recipe tells you to send it two ways and count what came back. Then the recipe gets updated by what the count shows.
Here’s the thing that hides the trap. The rewrite feels like quality control right up until the moment you measure it. The moment you measure it, some of it turns out to be. Some of it turns out to be a person paying themselves in confidence for a change the customer literally could not see. Both are possible. Both live in the same team, the same week, sometimes the same email. What sorts them is not opinion. What sorts them is response rate, booked next step, or customer completion.
Does that make sense? The rewrite has to earn its cost against a metric. Not against a feeling.
The Monday Move: run the six-draft response test
Pick one recurring email your team rewrites every week. Sales follow-up. Estimate reminder. Renewal nudge. Intake reply. Something routine, low-risk, where the AI version is already reading close because the Examples are already in the data.
Owner of the move is two people. The person who usually rewrites the AI draft, plus the person who can see the customer response. It has to be both, because the writer sees the effort and the observer sees the outcome.
Metric: response rate, booked call rate, or customer next-step completion. Pick one. Track it for both sets.
Guardrail: do not run this on angry customers, pricing disputes, legal issues, or anything with sensitive detail. This test is for the routine work where the AI version already looks close, not for the high-stakes work where a wrong word matters.
Next step: send three AI drafts as-is. Send three manual rewrites in the same week. Compare what the customer did. If the metric is the same, stop paying for phantom polish on that email type. If the AI underperforms, update the Examples set or the voice rule and test again.
That is not speed over polish. We already wrote about that in Speed Beats Polish. This is measurement over gut. A different axis. A different move. This one lives on the counting side of the recipe, not the drafting side.
It also lives on the second hill above Mount Stupid. The parent piece was about the operator who cannot see the next climb because the first win feels like the whole map. Phantom polish is one specific first-hill motion. The rewrite feels like the win. The test is what shows you whether the win moved the customer.
So.
The customer cannot see the polish. Measure what the customer can see.
If the test says the rewrite matters, keep rewriting. Now you know it is quality control. If the test says the rewrite does not move the customer, stop paying yourself in confidence and get twenty minutes back on Tuesday morning.
Either way, the feeling is not the answer. The count is.
Original framework. Distilled from client work.
Framework spine: The Professional Recipe, the Feedback ingredient’s instrumentation half, with a callback to Examples. Read the Professional Recipe.
~ source material · Original framework. Distilled from client work.
