Your Sent Folder Is the Training Library

Last month I ran an audit on an eight-person insurance agency. Office manager had been quietly running the place for twelve years. Knew every carrier, every renewal window, every client who liked a phone call instead of an email. She pulled up their AI tool to show me their “AI deployment.” Four months in. Tool on the verge of being cancelled.

She typed in a renewal reminder for a long-time homeowner client. Draft came back — grammatically clean, politely generic, the kind of thing a stranger would write. She shook her head, clicked the draft, and rewrote the whole thing in about ninety seconds. Then she hit send.

I asked what she’d just done. She said “I fix it every time. That’s why I’m not sure this thing is worth it.”

Right. I asked her to open her sent folder. Scroll back. Twelve years of literally the exact voice she wanted the AI to write in. Hundreds of renewal emails she’d be proud to sign her name on. A pile of evidence the AI had never seen.

She’d been doing the hardest part of this work for twelve years. She just never told the AI.

The Examples ingredient, and why it’s the one

Quick callback, then we go. In Backfill #1 we laid out the Professional Recipe — the seven things any operator with sense gives a new hire on day one, and the seven most operators skip with AI. Training. Context. Guardrails. Examples. Output over process. Measurement. Feedback loop.

When I audit AI deployments at small and mid-sized businesses, the lowest-scoring ingredient — by a long shot — is Examples. Almost every operator scores zero or one out of three. Across industries, company sizes, owner backgrounds. It’s the universal gap.

It’s also the single highest-leverage gap. You close this one ingredient and output quality changes in an afternoon. Not gets better over weeks. Changes by lunch.

So if this is the move that fixes most underperforming AI, and it doesn’t take months, and it doesn’t take a developer — why does nobody do it?

That’s the piece. Not “here’s how to write examples for your AI.” That piece is already everywhere. The real question is — if Examples is cheap, fast, and high-impact, why don’t operators do it?

Three ways I’ve heard operators describe this — only one is right

Framing one — “AI is just kind of generic. That’s how it is.” This one’s comfortable. It puts the limitation on the tool. Nothing for the operator to do about it except wait for a better tool.

Framing two — “We haven’t shown it our best work yet.” Closer. At least this one names an action. But it implies a project — sometime we’ll create some examples — and that project slides down the list and never gets done.

Framing three — “I’ve been showing it every time. I just never put the showing anywhere it could remember.” That’s the one. That’s literally what was happening with the office manager. She wasn’t failing to teach the AI. She was teaching it in private — every rewrite — and then throwing the lesson away. The AI didn’t sound generic because she hadn’t taught it. It sounded generic because her teaching landed in the customer’s inbox, not in the system.

If your AI sounds generic in your business, I’d bet you’re running framing three too. You just haven’t named it yet.

The invisible work you’ve been doing

Here’s the thing. Watch yourself for a day. Any time AI drafts something — an email, a note, a quote, a summary — and you tweak it before it goes out, that tweak is an example. You just made one. You just decided what “good” looks like in your shop. You did the highest-skill, highest-value piece of the whole workflow.

Then you sent the email. The tweak went with it. Nothing went back to the AI.

Do that twenty times a week for four months and you have eighty hours of teaching that never got filed. The AI is running on the same blank starter kit it had in month one. You’re running on growing frustration. And when somebody asks how AI is going in your business, you say “eh — it’s fine. I rewrite most of it anyway.”

That’s not an AI problem. That’s a storage problem. The lessons are being generated. They’re just not being kept.

Your sent folder, your templates, your drafts

Here’s the shift. You don’t need to sit down and create examples from scratch. You already have them.

Your sent folder is a library. Every email you sent that a client responded well to is an example. Every renewal note that got signed on time is an example. Every complaint response that ended with the customer staying is an example. Years of them. Thousands.

Same story for your templates folder. Same story for the five phrases you’ve told every new hire never to change. Same story for the three email openers you always use because they work. You’ve spent your entire career curating the voice of this agency. The voice is already documented. It’s in your own writing.

You don’t need to create the examples. You need to curate them. Pick the good ones. Paste them in. Tell the AI this is the standard. That’s the whole move.

Twenty minutes of work. Output changes by the next email.

The part everyone forgets — show the misses, too

One wrinkle that makes the difference between a decent Examples layer and a sharp one.

Don’t just show the AI five great emails. Show it two emails that missed — ones that got a complaint, got ignored, or got you that “reply to me directly next time” phone call from a client. Add a short note. “This one missed because it sounded corporate. We’re warmer than that.” Or “This one missed because it quoted a price. We never quote a price over email.”

What good looks like, and what bad looks like, side by side. The AI learns the line between them. Does that make sense?

This is how experienced managers train people. You show them the work you’re proud of and the work you’re not. You point at the gap. The gap is where your judgment lives, and the AI can’t see your judgment unless you make it visible.

Three or four good. One or two bad. Short notes on why. That’s an Examples layer.

Why this one ingredient fixes most of the underperformance

If you’ve read Backfill #1, you know there are six other ingredients in the framework and they all matter. This isn’t a just do Examples and you’re done pitch. But here’s the honest thing I see in the field.

Other ingredients mostly control whether the AI knows your business in general — the shape of it, the boundaries, the goals. Examples control whether the AI knows your business specifically — the voice, the taste, the cadence that makes your customer think they got a response from a human who knew them.

Without Examples, AI runs on the average of the internet. Every business on earth gets the same average. That’s why every AI email in America reads like it was written by the same slightly over-eager intern. It was. The intern is the internet.

Minute you bake in examples, the AI stops drawing from the average and starts drawing from you. That’s where leverage shows up. That’s why this ingredient moves the needle in an afternoon when the others take weeks.

One more thing worth naming. Once you’ve added examples, keep them alive. A company changes — new services, new positioning, new kinds of customers. If your examples are frozen from eighteen months ago, the AI will sound like the company you used to be. Put a quarterly reminder on the calendar. Retire stale samples. Add new ones that reflect the shop you are now. Examples are not a one-time install.

The Monday Move

Twenty minutes. Don’t schedule it. Don’t make it a project. Do it this week, in one sitting, before it becomes one more thing on the list.

1. Open your sent folder. Pick five emails you’d be proud to hand to a new employee and say “this is how we sound.”
2. Pick two you’d flag as “please don’t do it this way.” Add a one-sentence note on why.
3. Open the AI tool you use most. Find its system prompt, custom GPT settings, or project instructions. If you don’t have one, make one — takes two minutes.
4. Paste the five good examples in. Paste the two misses in with the notes. Add one line — “The examples above are the standard. Write like this.”
5. Run your next three drafts through it. See what changed.

If nothing changed, you’re probably in a different ingredient gap (usually Output Over Process — you’re writing step-by-step prompts instead of describing the outcome). But almost always, something changes. Often everything changes.

That’s the test. And you’re not starting from scratch. You’re just putting the work you’ve been doing in public for years, for the first time, into the system that can actually use it.

The shift

So stop fixing your AI in private. The lessons you’ve been writing into your sent folder for years are the training material the AI has been starving for. You don’t have an examples problem. You have a filing problem.

Pull the work out of the inbox. Put it in the system. Let it compound.

That’s the recipe.

Framework: The Professional Recipe — Examples ingredient (#4). Related failure modes: the Generic Output Trap and the Lonely Examples Problem (examples added once, never refreshed). Cataloged in AI Onboarding Failure Modes.

Companion piece: Stop Hiring AI. Start Building It. — the parent argument. If that one set the posture (you’re building, not hiring), this one closes the single highest-impact part of the build.