A month ago I walked through an independent pharmacy in Denver. Office manager I’ve known for two years. Twelve years in the chair. Knows every customer. Runs the place with quiet precision.
She pulled up the AI system they’d set up for patient communication — medication-ready notifications, insurance-denial explanations, refill reminders. Four months live.
She drafted a notification to tell a regular patient her prescription was ready. AI output came back — polite, generic, the kind of thing every other pharmacy’s AI would write. The kind of thing that sounds like nobody wrote it at all.
She shook her head. Took ninety seconds. Rewrote it completely in her own voice. Sent it.
I asked her what she’d just done.
“I fix it every time,” she said. “That’s why I’m not sure this is worth it.”
Here’s the thing — she’s right about the pattern. She’s completely wrong about the cause.
The failure mode that doesn’t announce itself
If you’ve read Backfill #1, you know the Professional Recipe — seven things you’d give any new hire on day one, seven things that make the difference between an underperforming AI tool and one that actually works.
When I audit AI deployments across small and mid-market businesses — I’ve done dozens now — one ingredient scores lower than all the others combined. Examples. Ingredient #4.
Almost every operator. Zero out of three. Across industries. Across business sizes. Across vertical. It’s the universal miss.
It’s also the single highest-leverage fix. Close this gap, and output quality changes in an afternoon. By lunch.
Right? So the question that matters is — if it’s that high-leverage and that fast, why doesn’t anyone do it?
I thought the answer was “they haven’t made examples yet.” That they’re waiting for time that never comes.
Eighteen months of audits taught me different.
The invisible work you’ve been doing
That office manager didn’t have an examples problem. She had a filing problem.
Watch what happened in four months.
Week one. AI drafts renewal reminder. She rewrote it. That rewrite — that’s an example. She just decided what “good” sounds like in her pharmacy. She did the highest-skill, highest-judgment work in the whole workflow.
Week two. Insurance-denial explanation. AI draft — corporate, unhelpful. She rewrote it with specific detail and a clear next step. Example number two. Filed nowhere.
By month four. She’d rewritten AI output maybe eighty times. Eighty lessons generated. Eighty times she demonstrated her judgment. Eighty times she taught the system what this pharmacy sounds like.
Then the lesson evaporated.
AI never saw it. Output went to the customer. Teaching went with it. System learned nothing.
So next month, week one — generic renewal reminder. Rewrite. Again.
That’s the Lonely Examples Problem. And it’s not what it sounds like.
Here’s the thing — it’s not that you lack examples. It’s that you’ve been teaching all along. Every rewrite. Every “no, say it warmer.” Every time you tweaked something before sending. You’ve been running an unpaid training program for an AI that never gets the lesson.
Office manager I talked to had been doing this alone for months, building muscle memory of what “good” sounds like, and getting zero compounding because the teaching never made it into the system.
The specific failure — frozen voice
Here’s where it gets sharper. Once you add examples to an AI system, you have to keep them alive. Your business changes. You add new services. You rebrand. Your customer mix shifts. Your tone evolves.
If your examples are frozen from eighteen months ago, the AI will sound like the company you were eighteen months ago.
Another office manager at an independent pharmacy I know added five examples back in October 2024. Good examples — actual patient communications she’d be proud of. AI improved overnight. She felt vindicated.
Then came Q1 2026. Pharmacy launched a new medication therapy management service. New positioning. New language. New patient types.
AI had never heard about it.
Examples didn’t refresh. AI was faithfully reproducing the voice of a business that no longer existed. New services, new customers, new positioning — none of it showed up. AI sounded outdated without being wrong.
Office manager didn’t notice for two months. When she did, she blamed the tool. “AI was working great, now it’s stale.”
Yeah. No. Example set was unmaintained.
This is the lonely part of the name. Examples don’t die quietly. You add them once, and they sit there, frozen. They look fine. They sound fine. But they’re slowly becoming a portrait of who you were.
What this looks like in a pharmacy
For independent pharmacies, this shows up most in patient communication.
You develop a voice over twelve years. Warm. Specific. Patient-centered. You know your customers by name. You explain insurance denials with detail. You don’t hide behind corporate language.
That voice is gold. Customers trust it because it’s real.
When you deploy AI to handle patient communication — refill reminders, status notifications, problem resolution — the AI doesn’t have that voice unless you show it.
You show it by adding examples. Five or six messages you’d be proud of. AI learns the pattern. Next draft — sounds like you instead of like the internet.
But six months later, the pharmacy adds a new service. Patients shift. Tone needs to warm up for a certain segment. Examples are frozen. They’re still good examples, but they’re examples of the old pharmacy.
AI sounds like it’s stuck in 2024. Which it is.
The Monday Move
So. Calendar this week. Recurring reminder. Quarterly.
When it fires:
- Open your sent folder. Pull five customer communications from the last ninety days that feel like your best work.
- Pull two that didn’t land well. Add a one-sentence note on why.
- Open the AI tool’s system prompt or custom GPT settings. If you don’t have one, build one in two minutes.
- Paste the five good examples in. Paste the two misses. Refresh the instruction — “The examples above are the standard. Write like this.”
- Hit save. Done.
Next time you draft communication, it will sound fresher because the examples are fresher.
This is not a one-time install. This is a seasonal maintenance task. Set the reminder. Do it every three months. Retire old examples, add new ones.
Cost — thirty minutes per quarter. Impact — your AI stays anchored to who you are now, not who you were.
The shift
Stop fixing your AI in private. The moment you catch yourself rewriting a draft — that’s data. That’s an example. That’s evidence of what your judgment looks like.
Right now that judgment lands in the customer’s inbox and disappears. Start filing it. Every rewrite that gets sent is a missed lesson the system should have seen.
Your sent folder is your training library. You’ve been building it for years without calling it that. Pull the best work out. Show it to the AI. Keep it current.
So. Stop teaching in secret. Start filing the lessons.
Framework: The Professional Recipe — Examples ingredient (#4). Failure Mode: The Lonely Examples Problem. Companion piece: Your Sent Folder Is the Training Library.
