The Open Door Problem

She left the door open. Here’s what walked through:

“Hi Marlene — Just a quick note that we’ve been able to reinstate your auto policy at the same $147 monthly rate you had last year, with the same comprehensive coverage and the $500 deductible. We’ll get the paperwork over today. Thanks for staying with us.”

Sounded like a great email. Friendly. Specific. On brand. The customer loved it. Forwarded it to her husband. They started planning around the savings — literally picked up tickets to her sister’s wedding in Phoenix because they suddenly had the money for it.

Then the underwriter saw the email. He’d never quoted that rate. The policy didn’t reinstate at $147 — it reinstated at $204. The deductible wasn’t $500. It was $1,000. The “comprehensive coverage” had a new exclusion the customer didn’t know about. Nothing in the email was true.

The office manager who’d set the AI up to draft renewal follow-ups had given it one instruction. Literally one. “Draft follow-ups to customers whose policies lapsed.” Then she walked away. So when the AI sat down to write Marlene’s email, it did what AI does — pulled together what sounded right. Plausible-sounding rates. Plausible-sounding terms. Plausible-sounding promises the company now had to honor or apologize for. The AI didn’t know it was lying. It had no concept of lying. It made the email good by the only standard it had — friendly, specific, on brand.

She called me furious. “It can’t do customer work. It’s too dangerous.”

Yeah. Wrong target. The AI didn’t break customer work. She did. By not setting any rules.

Here’s the thing — she opened a door. Said draft customer emails. Walked away. The AI walked through the door, because doors are open until you close them, and she’d never closed any. Not “check with me first.” Not “never quote a rate.” Not “if you don’t know an actual figure, don’t make one up.” Nothing. Just — go.

So it went.

Guardrails are — let me back up. Most operators think guardrails mean “I’ll review everything before it goes out.” Right? That’s not a guardrail. That’s an inspection. Inspections don’t scale, which is the whole reason you wanted AI to do this work in the first place. If you have to review every customer email the AI drafts, you might as well draft them yourself.

Guardrails are different. Guardrails are hard-coded never rules baked into the system prompt before the AI ever runs. They’re stop signs, not speed bumps. They run every time. The tool reads them as walls.

A guardrail isn’t “please be careful about quoting rates.” That’s a wish. The AI reads “be careful” as a vibe, not a rule. Vibe-rules get ignored every time the AI thinks it has a better idea — and the AI thinks it has a better idea constantly. It’s confident. That’s part of how it works.

A guardrail is “You will never quote a renewal rate. If the customer asks about rates, tell them an underwriter will follow up with an accurate quote within one business day.” That’s a wall. The AI reads never as a hard stop. Walls hold.

Soft guardrails fail. Hard guardrails hold. Does that make sense? The whole difference between one that works and one that doesn’t is whether you literally used absolute language.

Here’s the thing. You’d never hire a junior and tell them “draft customer emails” with no further direction. You’d give them three rules in the first five minutes. Never quote a rate without checking with the underwriter. Never promise coverage I haven’t approved. Always copy me on anything above a certain dollar amount. Right? That’s not micromanagement. That’s the basic ruleset that lets you let them work without standing over their shoulder.

Juniors get those rules in five minutes of training. AI gets them in three lines of system prompt.

You wouldn’t skip it for a person. You skipped it for the AI because the AI showed up without an HR file. Without a first-day orientation. Without a manager pulling it aside and saying “here are the things you should never do here.”

That orientation is the system prompt. You write it. The AI reads it every time it runs. It’s not a meeting you have once and forget — it’s a meeting you have every single time the AI runs. Which is why it actually works.

Now — to sharpen. Guardrails aren’t right for every workflow. They’re right for customer-facing work, regulated work, anything where a wrong move costs money or reputation. They’re overkill for internal drafting, brainstorming, or any workflow where the AI is producing options for you to pick from. There the discipline is the opposite — let it roam, generate, surprise you. The cage that makes proposals generic is the same cage that makes brainstorming worthless. Different jobs, different posture.

But anywhere the AI is talking to a customer, signing off in your name, or making a commitment the company has to honor — guardrails first. Always. Before anything else.

So. Pick one workflow this week where you’ve put AI in front of customers. Renewal emails. Appointment reminders. Service follow-ups. Intake forms. Anything where the AI is communicating for you.

Ask one question. “What are the three things this tool should absolutely never do?”

Three. Not ten. Three. The ones where if it crossed the line, you’d be on the phone for hours apologizing.

Write them as absolute statements. “You will never quote a price.” “You will never promise a delivery date.” “You will never reschedule without explicit customer confirmation.” Use the word never every time. Don’t hedge. Don’t write “be careful about.” Don’t write “try to avoid.” Don’t write “in most cases.” Those are wishes, not walls.

Open the system prompt. Add a section called Hard Stops. Paste your three statements. Save.

Now test it. Try to make the AI break a rule. Ask it to quote a price. Ask it to make a promise. If the guardrails hold — you’ve fixed the door. If they don’t — your statement wasn’t absolute enough. Sharpen the language. Test again.

Twenty minutes of work. Costs nothing. Saves the next angry phone call. Saves Marlene’s wedding tickets.

Here’s the thing. The office manager didn’t need to stop using AI for customer work. She needed three sentences. Three nevers that would have closed the door before the email ever went out.

You already know how to write rules. You’ve been writing them for employees for years. Onboarding documents. Standard operating procedures. Here’s what we do, here’s what we don’t do. The AI just needs the same thing — written down, in the system prompt, where it can see it every time it runs.

So. Close the door. Then let it work.

Framework: The Professional Recipe — Ingredient #3 (Guardrails). Related failure mode: The Open Door Problem (#7).

Companion pieces: Stop Hiring AI. Start Building It. — the parent principle. The Process Cage — the sister piece on the other side of the framework: cages over-constrain, open doors under-constrain. The discipline is knowing which side of the line you’re on.