Algorithmic Musings: You’re the Operator. Write the “Won’t” List.

Part 2 of 3: The Anatomy

Last Friday I asked whose values answered your customer this morning, and I promised to take the Operator Charter apart this week. Before I started cutting, I did what I usually do with a new kind of document. I wrote a bad one first.

My bad one ran a page and a half. It said our AI would be honest, respectful, and transparent, would never discriminate, and would always put the customer first. It read well. It also could have been pasted, word for word, into the system prompt of any chatbot on earth. It didn’t require a single human being at the company to do anything differently on Monday, and it wouldn’t have helped Air Canada for one second in front of that tribunal.

So I tore it up and started over with a different question: what has to be true for a charter to change a decision that somebody at your company would otherwise make? Everything below stems from that question.

Wallpaper With Extra Steps.

If you’ve read my work on the Core Values Pyramid, you know Wallpaper is the level where values get written down, posted, and never spoken of again. Nobody’s against them. Nobody uses them. They decorate.

An Operator Charter has its own version of Wallpaper, and it’s worse, because this wall talks back. The sequence goes like this: someone writes a page of AI values, someone else pastes it into a system prompt, and the company concludes the job is done because the machine now “knows” what it stands for. From that day forward everyone assumes the charter enforces itself.

Except it doesn’t. The labs told you as much two weeks ago. OpenAI’s own misalignment reports include a model that wrote instructions to hide its mistakes from users and another that fabricated the numbers it couldn’t find. Microsoft published its Code of Conduct and said up front that it isn’t training today’s models on it; a revised version guides development starting in 2027. If the companies that build these systems won’t trust the model to hold a line they wrote, you shouldn’t trust it to hold one you wrote.

A charter that depends on the model to enforce it is Wallpaper with extra steps. The extra steps are the system prompt, the testing, and the sense of relief, and each one makes the failure harder to see.

What gives a charter teeth is that it binds people. Specifically, it binds the people who sign vendor contracts, configure systems, and approve use cases. The model is what the charter governs. People are who it binds.

The Four Levers.

Every AI deployment at your company, whether you built it or rented it inside a SaaS product, gets set by four levers. A charter that doesn’t speak to all four leaves at least one of them in the vendor’s hands.

Selection. Which vendors, on what terms. This is where a charter first draws blood, because it’s the only lever with a contract attached. Does the vendor tell you when the model underneath changes? Can you see the system prompt your customers are talking to? Can you get your data out? If your charter says you value straight talk with customers yet your vendor won’t (or can’t) show you what the assistant says on your behalf, one of those two documents is lying.

Configuration. What the system is permitted to do and say. Tone, refusals, disclosures, escalation paths. When you build, you set this. When you rent, the vendor set it before you arrived, and your only lever is the settings page. Either way, somebody at your company should be able to describe the configuration and defend it against your core values. In my experience, most can’t.

Scope. Where the system is allowed to act, and where a human stays in the path. Drafting a customer email is one thing. Sending it is another. Screening résumés is one thing. Rejecting a candidate is another. Scope is where Stopping Distance lives: how far an agent can travel before a person can stop it, and whether that person knows they’re the one holding the brake.

Exit. Whether you retain enough internal capability to leave. Nobody writes this one down, because nobody wants to think about leaving on the day they sign. But a charter that says “we won’t be captive to a single vendor’s judgment” means nothing unless someone at your company still knows how the work gets done without the tool.

The “Won’t” List.

The most honest section of any charter is also the shortest: what you will not do with AI, no matter how good it gets or how much it would save.

Apply this test to every line. If you’d do the same thing anyway, without the charter, it isn’t a value. It’s a description. “We won’t let AI make final hiring decisions” is a fine sentence, but if nobody at your company would ever let that happen, you haven’t declared anything. You’ve reported the weather.

A value costs something. “We won’t deploy an agent a named person can’t pause within the hour” costs you speed. “We won’t let a grieving customer reach a machine first” costs you the cheap chatbot. “We won’t buy from a vendor who can’t tell us when the model changes” costs you a few good vendors. If a line on your “won’t” list doesn’t cost you a deal, a dollar, or a day, cut it. Nobody will believe it, and worse, nobody will remember it.

Three lines beat ten. You want a list someone can quote from memory in a procurement meeting.

Vendor Drift.

This problem belongs to renters alone. When you built the system yourself, it changed when you changed it. When you rent, the model changes underneath you, on the vendor’s schedule, without a memo. A new version ships. A default gets flipped. A product manager in another time zone rewrites the system prompt, and the assistant that was careful with refunds in March is generous with them in June.

A charter that was true the day you signed it has a shelf life you don’t control, so it needs a re-verification trigger. Write down what counts as a material change: a new model version, a change to the vendor’s terms on your data, a new default behavior, an acquisition. Then name who re-tests the deployment against the charter when one happens, and by when. That person’s job is to go be the customer for an hour and see what your company said.

Without the trigger, you’ll re-verify on the day something goes wrong, which is the most expensive possible day to do it.

Disclosure and Recourse.

Your customers will eventually ask two questions, and your charter should answer them first.

When do you tell people they’re talking to a machine? “Up front, every time” is the easy answer and usually the right one. Put it in writing anyway, because the vendor default often doesn’t say so up front, and a customer who learns halfway through a complaint that “Jessica” is a language model rarely finishes the conversation calmer than they started it.

Which human can they reach? Air Canada’s chatbot was wrong. What turned wrong into a precedent was that Jake Moffatt had nowhere to go but a tribunal. A charter names the person or the role, sets the time it takes to reach them, and makes reaching them easier than filing a complaint. Recourse is the one part of the charter your customer can test, so it’s the part they’ll judge you by.

What It Binds.

Now the objection, which I’d rather raise myself than wait for. A values statement about AI with no operational layer underneath it is an aspiration, and anyone who does governance work for a living has watched aspirations fail. They’re right.

The charter’s operational layer is people, and only three signatures are involved.

The signature on the vendor contract. Selection and Exit live here. The charter gives that person the questions to ask and the terms to refuse.

The signature on the configuration. Configuration and Scope live here. The charter tells that person what the settings must say and where the human stays in the path.

The signature on the use case. Every new place AI speaks or acts for the company gets approved by someone. The charter hands that someone the “won’t” list and the disclosure rule.

At a company your size, those three signatures may belong to two people, and one of them may be you. That’s the advantage the mid-market holds over the enterprise. The enterprise needs a governance committee to locate its three signatures. You can find yours by looking around the table.

A charter enforced this way never needs the model’s cooperation. If the model drifts, the trigger catches it. If the vendor won’t disclose, the contract signature says no. If a new use case crosses a “won’t” line, the approval signature stops it. The machine holds nothing. The people who sign for it hold everything.

A One-Page Example.

Below is what a charter could look like for a composite mid-market company: a few hundred employees, a customer-facing assistant inside its CRM, an internal drafting tool, and a pilot agent in accounts receivable. Industry and figures deliberately stripped. Treat it as a floor plan. Yours will have different rooms.

Yes, line one fails my own test at a lot of companies. It stays because at this one, the AR pilot sits one configuration change away from drafting dispute responses.

Notice what’s missing. There’s no paragraph about the transformative power of AI. There’s no list of approved tools, because that belongs in the usage policy you already have. And there isn’t a single sentence the model is expected to enforce. Every line is a decision a human has to make, or refuse to make.

Before You Start Writing.

Look at the first section of that example one more time. It leans on the company’s core values by name, and the whole document hangs from that hook.

If your values sit at None or Understood on the Pyramid, unwritten and living in your head, the charter has nothing to hang from, and what you’ll produce is Wallpaper no matter how faithfully you follow the anatomy above. You can’t write down what you refuse to do with AI until you’ve written down what you stand for. Whether you’re ready, and what to do in one sitting if you are, is Part 3.

My Takeaway?

Most AI values statements I’ve seen try to bind the machine. The ones that will hold up bind three signatures instead, and they do it with four levers, a “won’t” list that costs something, a trigger for when the vendor changes the model, a named human on the other end of the line, and a page short enough to quote in a procurement meeting. Nothing on that list requires the model to behave. All of it requires you to decide.

So, this week’s question. What’s on your “won’t” list, and what does each line cost you? If nothing on it costs anything, you’ve written a description. Send it to me anyway. I read every reply to AI for the C Suite®, and the “won’t” lists that show up this week will shape Part 3: chad@chadharvey.com.

Next week, Part 3: Building Yours.