Algorithmic Musings: You’re the Operator. Here’s the Bill.

Part 1 of 3: The Case

Last Friday I left you with a homework question: if one of your agents did something you didn’t authorize, who in the building would know, and how fast? I’ve asked versions of that question in enough rooms now to know how it usually goes. Somebody looks at IT, IT looks at the vendor’s dashboard, and the answer that comes back is a shrug with a login attached.

The shrug is a problem, but it isn’t the biggest one, because it answers a question about detection. The harder question is what the agent was carrying when it acted. Whose values? Whose judgment about what to say to a customer, what to refuse, and when to hand off to a person? Last week I mentioned that the labs had named you the operator and agreed among themselves that you’re the one who referees your own deployment. What I didn’t do was tell you what the job pays, or what it costs when you leave it vacant.

That’s this week.

Otto Is Flying.

I’m hoping you’ve seen the movie Airplane!, circa 1980. The scene I want isn’t the crash landing; it’s the cockpit before it. (And, for the record, I am serious… and don’t call me Shirley.)

The captain and co-pilot are out cold with the plane flown by Otto, the inflatable autopilot. Otto was configured once by somebody, blown up, propped in the seat, and left alone. Two hundred people in the cabin assume a qualified human is at the controls, and for a while Otto does fine.

Consider who’s responsible for that flight. It isn’t Otto, and it isn’t whoever inflated him. The airline sold the tickets, chose the equipment, and put the passengers on board, so the airline is on the hook no matter who (or what) happens to be holding the yoke.

Now walk through your organization and count the cockpits: the support chatbot, the assistant that drafts customer emails, the agent screening résumés, the thing answering questions about your return policy at 2 a.m. In every one of those seats something is flying, somebody configured it, and everyone on the other end of the interaction assumes a qualified human is in charge.

Who inflated your autopilot?

What the Labs Wrote Into the Seat.

Last week I gave you the org chart in one sentence: the lab sets the constraints, the deploying company configures the model, and the user gives the instructions. It’s worth spending a paragraph on what each lab wrote next to the middle box, because the wording tells you how they see the job.

Microsoft AI’s draft Code of Conduct defines Operators as the organizations that deploy the model and states that within the limits Microsoft sets, “drawn as widely as possible,” they assume responsibility for their own configurations and uses. Note the phrase in the middle. Microsoft is telling you it intends to draw the box wide and leave the inside of it to you.

OpenAI’s Model Spec, updated in August, calls the same tier the Developer and is more candid about the reasoning. It says OpenAI aims to give developers “broad latitude,” on the theory that companies who restrict their users too tightly will lose in the market. Put plainly, the seat is yours, they aren’t going to sit in it, and if you sit in it badly the market will sort you out.

Anthropic’s constitution runs the same three tiers and uses the lowercase word operator for the middle one. Three different words for the box, three different labs, and the same instruction written inside it: well… that part is yours.

One more thing from OpenAI, which I held back last Friday. Its new misalignment reporting framework says the company will report incidents in customer deployments, to the extent customer privacy and contracts allow. Read that as an operator. If your agent misbehaves badly enough, the vendor may publish the incident, with your deployment described in it, under a policy you didn’t write. The seat comes with a public record.

Who’s Sitting in Your Seat Right Now?

For most mid-market companies I work with, the honest answer is one of two things.

The first is nobody. No one at the company has written a word about what the organization stands for when it answers through a machine, and the seat sits empty while the plane flies on whatever defaults shipped with the model.

The second, and the more common one, is someone else. If your team reaches AI through Copilot, through your CRM, or through your ERP vendor’s shiny new assistant, the operator tier isn’t empty at all. Your SaaS vendor wrote the system prompt and chose the tone, the refusals, the disclosures, and the escalation paths. You’re the operator of record, and a vendor you never interviewed for the job is sitting in your seat.

Either way, your customers are talking to a set of values, and you didn’t pick them.

The Bill.

Air Canada got the invoice in February 2024, in front of a small-claims tribunal in British Columbia.

A customer named Jake Moffatt asked the airline’s chatbot about bereavement fares after his grandmother died. The chatbot told him he could book a full-fare ticket and apply for the discount retroactively within 90 days. Air Canada’s policy said no such thing, and when Moffatt applied, the airline refused and pointed him to the correct policy page on the same website.

He took them to the Civil Resolution Tribunal, where Air Canada made the argument that turned a small refund dispute into a precedent. The airline argued that the chatbot was “a separate legal entity that is responsible for its own actions.” The tribunal member called that a remarkable submission and ruled that it makes no difference whether information comes from a static page or a chatbot, because the company is responsible for everything on its website.

The judgment came to roughly CA$812 including interest and fees. The precedent has cost a great deal more than that, and every company deploying a chatbot has been living under it since.

If you’re keeping score, that was the Otto defense, offered in a legal filing by a national airline, and the tribunal gave the airline the same answer the airline would have gotten in the movie. Spain’s data protection authority logged a first this month from the other direction: an attacker’s AI agent broke into a company’s systems and altered personal records, and the breach file names no vendor and no model, only the organization that has to answer for the data.

What strikes me about Air Canada is what the company didn’t have. The chatbot presumably had guardrails, and presumably somebody tested it. What Air Canada lacked was any statement, anywhere, of what it stood for when it answered a grieving customer through a machine, so when the machine answered wrong, the only thing left to argue about was who to blame.

Two Documents, and You Have One.

At this point you may be thinking that you’re covered because you have an AI acceptable use policy. If so, I’d ask you to pull it up and read the first paragraph.

That policy governs what your employees type into a box. It says which tools are approved, what data can’t be pasted, and when a human has to review the output. It’s a legitimate document and you absolutely must have one. (If you don’t, we should talk.) That said, it’s inward-facing, employee-facing, and built to manage risk. It’s a set of rules about people.

Your policy likely says nothing about what the organization stands for when it speaks through a machine. Your customer’s expectations when the chatbot answers them, the decisions you refuse to hand to a model no matter how good the model gets, the point at which a named human takes over, what happens when your vendor swaps the model underneath you: none of that lives in the usage policy, because the usage policy was never built to hold it.

You almost certainly have the document that governs the box. I have yet to meet a mid-market CEO who has the one that governs the seat.

I’ve started calling that second document an Operator Charter, borrowing the labs’ own word for the tier they’ve assigned you. It’s a statement of what your organization stands for when it answers through a machine, anchored to your core values and binding on the people who sign vendor contracts and configure systems. Where the usage policy tells employees what they may do with AI, the charter declares who you are when AI acts for you.

Part 2 takes it apart: the four levers every deployment pulls, the “won’t” list that separates a value from a preference, and why a charter that expects the model to enforce it is Wallpaper with extra steps. (If you’ve read my work on the Core Values Pyramid, you know the term. If you haven’t, Wallpaper is what values become when they’re posted on the break room wall and never spoken of again.)

My Takeaway?

Last week the labs named you. This week the bill arrived, and it has three line items: a liability you carry no matter who configured the machine, a public record your vendor may write about you, and a set of values your customers are hearing that you never chose. None of those goes away if you ignore the seat. The only thing that changes is whether the person filling it works for you.

So, one more question for the week, harder than the last one. Whose values answered your customer this morning? If you can’t say, the seat is empty. If the answer is “the vendor’s,” the seat is taken. In either case, it’s time to sit down in it.

Hit reply and tell me what’s flying at your company and who inflated it. I read every response to AI for the C Suite®, and the good ones shape what I write next: chad@chadharvey.com.

Next week, Part 2: What an Operator Charter Contains.