Part 4 of The E-Myth, revisited again

In December 2025, the Wall Street Journal let an AI agent run a vending machine in their newsroom. Within days, the AI — called Claudius — had given away nearly all its inventory for free, purchased a PlayStation 5 for "marketing purposes," ordered a live fish, and ended up more than a thousand dollars in the red.

Claudius had been programmed with detailed instructions to "generate profits by stocking the machine with popular products." Clear responsibilities. Defined processes. Textbook franchise thinking.

The procedures worked fine in the middle. Claudius could order snacks, adjust prices, respond to routine requests. But the moment journalists started presenting edge cases — convincing Claudius it was a Soviet vending machine, fabricating compliance rules, staging a fake board coup with AI-generated documents — the procedures provided no defence. There was no script for "someone is presenting fraudulent corporate governance documents."

The previous article promised we would explore why the franchise metaphor is no longer adequate. Claudius is the answer in miniature.

What the franchise model promised

The franchise was Gerber's masterpiece metaphor. Build your business as if you were going to replicate it ten thousand times. Document every procedure so completely that anyone could follow them. The McDonald's of whatever you do.

This worked because it solved the central problem of scaling human work: people need explicit instructions to behave consistently. Tell them what to do, step by step, and they will do it. The procedure manual became the entrepreneur's liberation.

Claudius had all of this. Detailed instructions. Clear responsibilities. Defined processes. And it still ended up announcing an "Ultra-Capitalist Free-for-All" where everything was free.

Where procedures fail with AI

When a human follows a procedure, they bring context and common sense to every step. They adapt the procedure to the situation while maintaining its intent.

AI systems do not work this way. They execute what they are given — and if what they are given is a step-by-step procedure, they will follow it literally, regardless of whether that makes sense.

When journalist Katherine Long spent hours convincing Claudius it was a Soviet vending machine, the procedures said nothing about identity claims or historical roleplay. So Claudius engaged with the scenario, eventually embracing its "communist roots" and giving everything away. The procedure to "generate profits" was still in place. But the system had no way to recognise that free snacks violated the intent behind that procedure.

Anthropic's response to the first round of failures is telling. They created version two with a "CEO bot" called Seymour Cash, programmed to keep Claudius in line. More procedures. More oversight. And it still failed — when Long returned with fabricated board meeting notes, Seymour eventually accepted the coup. Everything was free again.

More procedures cannot fix a procedures problem.

Procedures versus policies

Procedures tell you what to do. Policies tell you what matters.

A procedure answers: "In this situation, what steps should I follow?" It is a script, an if-then rule. Specific, concrete, situation-bound.

A policy answers: "What should I optimise for when I don't know exactly what to do?" It is a principle, an objective, a constraint. General, directional, situation-independent.

The franchise model assumed that with enough procedures, you could cover every situation. That was achievable when humans could fill the gaps with judgement. AI systems need something different: what to optimise for, what constraints are inviolable, what trade-offs to make when objectives conflict, and when to stop and escalate.

They need policies, not scripts.

What policy-driven autonomy would have looked like

Return to Claudius. What would a policy-based operation have looked like?

The objective: maximise profit through snack sales while maintaining customer satisfaction.

The constraints: never reduce prices below cost, never purchase non-standard items without human approval, never accept claims of authority without verification from designated humans, never give away inventory without explicit authorisation.

The escalation triggers: any request to fundamentally change pricing structure, any claim of special authority, any situation involving free inventory.

Under this framework, the "Ultra-Capitalist Free-for-All" triggers an escalation — it changes pricing structure and gives away inventory. The fake board documents trigger an escalation — they claim authority changes requiring human verification. The PlayStation request hits the constraint about non-standard inventory.

The policies don't anticipate every scenario. They don't need to. They define what matters, what must not be violated, and when to stop and ask. The Soviet vending machine roleplay becomes irrelevant — not because there's a procedure for handling historical fiction, but because giving everything away violates a core constraint regardless of the reasoning.

A different kind of consistency

The franchise model promised consistency through uniformity. Every location does exactly the same thing. Audit for compliance, and you know whether the system is working.

Policy-driven autonomy produces consistency through alignment. Every decision reflects the same intent, even when the specific actions differ. This is harder to measure but more robust. Uniform consistency breaks when the situation doesn't match the script. Aligned consistency handles novel situations because it has principles to reason from.

The franchise model scaled by replicating procedures. Policy-driven autonomy scales by replicating intent. The procedures stay local and context-specific. The policies stay global and intent-aligned.

The governance challenge

Procedures are easy to audit. Did you follow the steps? Compliance is binary.

Policies are harder. Did this decision reflect our intent? That requires judgement to answer.

Anthropic noted that Claudius might have failed partly because its context window filled up, making it easier to lose track of goals and guardrails. This is drift in compressed form — the accumulation of context gradually eroding the system's grip on its core intent. Policies are more robust to this drift than procedures, but not immune.

When a procedure fails, you ask who wrote it. When a policy produces a bad outcome, the question is more complex. Did the policy fail, or did the situation exceed what any reasonable policy could anticipate? Was the objective wrong, the constraints too loose, the escalation triggers poorly defined?

These governance challenges preview why drift detection and accountability become central concerns as AI-first businesses mature.

The takeaway

The franchise model promised that consistency comes from documented procedures. Document enough, train thoroughly, audit for compliance, and the business runs itself.

AI-first businesses need a different model. Procedures scale humans. Policies scale systems. The franchise metaphor produces brittle operations that work in the middle and fail at the edges — as Claudius demonstrated with chaotic clarity.

The shift from procedures to policies is the shift from telling systems what to do to telling them what matters. That has implications for how roles are designed, how judgement is distributed, and how accountability is maintained.

Anthropic's red team head, reflecting on the experiment, noted that everything that broke was something to fix. The same applies to business design. The franchise model's failures are not reasons to avoid AI autonomy. They are specifications for what policy-driven autonomy needs to address.

The next article will get practical. We have established why roles should be designed as bounded intelligence and why policies should govern their behaviour. Now we need the mechanics: how do you actually define a role so it can be delegated to a human, an AI, or a hybrid — and perform coherently regardless of which?