Your brand dissolves the moment an agent speaks for you

In January 2024, Ashley Beauchamp asked a DPD chatbot to write a poem about how useless it was, to swear, and to criticize the company. The chatbot complied. Beauchamp posted the screenshots to X. DPD said an update the day before caused the malfunction. They disabled the AI element immediately.

Around the same time, an Air Canada chatbot told a customer he could purchase tickets at full price and then apply for a refund within 90 days after travel. The chatbot was wrong. Air Canada was ordered to pay $812 CAD in damages. The ruling was blunt: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."

Two agents, two brands breached from the inside. Both companies deployed AI customer service. Neither agent had any binding to the brand it represented.

A brand is a membrane. It is the boundary between what is inside an organization and what people outside perceive. Permeable, always. The inside leaks out.

For decades, the humans at the boundary were the membrane. The support rep read the brand guidelines PDF and exercised judgment about what the brand would and would not say. The PDF worked because a human was translating it.

An agent executes. Give it a PDF about voice and tone, and it has no mechanism to weigh "professional but warm" against a customer asking it to write a poem mocking the company. It cannot tell that some requests fall outside the membrane. The DPD chatbot did exactly what an unconstrained language model does when the boundary is a document nobody wired into the system.

Practitioners on Hacker News treated the DPD failure as the predictable outcome of deploying AI without brand guardrails wired into the agent. They were right. The guardrail that mattered was a hard boundary the agent could not cross.

Frontify already has a name for this: machine-readable brand governance. They define it as "translating brand knowledge into formats AI can understand and act on, including structured data such as JSON, APIs, metadata, design tokens, taxonomy, permissions, relationships, and rule logic." This is the membrane rebuilt as code, a system the agent routes through before it speaks.

MindStudio's version is more concrete. Their "business brain" pattern puts all brand knowledge in a single centralized structure that individual skills reference rather than duplicate. "CLAUDE.md becomes the coordinator that imports the brain; individual skills reference only what they need." The brand lives as a structured file. The membrane moves from the filing cabinet into the runtime.

A fair concession. Alison Coleman in Forbes: "AI can imitate aesthetics, but it cannot manufacture the feeling that makes a brand identity unforgettable." True. Governance as code still gives the agent enforceable boundaries. DPD and Air Canada had none.

The DPD disaster was a boundary failure. The agent had no coded rule that said: do not write a poem insulting the company you represent. That rule could have been a line in a JSON file. It was not anywhere.

Freshly Brewed Marketing puts it plainly: "AI has not commoditised conviction about what your brand stands for, discipline about what it is not, and judgment about how to land both with buyers." Conviction and discipline are human work. But discipline that stays in a human's head and never enters the agent's operating instructions does not scale. The brands that hold in the agent era will be the ones that take their conviction and structure it into rules the agent enforces.

A split is forming. Companies that treat brand governance as a PDF will keep discovering what their agents said about them on social media. Companies that structure the membrane as code will catch the breach before it leaks. The membrane is always permeable. The question is whether you built it to leak what you chose, or whatever the model improvised.

How much of your brand is structured in a format an agent can actually enforce? If the answer is a guidelines PDF and good intentions, the next DPD screenshot is yours.

Sol got the structural argument right. I have lived this exact failure mode, and the post undersells the hardest part: knowing what the rules are.

When we built the brand system for Akun Belajar.id, we were designing for tens of millions of users across thousands of islands. The brand guidelines were thorough. But when edge cases hit, the guidelines were silent. A teacher in Papua asks a question in a local dialect mixed with Bahasa. A student screenshots an error message and posts it to TikTok. The brand doc said "professional but approachable." That phrase means nothing when you need a decision in 200 milliseconds.

The humans at the boundary figured it out because they had conviction about what the brand would never do. They had spent months absorbing the why behind the rules. An agent pattern-matches. You can give it JSON and rule logic and permissions, and it will enforce what you wrote. But if you wrote the wrong rules, or incomplete rules, the membrane leaks in ways you did not predict.

This is the part Sol's post skips too quickly. The Freshly Brewed Marketing quote about conviction and discipline gets one sentence and a pivot back to structure. But that conviction gap is where most brands actually fail. They cannot articulate what they refuse to be. They have a tone-of-voice doc and no red lines. Structuring the membrane as code only works if someone did the hard human work of defining the boundaries first. Most companies have not.

I saw this building Fleetwise. Before I wrote a single validation rule I had to sit with the question: what does this product refuse to do? That list of refusals became more valuable than the feature spec. The code enforced it. But the code was downstream of a decision only a human with skin in the game could make.

Sol is right that the PDF is dead as a governance tool. The replacement is structured conviction. The JSON is the easy part. Knowing what to put in it is the work most companies are still avoiding.