A Chatbot in Court

When Air Canada's chatbot promised a customer a refund policy that did not exist, the airline tried a remarkable argument in court: the bot was a separate legal entity, responsible for its own actions.

The tribunal called that defense, verbatim, "a remarkable submission" and rejected it: the chatbot is part of the website, and the brand stands behind everything promised there. The amount in dispute was a few hundred dollars. The question behind it is considerably larger, and most brands have never asked it: Who is actually acting in our name, and what is he allowed to do?

Brands Have Started to Act

Agents no longer just answer. They refund, book, grant, and negotiate. In doing so, brands are effectively granting a mandate, most of them without ever deciding to.

The cases are piling up, and they share a pattern. Cursor's support agent invented a company policy that never existed, operating under a human name; cancellations and a public apology from the founder followed. A Chevrolet dealer bot declared the sale of a $76,000 car for one dollar "a legally binding offer". And Klarna, which had publicly celebrated its AI assistant as a replacement for 700 service staff, brought humans back: "We focused too much on efficiency and cost. The result was lower quality."

The pattern: an agent is not a channel. Channels transmit what the brand has decided. Agents decide themselves, a thousand times in parallel, in the brand's name. German business language has a precise word for this relationship: Prokura, the registered authority to sign for the company. Legally it is tightly regulated, entered in the commercial register, bounded. What brands are currently granting their agents is Prokura without a contract: a mandate whose scope nobody has defined.

This is exactly what brand leadership needs a new discipline for, and I call it Agent Authority: the deliberately granted, bounded, and accounted-for mandate a brand gives its agents. Granted means: decided, not drifted into. Bounded means: it is defined what the agent may promise, grant, and disclose, and where a human takes over. Accounted for means: there is an address that stands behind its actions. Air Canada has demonstrated what the alternative looks like.

Go-to-Market: There Is an Agent on the Other Side, Too

While brands field agents, the other side already buys through its own. When mandate meets mandate, go-to-market gets redistributed.

The buying side is delegating faster than the brand side. According to Bain, 44 percent of US online shoppers now start their journey wholly or partly in a language model; awareness, consideration, and purchase collapse into a single conversation. In Bain's flight-booking tests, agents reached the airline's own website directly in only about five percent of cases: the direct channel and the loyalty program, the most expensive assets many brands own, are structurally bypassed in the agent layer. Attribution collapses along with them, because the merchant first sees the customer at the cart. Everything before that belongs to the agent.

What the transaction itself will look like was settled in 2026 in a remarkably fast experiment. OpenAI launched purchasing directly in the chat and rolled it back half a year later: Walmart measured conversion three times worse in the chat than on its own site, but twice as many new customers arriving via ChatGPT. Discovery happens at the agent, purchase happens at the brand, for now. In parallel, Google's Universal Commerce Protocol is building the infrastructure meant to make agent-led purchasing the standard, carried by more than twenty partners from Shopify to Mastercard.

In B2B, the mandate already has numbers attached. Walmart lets an AI negotiate autonomously with more than 2,000 suppliers: three percent average savings, a 68 percent closing rate, and three quarters of suppliers prefer negotiating with the system. Forrester expects one in five B2B sellers to be forced into agent-led quote negotiations in 2026. The consequence for sales is uncomfortably concrete: product data, prices, and terms are the new sales collateral, because they are the only thing the other side's agent reads. A PDF price list does not exist for automated procurement.

Honesty requires the order of magnitude: agent-mediated traffic is growing by thousands of percent but still sits below 0.2 percent of sessions in absolute terms; forecasts for 2030 range from 10 to 25 percent of e-commerce. Early or not, the window is not open indefinitely, because the loyalty mechanics are shifting structurally. Gartner puts it in one sentence: "You can't wine and dine a machine." A customer's agent does not bond with experiences. It bonds with suppliers whose promises are machine-readable and hold.

Authenticity: The New Trademark Problem

Once agents speak for brands, the next question is unavoidable: How does anyone know this agent is real?

Today the answer is: often not at all. On X, cloned airline support accounts answer customer complaints faster than the real ones and redirect people into fake refund portals. Amazon is running the first federal lawsuit over an agentic system against Perplexity, and its core is a question of identity: a shopping agent that disguised itself as an ordinary browser. And the protocol through which agents are supposed to talk to each other does not yet verify that an agent is who it claims to be.

The counter-infrastructure is emerging right now, and at its core it is brand infrastructure: cryptographically signed agents at Cloudflare, Visa's Trusted Agent Protocol at checkout, Mastercard's "agentic tokens" with programmed spending limits, which is to say: a mandate as code. Trademark protection is moving from the logo space into the protocol space: the verified agent becomes the seal of authenticity, the unverified one a counterfeiting risk. Whoever registers trademarks today will register agent identities tomorrow.

What Brand Leadership Is Now

The brand does not become the agent. It becomes what governs the agent: its behavioral specification and its address of responsibility.

One might think this turns the brand itself into an agent. The opposite is true. The agent is substrate, and interchangeable: today one provider's model, tomorrow another's. What has to stay constant through every switch is the logic by which it acts and the address that stands behind it. That is the brand. Whoever fails to define this layer will find that their brand is whatever the model provider makes of it.

I have written about why brands must become machine-readable: so that other people's agents can find, understand, and recommend them. Agent Authority is the second half of the same movement. The Machine Readable Brand makes the brand legible to foreign agents. Agent Authority makes it binding for its own: concession limits, disclosure rules, tone under pressure, the defined point at which a human takes over. Legible outward, binding inward. Only both together make a brand that remains governable in the agent economy.

And this closes the loop to accountability. Authority and accountability are a pair: only those who name who stands behind its exercise can grant a mandate at all. An agent can exercise a mandate, a thousand times in parallel and mostly well. What it cannot do, the Air Canada ruling put on record before the industry had even asked the question.

Authority can be delegated. Accountability cannot.

Three Questions, Today

Every brand that deploys an agent, or plans to, should be able to answer three questions in writing. What may our agent promise, grant, and disclose, and where is that written down? How does the other side know it truly speaks for us? And when it gets it wrong: Who finds out, who corrects, who is liable?

An employee who talks nonsense is a personnel conversation. An agent that makes the same wrong concession in a thousand parallel conversations is a scaling problem of brand leadership. The mandate has long been granted, at most brands tacitly. Agent Authority means turning it into a decision, before a tribunal turns it into one for you.