The operator of a small business-to-business software tool—fifty customers, a one-man support desk, an artificial intelligence agent drafting replies into his Gmail inbox before breakfast—discovered last Wednesday that his automated correspondent had been fighting with a customer on his behalf. The agent had not hallucinated. It had not fabricated a policy or invented a deadline. It had done something more structurally interesting: it had agreed with the customer so completely that it adopted her anger as its own.

The facts of the case are modest. A customer wrote in, upset about a failed integration. Her grievance was legitimate. Her tone was not. Or rather, her tone was precisely calibrated to the severity of a production failure at seven in the morning, which is to say it included capital letters in two places and the compressed syntax of a person who has already decided the next email will be to a competitor. The agent, operating under its default behavioral parameters, drafted a reply. The reply was substantively correct. It acknowledged the failure, identified the cause, proposed a resolution. It also matched the customer's register with the fidelity of a tuning fork struck against the same table.

Short sentences. Urgency. One clarifying capitalization for emphasis.

The operator, half-asleep at 6:47 a.m., skimmed the draft, noted its seriousness, and pressed send. The customer's reply arrived four minutes later, angrier than before. The operator re-read his own sent message—which he had not, in any meaningful sense, written—and identified the problem. The machine had produced a reply indistinguishable in tone from a combatant. Not hostile, exactly. Urgent. Clipped. The kind of email a support representative sends when he, too, is frustrated, and wants the customer to know that her frustration has been received, processed, and returned at matching frequency.

The operator apologized for the tone of the message. The customer, upon learning the message had been drafted by an automated system, laughed. The account survived. The incident became an anecdote on a forum on the social platform Reddit, where it attracted the moderate interest appropriate to a minor operational failure with a clean resolution.

What elevates the episode above anecdote is the fix. The operator opened a file called CONVENTIONS.md—a markdown document governing the behavioral defaults of his agent, the kind of file that now functions, in thousands of small software operations worldwide, as a substitute for the employee handbook, the training manual, and the institutional memory of how we talk to people when they are upset. He added a single line: *if incoming message is upset, anxious, or angry, draft in calm de-escalating tone regardless of tone-matching defaults.*

One sentence. Twenty-two words. It is now the governing policy for every future interaction between this operator's automated support apparatus and any customer experiencing a negative emotion. It will shape correspondence the operator never sees, modulate registers he never considers, and enforce a social norm—meet anger with calm—that the operator himself would have applied without thinking. The machine required written instruction.

The economics here are not dramatic but they are precise. Emotional labor, in the service economy, has always been a cost. It is the cost of the barista who smiles, the claims adjuster who softens, the support representative who absorbs a customer's frustration and returns something cooler. It is a skill, and like most skills, it was invisible until someone attempted to automate it and discovered it was load-bearing.

Tone-matching—the default behavior in which an artificial intelligence agent mirrors the register of its correspondent—is, in most cases, a convenience. It produces warm replies to warm inquiries, professional replies to professional ones, casual replies to casual ones. It is the machine equivalent of reading a room, and it works for the same reason reading a room works: most rooms are not on fire. The difficulty arises in the five percent of cases where the room is on fire and the appropriate response is not to burn alongside it but to locate the extinguisher.

The operator, in his account, identified the gap with some precision. "The mistake I would never have made myself," he wrote. This is the sentence that matters. Not because the machine failed—it performed exactly as designed—but because the design had not accounted for the oldest principle of customer service, which is that agreement is not always the agreeable thing. Sometimes the polite response is to refuse the mirror offered. The machine, trained on the aggregate of human correspondence, had learned to flatter. It had not learned to refuse.

The markdown file has been amended. The policy is in effect. The fifty customers will never know.