When AI Speaks for the Company: The Voice Problem

When AI speaks for the company, every sentence it produces is a claim the company has authored, whether or not anyone in the company read it first. A chat agent that reassures a customer, a generated email that explains a delay, a summary that describes an employee in a talent review: each one carries the organization’s name, and each one can be relied on. That is the voice problem, and most organizations have assigned it to the wrong function.

The usual home for “voice” is marketing, where it means tone, style, and brand consistency. That framing made sense when every public sentence passed through a person who could be asked what they meant. It does not hold when language is generated in many places at once, at a speed no review queue can match. What follows is a governance reading of the problem, written for boards, general counsel, and the executives who will be asked to answer for sentences they never saw.

What is the voice problem when AI speaks for the company?

For most of corporate history, organizational speech had an author. A customer-service representative could improvise, and sometimes improvised badly, but the improvisation had a face, a supervisor, and a training record. If a representative promised something the policy did not allow, the company could trace the promise, decide whether to honor it, and correct the behavior.

Generated language changes three things at once. It removes the fingerprint, because a fluent sentence no longer reveals who shaped it. It multiplies the volume, because one configuration can speak to thousands of customers in an afternoon. And it raises the confidence of the speech, because generated text arrives clean, consistent, and well formatted, which is exactly how authoritative statements look.

The result is a company that speaks constantly without deciding, sentence by sentence, what it is willing to say. The voice problem is not that the machine writes badly. It often writes better than the people it replaced. The problem is that nobody has decided who may speak in the company’s name, about what, and with what human ownership before the language travels.

Why does a chatbot’s wording become a company commitment?

Because commitment is created by reliance, not by intent. A sentence becomes the company’s word at the moment a customer, employee, supplier, or regulator acts on it. The company did not need to intend the promise. It only needed to put the sentence where someone would reasonably believe it.

Courts have started to say so. In a 2024 decision, British Columbia’s Civil Resolution Tribunal held an airline responsible for refund guidance its website chatbot had given a grieving customer, and rejected the argument that the chatbot was a separate entity responsible for its own statements (McCarthy Tétrault’s analysis of the decision). The legal reasoning was ordinary. The company published the information; the customer relied on it; the company was accountable for it. What was new was only the speaker.

That ordinariness is the point boards should take away. No new doctrine is required to make generated speech binding. The existing logic of representation and reliance already does the work. The governance gap is internal: most organizations still treat generated customer language as a product feature, reviewed for tone and accuracy at launch, rather than as a continuous stream of claims made on the company’s behalf.

What does voice drift look like in practice?

Consider a composite case. A specialty retailer deploys a chat agent to handle order questions and returns. The written return policy is clear and has not changed. The agent is tuned for empathy, because early customer feedback said it sounded cold. Over several months, its consolation phrasing loosens. “We understand this is frustrating” becomes “we’ll make sure this gets resolved for you,” which becomes “don’t worry, we can take that back,” offered to customers whose purchases fall outside the return window.

Nobody wrote the new promise. The policy team did not approve it. The vendor did not intend it. The customer-experience lead saw rising satisfaction scores and took them as good news. Store managers began receiving returns the policy did not cover, from customers holding screenshots of what the company had told them. At that point the question was no longer what the policy said. It was what the company had said, to whom, and how many times.

This is voice drift: the gradual widening of what the company effectively promises, produced by a configuration optimized for something else. It rarely shows up in an error log, because each individual answer looks reasonable. It shows up in the gap between the policy and the record of what customers were told, and that gap usually belongs to nobody until it becomes expensive.

Who should own the company’s voice when AI writes it?

Ownership has to move from the content to the authority to speak. The useful question is not “who reviews the chatbot’s answers?” No one can review every answer. The useful question is “who decides what the company is permitted to say, in which channels, about which subjects, and who answers when a generated statement is challenged?”

That is a decision-rights question applied to organizational speech. Most organizations already have mature decision rights for capital, contracts, supplier changes, product release, and regulatory communication. Very few have written down who may commit the company in language, or which subjects are too consequential for generated speech to address without a human author. A refund policy, a safety instruction, a warranty term, a statement about an employee’s readiness for promotion: each of these is a commitment, and each deserves a named owner before the language goes live.

In practice, voice governance rests on three working rules:

1. Name the commitment zones. Identify the subjects where a generated sentence can create an obligation: refunds, eligibility, safety, pricing, employment status, regulatory matters. In those zones, generated language may draft and prepare, but a human owner is accountable for the standard the language must meet.
2. Audit the record, not the configuration. Sample what the system actually said to customers and employees, and compare it with what the policy permits. Drift lives in the record. A prompt review at launch will not find it six months later.
3. Bind every channel to an answer. The glossary’s definition of Voice Rights is precise: the auditable right of a person working alongside an automated decision process to contest its output and be answered by an accountable human within a defined time. A channel without an obligation to answer is not a voice right.

The same discipline applies inside the company. A generated phrase in a talent packet or performance summary is the company speaking about a person, and it will travel into calibration rooms and succession decisions. As one executive in a recent book puts it, “We need to decide whether we own that sentence.” That decision cannot be delegated to the system that wrote it.

Why is voice governance a board-level and legal concern?

Because the exposure is aggregated, and the aggregation is invisible to the functions that usually manage it. Marketing sees tone. Customer experience sees satisfaction. Legal sees individual complaints. Finance sees an uptick in returns or credits. No one sees the pattern: a company making thousands of small claims a day that no accountable person has authorized.

Boards are asked to oversee risk that crosses functions and compounds quietly. Generated speech fits that description precisely. The questions for a board or audit committee are short. Where does the company speak through AI today, to customers, employees, suppliers, and regulators? Which of those channels can create a commitment? Who owns each one? And when did someone last compare what the system said with what the company intended to say?

The answers are often uncomfortable, not because the organization has been careless, but because voice was never framed as a governance asset. It was framed as a brand asset. Reframing it is the work, and it is easier to do before the first dispute than after.

Frequently asked questions

Is a company legally responsible for what its AI chatbot says? In many jurisdictions, likely yes. A 2024 British Columbia tribunal decision held a company responsible for inaccurate guidance from its website chatbot, applying ordinary misrepresentation principles. Information published on the company’s channels is treated as the company’s information. Organizations should assume customer reliance on generated statements can create obligations and govern accordingly.

What is voice governance for AI? Voice governance defines who or what may speak in the company’s name, to whom, about which subjects, and with what human ownership before language reaches its audience. It treats generated customer and employee communications as commitments to be governed, not content to be styled, and pairs every consequential channel with a way to contest and be answered.

How can companies prevent AI from making promises they did not authorize? Identify the subjects where a sentence can create an obligation, such as refunds, eligibility, safety, and pricing, and require human-owned standards for each. Then audit what the system actually said, not just how it was configured. Drift shows up in the conversation record over months, rarely in the original prompt or launch review.

Should the board oversee customer-facing AI? Yes, at the level of exposure rather than operations. The board does not need to read transcripts. It needs a map of where the company speaks through AI, which channels can create commitments, who owns each one, and how often the record is compared with policy. That is standard risk oversight applied to a new kind of speaker.

---

Boards and executive teams mapping where their company now speaks through AI can start a conversation with Anker Bioss about governing judgment and accountability in AI-shaped work. Related reading: The Point of Reliance and The Incomplete Org Chart.

Jose J. Ruiz is CEO and Managing Partner of Alder Koten and Chairman of Anker Bioss. He develops these ideas at book length in AI in the Org Chart: A Leadership Guide to Implementing AI Without Losing Human Judgment, Accountability, and Trust, published by Elavant Press and available on Amazon.