There is a document in most organisations that was written by a small group of thoughtful people who genuinely cared about language. It describes the brand’s personality in four to six adjectives. It shows examples of good sentences and bad sentences. It has a section on exclamation marks. Someone printed it out and put it in a shared folder. It is, for all practical purposes, decorative.
This is the brand voice document. And in an age when AI is generating content at the scale of a medium-sized newspaper every hour, it is no longer adequate as the main instrument of communication governance. The issue is that they were solving a 2015 problem with a 2015 tool. The world has moved, and the document has not.
A tone of voice document is the most instituted object in brand communication. It crystallises a set of decisions made at a particular moment and offers them, unchanging, to a world in permanent motion.
The philosopher Cornelius Castoriadis, whose work on social imaginaries is still underappreciated outside France, made a distinction between the “instituted” and the “instituting”. The instituted is the form things have settled into: the rules, the roles, the conventions. The instituting is the living force that keeps creating and re-creating meaning. Organisations freeze into the instituted and forget the instituting energy that gave the institution its original coherence.
When your workforce was ten people writing copy by hand, the document was enough. The people interpreting it brought their own judgment, their own ear for the brand, their own sense of when a sentence was right. The document was scaffolding for that human judgment. It never needed to be more than scaffolding because human judgment was always doing the real work.
Generative AI removes that assumption entirely. The model has no ear. It has no sense of when something is off. It has pattern recognition at extraordinary scale and no felt understanding of what your organisation is actually trying to say. Feed it a style guide and it will produce content that follows the guide. Feed it nothing and it will produce content that follows the statistical average of everything it has seen. The output is your brand, for better or worse.
The worse is what organisations are discovering, slowly and often painfully. AI as a content production system has exposed the difference between “describing” a voice and “governing” one.
A description tells you what things should look like. A governance system tells you what things must do, what rules constrain them, what decisions they are authorised to make, and what happens when they go wrong. Governance is structural. It operates at the level of the organisation. Describing a voice in four adjectives is as useful for governing AI content as a list of architectural aesthetics is for running a city’s building code.
The stoic emperor Marcus Aurelius, who spent twelve years writing “Meditations” as a private governance manual for his own mind, understood this distinction intuitively. He was not interested in describing what a virtuous man looked like. He was constructing a repeatable internal system that would make virtuous decisions automatic under pressure.
The “Meditations” are not a style guide. They are an operating protocol. Every entry is an instruction, a principle tested against a specific temptation, a rule stress-tested by circumstance. That is the level of architecture brand voice needs now.
So what does an actual voice architecture look like?
It begins, as governance systems do, with authority. Someone, or some function, holds the decision right over what the brand says. This is not a committee that reviews the document every two years. It is a living function with clear accountability, the power to reject outputs, and the obligation to update the system when the world changes. Governance without authority is decoration under a different name.
The word choices in a press release and the word choices in an automated customer service reply are governed by different pressures, different audiences, different risks. A voice architecture maps these contexts and assigns different levels of constraint to each. Some outputs need tight guardrails and human review. Others can be automated with confidence once the parameters are set. The map of those decisions is a governance instrument. A list of adjectives is not.
Not all communication choices are equal.
The AI learns from the whole. A voice architecture specifies the canon: the set of outputs that actually embody the brand at its best, curated with the same care you would apply to a legal precedent, because that is what it is.
When you feed an AI system the content it should emulate, you are making a consequential architectural choice. Most organisations feed it whatever is available, the content that already exists, which is a mix of everything the brand has ever produced, including the content that was off-voice, rushed, inconsistent, written by someone who had not read the style guide.
The Confucian philosopher Xunzi, writing in the third century BCE on the nature of ritual propriety, argued that “li”, the system of rites and forms that structured social conduct, was not a constraint on human nature but its completion. Without external structure, he wrote, the individual has no reliable way to act well consistently. The structure was not what prevented good conduct. The structure was what made good conduct reproducible. Xunzi understood governance as enabling technology, not limiting technology.
This is the reframe organisations need to make. A voice architecture is not a restriction on what AI can produce. It is what makes the AI capable of producing work that carries real brand identity. The constraint is the capability. Without a governance structure, AI produces content that is technically correct and strategically nowhere.
Organisations that have not built a voice architecture are not just producing inconsistent content. They are outsourcing identity formation to a probabilistic text generator. Every output that goes out without governance is a micro-vote for what the brand is. The accumulation of those votes, at the volume AI now enables, shapes perception faster than any positioning exercise can correct. Brand dilution through AI is not dramatic. It is incremental. It is the slow replacement of a specific voice with a generic one, sentence by sentence, week by week, until the brand sounds like everybody else.
The philosopher of technology Gilbert Simondon, writing in the 1950s in work that has only recently received the attention it deserves, argued that understanding a technical object properly means understanding its “mode of existence”: what makes it function as a coherent, individualised thing rather than a collection of parts. A brand voice, properly understood, has a mode of existence. It is not a set of rules. It is a way of being present in language. When you build a governance system around it, you are not bureaucratising the voice. You are giving it the structure it needs to survive contact with scale.
The operating system argument Intradiegetic makes is about the brand as the decision logic of the whole organisation. This is its communication correlate. If the brand is the operating system, then voice is the language the operating system speaks. And a language without syntax, without rules of formation, without a governing logic that distinguishes valid statements from invalid ones, is not a language. It is noise wearing the costume of a language.
The AI is not the threat. The absence of architecture is the threat. The AI simply makes that absence visible, loudly and at speed.
Build the governance system before the volume builds. The window in which it is still cheaper to build it than to repair the damage from not building it is closing faster than most organisations think.
Leave a Reply
You must be logged in to post a comment.