Brand voice is easy to describe and hard to protect. Most teams can articulate the qualities they want—clear, confident, human, distinctive. Fewer can maintain those qualities once the work moves into regular production. AI sits in the middle of this problem. It can accelerate the discovery and testing of voice. It can also accelerate the drift toward averaged professional language.
I have used AI both ways. Early on I watched it flatten voices that had started with real particularity. Later I learned to treat it as a pressure-testing and expansion tool rather than as a source of the voice itself. The difference is not in the model. It is in the sequence of decisions and the standards applied to the output.
This article lays out a practical approach for using AI in brand voice work without letting the tool sand down the edges that make the voice worth having.
Why Voice Flattens
Voice flattens for predictable reasons. Teams optimize for clarity and broad acceptance. Legal and stakeholder reviews remove risk. Writers working at speed reach for familiar constructions. Over time the language becomes competent and interchangeable.
AI introduces a new flattening mechanism. Models are trained on large volumes of existing text and tend to reproduce the statistical center of professional communication. When asked to write in a brand voice, they often deliver a smoothed version of the input examples. The result sounds consistent and on-brief while losing the specific rhythms, asymmetries, and points of view that made the original samples distinctive.
The flattening is rarely dramatic. It happens in small substitutions: a sharper verb replaced by a safer one, a particular sentence shape replaced by a more balanced construction, an opinion softened into a general observation. Each change is defensible. The cumulative effect is a voice that no longer belongs to anyone in particular.
What AI Can Actually Contribute
Used carefully, AI can help with several parts of voice work.
It can surface patterns in existing writing that are hard to see at scale. Feeding a body of strong brand writing into a model and asking for recurring syntactic habits, preferred sentence lengths, or distinctive word choices can make the voice more explicit and teachable.
It can generate controlled variations. Once the core voice is defined, the model can produce multiple versions of the same message at different levels of formality, length, or intensity. These variations help teams understand the range the voice can sustain without breaking.
It can act as a stress test. Asking the model to write the brand’s message in the voice of a competitor, or to restate it for a skeptical audience, often reveals where the original voice is soft or generic.
It can accelerate drafting inside an already-defined system. When the rules, examples, and rejection criteria are clear, AI can produce first passes that a human editor then sharpens. This is different from asking the model to invent the voice.
The common requirement across these uses is that the distinctive core must already exist or be actively protected. AI does not originate particularity. It can help develop, test, and scale it.
A Sequence That Protects Specificity
The workflow I use now keeps the human decisions about voice upstream of generation.
First, collect and protect the source material. I gather the strongest existing examples of the brand’s writing—the pieces that already feel most like the voice at its best. These examples are treated as primary evidence. They are not summarized or generalized too early.
Second, make the voice explicit in short, concrete terms. Instead of lists of adjectives, I write notes about sentence behavior, stance toward the reader, preferred and forbidden constructions, and the kinds of claims the brand is willing to make. The more operational the description, the more useful it becomes as a constraint.
Third, use AI to expand and test inside those constraints. I provide the source examples and the operational notes, then ask for new material that stays within the defined territory. I also ask for deliberate violations so the boundaries become clearer.
Fourth, edit with the original source material as the standard. Generated drafts are compared against the strongest human examples, not against the prompt. Anything that softens, generalizes, or balances away the particularity is rewritten or cut.
This sequence treats AI as a tool for range and pressure rather than as the origin of the voice. The particularity has to be supplied and then defended.
Practical Techniques Against Flattening
A few specific techniques help keep the edges intact.
Keep a living anti-voice list. Record the phrases, rhythms, and moves that signal averaged professional language for this brand. Update the list as new defaults appear in AI output. The list becomes a fast filter during editing.
Require concrete anchors in every piece. Abstract brand language is the first to flatten. Insisting on specific observations, examples, or claims forces the writing to stay closer to lived detail and farther from generic elevation.
Separate generation from evaluation. Reviewing AI output in the same session in which it was produced tends to lower standards. Distance makes flattening easier to detect.
Use the model to write the wrong version on purpose. Asking for the most generic possible version of a message, or the version a competitor would write, makes the distinctive version clearer by contrast.
Protect asymmetrical elements. Distinctive voices often contain slight imbalances—unexpected sentence lengths, recurring unusual constructions, a willingness to leave certain things unsaid. AI tends to regularize these. Human editing has to restore them deliberately.
When AI Is the Wrong Tool
There are stages of voice work where AI adds little value and some risk.
Early discovery of a new voice is usually better handled through human writing and conversation. The model can only recombine what it is given. If the distinctive core does not yet exist, generation produces plausible averages.
High-stakes definitional pieces—positioning statements, manifesto-length arguments, foundational about pages—benefit from slower human drafting. AI can later help adapt these pieces, but the original articulation is better done without the smoothing effect of the model.
Situations that require a strong personal or authorial presence also resist useful AI involvement. The particularity in those cases is often inseparable from a specific person’s syntax and experience. The model can imitate surface features. It cannot supply the underlying stake.
Knowing when to leave the tool closed is part of using it well.
Measuring Whether the Voice Is Surviving
I use a few practical signals to check whether AI involvement is helping or eroding the voice.
Do new pieces still sound continuous with the strongest older pieces when read side by side?
Are people inside the brand able to recognize the voice immediately, or does it require explanation?
When the writing is stripped of brand names and visual identity, does it still feel specific to this organization?
Is the editing process mainly removing generic language, or is it mainly adding distinctive language back in?
If the answers start trending in the wrong direction, the process needs tightening. Usually the source examples have been neglected, the operational description of the voice has become too abstract, or the editing standards have loosened under time pressure.
Building a Sustainable Practice
Voice work is ongoing. Brands change, audiences shift, and the surrounding language environment evolves. A system that protects voice has to accommodate revision without losing continuity.
I keep the source examples and the operational voice notes in a place that is easy to revisit. I update them when the brand’s stance genuinely changes, not when a single piece of content feels difficult to write. I also keep examples of failed AI output that flattened the voice, so the failure modes stay visible.
Over time the combination of clear source material, explicit constraints, controlled generation, and strict editing produces a more stable result. The voice can flex across formats and topics without dissolving into the general professional register that AI reproduces so easily.
Closing
AI can help find and develop a brand voice. It can make patterns visible, generate useful variations, and stress-test boundaries. It can also flatten the very particularity that makes a voice worth protecting. The outcome depends on the sequence and the standards.
Keep the distinctive source material primary. Make the voice operational rather than purely adjectival. Use the model for expansion and pressure inside clear constraints. Edit against the strongest human examples rather than against the prompt. Protect the asymmetrical elements that give the voice its character.
The tools will keep improving at imitating style. The responsibility for deciding what is worth imitating, and what must remain non-negotiable, stays human. Brands that treat that responsibility seriously will be able to use AI without sounding like everyone else who is using the same tools. The ones that do not will find their language becoming more fluent and less specific at the same time.
That trade-off is avoidable. It simply requires treating voice as a set of decisions to be defended rather than as a tone to be approximated.
Travellers Write
No letters yet — be the first traveller to write.