Your sales lead writes a proposal with AI help. Your office manager drafts a customer email with AI help. Your marketing person puts together a newsletter with AI help. Read any one of them alone and it is fine. Put all three side by side and they read like three different companies wrote them. One is chirpy and full of exclamation points. One sounds like a legal filing. One is stuffed with words nobody at your company has ever said out loud.
This is one of the most common issues we run into once a team gets comfortable drafting with AI, and it is almost never a tooling problem. It is an instruction problem. Harrison likes to say you should treat AI like a junior staff member, and this is a clean example of why. A new hire who has never read a word your company published will write in whatever voice they brought from their last job. The fix is the same for the software as it is for the person: hand them the house style, show them real examples of good work, and keep it somewhere they can actually find it.
Why Everyone’s Output Sounds Like a Different Company
Three things are happening at once. The tool has not read your website, your proposals, or your last two years of customer emails unless somebody put them in front of it. Every person on your team is writing their own brief every time they open a chat window, and “make it sound professional” means something different to each of them. And when nobody specifies a voice, these tools fall back on a generic business register that leans hard on words like leverage, robust, and seamless. That default is what people mean when they say AI writing sounds like AI writing.
The vendors are direct about this. Anthropic’s prompt engineering documentation, which we reviewed in 2026, says that setting a role in the system instructions “focuses Claude’s behavior and tone for your use case” and that “even a single sentence makes a difference.” OpenAI’s prompt engineering guide, also reviewed in 2026, describes its instructions field as giving the model “high-level instructions on how it should behave while generating a response, including tone, goals, and examples of correct responses.” Tone is not something you clean up afterward. It is an input, and if you leave it blank the tool fills it in for you.
Build a Style Guide Short Enough That People Read It
You do not need a brand book. You need about one page that a busy person will skim once and paste into a tool. Five things belong on it.
- Words we use. The handful of terms your customers actually say. If your clients call it the server room and not the data center, write that down.
- Words we never use. This is the highest value section on the page. Ban the filler. Ours rules out cutting-edge, synergy, best-in-class, and any sentence that opens with “In today’s fast-paced business world.”
- How formal we are. Pick a marker people can picture rather than an adjective. We aim for how we would explain something to a client across a conference table: plain, direct, no performance.
- How we open and close. Openers and sign-offs are where inconsistency shows up first. Decide whether you lead with the answer or the background, and settle on two or three closings everyone uses.
- Our stance on jargon. Ours is simple. If a technical term has to appear, it gets translated in the same sentence. Nobody should have to look anything up to finish a paragraph.
Feed It Real Examples, Not Adjectives
Here is the part most teams skip. Describing a voice in the abstract barely works. Telling a tool to be “friendly but authoritative” gets you an average of every friendly and authoritative thing ever written. Showing it three real emails you were proud of gets you your voice.
Anthropic’s prompt engineering documentation, reviewed in 2026, recommends including three to five examples for best results and calls examples “one of the most reliable ways to steer Claude’s output format, tone, and structure.” It advises that examples be relevant, mirroring your actual use case, and diverse enough that the tool does not latch onto a pattern you never intended. OpenAI’s guide agrees, recommending examples that show “a diverse range of possible inputs with the desired outputs.”
In practice this takes twenty minutes. Pull three or four pieces of real writing that landed well: a proposal cover note, a customer apology, a short internal announcement. Strip out client names, then paste them under a heading saying these are examples of how we write. If you have a before and after, use it. A weak draft next to the version you actually sent teaches more than either alone, because it shows the tool what you edit out.
Put It Where the Whole Team Actually Works
A style guide living in one person’s downloads folder is not a style guide. The test is whether a new employee can find it in their first week without asking anyone.
- Use the saved instruction feature if your tool has one. Most business AI tools now offer some form of persistent custom instructions or a reusable shared workspace that applies your rules to every conversation automatically. Look in the settings or workspace area of the admin console. Best option, because it removes the human step.
- Otherwise, pin the document. A single shared file in your company’s main document library, linked from wherever your team starts the day, works fine.
- Give it a named owner. One person keeps it current. Shared ownership of a document means no ownership of a document.
- Standardize the tool while you are in there. If four people are on four different AI subscriptions they expensed themselves, an inconsistent voice is the least of your problems.
That last point matters more than it sounds. Personal AI accounts paid on personal cards are a textbook case of shadow IT, and you cannot push a shared instruction set to a tool you cannot see.
Review It Every Quarter
Your voice moves. You add a service line, you shift who you sell to, somebody finally points out that a phrase you have loved for years sounds condescending to customers. Put a recurring thirty minute item on the calendar once a quarter and ask four questions. What did we publish this quarter that we were proud of, and should it become an example? What went out that sounded wrong, and why? Have we added any words to the never list? Did the tool change in a way that affects our instructions?
Thirty minutes, four times a year. That is the entire maintenance burden, and it fits the broader habit of preparing your team for AI without the hype: set clear expectations, then actually revisit them.
The Bottom Line
Inconsistent AI output is a documentation problem wearing a technology costume, and it is one of the cheapest problems on your list to fix. Write one page. Attach three real examples. Store it where everyone works. Revisit it quarterly.
And keep the framing in perspective: AI is the dumbest it will ever be today. The tools will keep getting better at picking up nuance from less input, and your written voice guide is not wasted when that happens. It is what makes each better tool useful on day one, because you will already know, in writing, exactly what your company sounds like.
If your team is drafting with AI and the output is all over the map, we can help you build the one page guide, load it into the tools your people already use, and find the accounts nobody is tracking. We work with small and mid-sized businesses across Denton County. Contact us today.
Sources:
Comments are closed