Open a typical set of brand voice guidelines and you find a row of adjectives. Friendly, professional, approachable, expert. Hand that list to a new member of staff and they will nod politely and carry on writing exactly as they did before. Hand it to a website assistant and much the same happens.
The problem is that every business on earth wants to be friendly and professional. The words do not describe anything. What makes a local bakery sound different from a firm of surveyors is a set of specific habits: how long the sentences are, whether they say "we" or use the company name, how they deliver bad news, whether they ever use an exclamation mark.
Keeping an ai brand voice consistent means writing those habits down as instructions and examples the assistant can actually follow. It is less mysterious than it sounds, and it is about an afternoon's work.
Why the assistant drifts off-voice
An assistant that answers from your content tends to pick up the style of that content. If your website was written by an agency in a polished corporate register, and your team actually speaks to customers in a warm, plain way, the assistant will sound like the website, not like your team.
Two other things push it off-voice. Long answers, because a three-paragraph reply gives far more room for generic phrasing than a two-line one. And awkward situations such as complaints, refunds and "why is it so expensive", where assistants default to a careful, slightly stiff politeness unless told otherwise.
Turning adjectives into rules
Take each adjective in your current guidelines and ask what it means in a real reply. Then write it as a pair of do and do not.
| Vague guideline | Rule the assistant can follow |
|---|---|
| Friendly | Use the visitor's first name once if they give it. Do not open every reply with an exclamation. |
| Professional | Give prices and policies plainly, with no softening words. Do not use slang. |
| Approachable | Short sentences. Explain any trade term the first time it comes up. |
| Expert | Say what we recommend and why in one sentence. Do not hedge with 'it depends' without saying on what. |
| Honest | If we do not offer something, say so in the first sentence, then suggest what we can do. |
Five or six of these rules will cover most of what makes your voice yours. More than ten and they start contradicting each other.
Notice that several rules are about length and structure rather than word choice. That is deliberate. The quickest way to make an assistant sound like a small, confident business is usually to make it say less: answer first, add one useful detail, offer a next step, stop. Long replies are where the generic filler creeps in.
Building the voice sheet, step by step
- Collect ten real replies you are proud of. Emails, messages, anything your team sent to customers that sounded exactly right. These are your raw material.
- Underline the habits. Sentence length, greeting, sign-off, how prices are mentioned, what words keep appearing.
- Write the rules table as above, one row per habit.
- List words you never use. Every business has them. A salon might ban "treatment package". A law firm might ban "no worries".
- Write two or three example exchanges for the situations where voice matters most: a price question, a complaint and a request you have to decline.
- Put it into the assistant. The rules and banned words go into the persona instructions. The example answers go in as Q&A pairs in your own wording.
In SpideyChat, the name, persona, welcome message, suggested prompts, brand colour and avatar are all set per assistant, so a business running several brands can give each its own voice. The AI reply generator is handy for drafting alternative wordings of a tricky reply that you then edit into your own style.
What on-voice sounds like in a hard moment
Declining is where voice shows most. Here is a garden centre, whose voice rules say: short sentences, warm but not gushing, always offer something useful.
Visitor: Do you deliver to Cornwall? I'm looking at the olive trees.
Assistant: Sadly not. We only deliver within 30 miles of the shop, because the big trees go on our own van. If you're ever passing, we can hold one for up to a week. Or I can take your details and let you know if that changes.
And the same reply with no voice guidance:
Assistant: Thank you for your enquiry! Unfortunately, we are unable to offer delivery services to Cornwall at this time. We apologise for any inconvenience this may cause. Please do not hesitate to contact us if you have any further questions.
Both are accurate. Only one sounds like a person who works at a garden centre, and only one gives the visitor anything to do next.
Where voice must give way
A brand voice is not a licence to be clever at the wrong moment. Some rules override style entirely:
- Money and policy get stated plainly, even for a playful brand. Nobody wants a joke next to a cancellation fee.
- Complaints get a short acknowledgement and a route to a person, not personality.
- Health, legal and safety topics stay factual, with a clear note that the assistant gives information rather than advice.
- Accuracy beats tone. If making a sentence sound more like you would make it less accurate, keep the accurate version.
It also helps to tell visitors plainly that they are talking to an assistant. A strong voice makes that easier, not harder; whether to tell customers the receptionist is AI covers the wording.
Your afternoon's work
Find those ten replies you are proud of, today, before anything else. They are the hardest part to gather and the most valuable. Build the rules table from them, pick the three hardest situations for your business and write the example exchanges. Then ask your assistant those three situations and compare the answers with yours, line by line.
After launch, pull five transcripts every few weeks and read them aloud. The off-voice lines are obvious when heard. For the persona side of this, naming and character, see choosing a name and persona for your AI assistant, and if you have not yet scoped what the assistant is for, write a job description for your AI employee first.