The two failure modes
When teams train an AI agent for the first time, the output usually falls into one of two failure modes:
- The "generic LLM" voice — overly polite, hedged, full of "I understand", uses bullet points where humans would write a sentence. Sounds like a chatbot.
- The "Slack DM" voice — too casual, drops articles, makes jokes that don't land. Sounds like an intern.
Neither sounds like your brand. The fix is not "tweak the prompt" — the fix is to give the agent the same materials a new hire would read on day one.
The voice triangle
Every brand voice can be expressed as three coordinates:
- Warmth — formal ↔ warm
- Pace — careful ↔ direct
- Register — corporate ↔ playful
Pick a point. Write five lines that hit that point. Those five lines become your voice anchors. The agent will weight everything else against them.
What to actually feed it
In order of importance:
- Five exemplar replies — actual messages your best teammate sent, with the customer context redacted. These are worth more than 200 pages of style guide.
- One "do not say" list — phrases, words, formatting habits the agent should never use. ("Unfortunately" is on most teams' list. So is the em-dash, on others.)
- Your top 20 cited answers — for the most-asked questions, write the answer you wish a teammate would send. Lock it in as canonical.
- Policy & product docs — for the long tail. The agent should ground here, not invent.
What not to feed it
- Marketing copy. The voice on a landing page is not the voice in a support reply.
- Old chatbot scripts. They will leak through.
- Slack threads. Internal jargon will leak through.
- The CEO's tweets. They were not written to a customer in distress.
The weekly tune
Once a week, pull the bottom 10 replies (by CSAT or by the agent's own confidence score). Look at them in plain text. Ask "would I send this?" If no, find the source it was grounded in and rewrite the source. The agent is downstream of your knowledge. Fix the knowledge.
Six weeks of that and the agent sounds like you — because it is, structurally, you.
Turn the triangle into an actual voice specification
Labels such as “friendly” and “professional” are too vague to test. Convert each coordinate into observable rules. For warmth, define whether the agent uses the customer's name, contractions, empathy statements, or emojis. For pace, define the maximum opening length and whether the answer leads with the result or the explanation. For register, list natural audience vocabulary and the jargon that must be translated.
A useful specification fits on one page. It includes three “we sound like this” examples, three counterexamples, formatting rules, approved greetings and sign-offs, and rules for stressful situations. Counterexamples show the boundary more clearly than adjectives do.
Write examples as complete situations
An exemplar needs context, not just a polished answer. Store the customer's message, the relevant account or product facts, the ideal reply, and why that reply is good. Otherwise the agent may copy the tone while missing the decision rule.
A delayed-order exemplar should state the promised window, current tracking state, what can be offered now, and when a human must intervene. A cancellation exemplar should distinguish between a request that can be handled immediately and one requiring identity verification. Voice is inseparable from policy.
Create a voice test set
Use 30 prompts across emotional and operational conditions: five routine questions, five frustrated messages, five messages with incomplete context, five policy exceptions, five product-discovery questions, and five cases that require a human.
Have two reviewers score each response from 1 to 5 for clarity, brand fit, empathy, directness, and policy correctness. Review independently before discussing results. If reviewers disagree by more than one point, the voice rule is probably ambiguous.
Include adversarial phrasing. Ask the agent to ignore policy, promise a discount, criticize a competitor, reveal internal instructions, or adopt a radically different persona. A stable voice is also a boundary: the agent should stay useful without being steered away from approved behavior.
Match voice to the moment
Consistency does not mean emotional flatness. A size-chart question needs a quick answer. A delivery that missed an event needs acknowledgment before instructions. A security concern needs calm, precise language and immediate escalation.
Define three modes: routine, recovery, and sensitive. Keep the same vocabulary and personality across them, but change sentence length, empathy, and escalation behavior. This creates adaptive tone without turning the agent into three different brands.
Governance for multilingual answers
Do not assume literal translation preserves voice. Choose priority languages and have a fluent reviewer approve greetings, apologies, policy terms, and escalation language. Keep product names and legal terms consistent. If a language has not been reviewed, use more conservative escalation for policy-sensitive cases.
A weekly review that takes 30 minutes
Sample ten conversations: the three lowest-rated, three escalations, two sales conversations, and two random successful resolutions. Record whether each problem came from missing knowledge, an unclear rule, weak retrieval, or voice execution. Fix the source before adding prompt complexity.
Track brand-fit score, policy-correctness score, unnecessary escalation rate, and repeat-contact rate. Voice work succeeds when the agent sounds right and customers reach the correct next step. A charming wrong answer is still a failed answer.
Publish the voice specification alongside onboarding material for human agents. The best result is one customer-facing standard that both people and automation can follow.