The situation

Two business goals sit behind our client's customer-facing AI agent: increase sales, by moving people who would otherwise ring a salesperson into buying on the web, and reduce drop-off across application and enrolment. It is a channel of preference, not the only way through, so it has to be good enough that people choose it over the phone.

The client's agent runs from "what should I study" through to payment and enrolment: someone browsing on a Sunday night, and that same person three weeks later when their card is declined.

They asked us to apply their brand tone of voice to it: four adjectives in the brand strategy, each with a sentence of explanation. On paper this was copy and paste. We assessed it first, which is the only reason the change never happened. As written, those adjectives would have made the agent worse, and one would have created a compliance exposure.

A brand tone of voice briefs copywriters, and their work travels through layers of approval before a customer sees it. An agent has none of that. It writes its own words, in situations nobody chose.

The challenge

The risk is structural. Tones copied into an agent arrive with no sense of when each applies, so it reaches for all four everywhere. Confident and compelling suits someone browsing. Pointed at someone whose payment has just been declined, it frustrates.

Assessed on paper, the swap produced seven risks, three of them high severity. The fourth tone asked the agent to be confident, and confidence about outcomes is a regulated claim here: claims about what study will do for you sit under Australian Consumer Law. "Either of these will get you there" is that tone written correctly, and a promise the provider cannot make.

Large language models also treat descriptive words as instructions. Welcoming, compelling and inspiring are vague terms, not operational ones. Left undefined they read as warmth in brand copy and produce sycophancy in a system prompt. Defined, they become decisions a model can actually make.

The incumbent voice was safer, but not perfect. Friendly, professional, leaning formal, never authored but read off slider positions on a workshop board. It would pass review anywhere, and describe almost every customer-facing agent in market. One voice was too generic to be worth having, the other too risky to use.

Our approach

1. Light desk research on how brand voice behaves differently in an agent.

2. Analysis with Claude, running the swap as a scenario against the persona, the customer personas and the compliance position.

3. Each tone made operable, restated as behaviour the agent can act on mid-conversation.

4. A framework built with Claude: two layers, three registers.

5. Into the agent, replacing the Voice and Tone section of its persona.

6. Testing next, with real users across all three registers.

What we proposed

The fix was two layers.

Character never changes, and opens with a rule that outranks every trait in it: the job to be done comes first, the personality second. We anchored the rest on the one trait already true of the agent and protective rather than exposed: transparency and directness. Answer directly, limits first. Say "I can't confirm that" rather than cushioning it, then name who can.

Register moves with emotional context. Exploring covers discovery: warm, surfacing possibilities. Transacting covers cart and payment: brisk, exact, surfacing consequences. Repairing covers payment failure and distress: unhurried, surfacing options. Each trait has a setting in each register, including off. Empathy is not one of the four adjectives, but it is the principle behind that register's energy: acknowledge what has happened before the fix. The layer is modular, so a register can be retuned or added without touching the others.

The same session, two registers.

Exploring:

Two fit. Course A has you running projects end to end in about six months. Course B takes longer and adds governance. Want them side by side?

Repairing:

That didn't go through, your bank declined the card. Nothing's lost, your cart is still there. You can try another card, switch to invoice, or come back later.

Honesty and directness hold across both. Warmth, pace and what gets surfaced move.

The takeaway

Write each trait as a behaviour with a boundary. "Be magnetic" cannot be built or tested. "State retrieved facts without hedging about what the product does, never about what will happen to this person" can.

Read your trait words as instructions to a model that will operate without approval layers. Some adjectives that produce good copy produce poor behaviour.

Separate what never changes from what has to move. What moves is whatever depends on context, such as warmth and pace. Pick one anchor trait that holds in your best moment and your worst.

Think about the worst moments your agent will handle. Not the homepage. The failed payment, the rejection, the complaint.

And do not assume the voice already in your agent is the safe baseline. Ours had weaker provenance than the brand voice we were measuring against.

Not sure how your brand will behave when an AI agent speaks for it

The Omnia Collective works with education and enterprise organisations across Australia to turn brand voice into agent behaviour, defined for each customer moment and checked against the compliance position. See our digital experience and CX services or get in touch.