Back to blog
Field Notes

What we mean by 'brand-faithful'

Canary team · May 5, 2026 · 5 min read

We started Canary because every customer-facing AI we tried sounded like the same person. A polite, slightly nervous assistant who writes in cleanly hedged sentences and ends every reply with 'is there anything else I can help you with?'. That voice is fine, until you put your logo on top of it. Then it becomes a problem -- because the agent is now writing more sentences in your name in a single afternoon than your blog has shipped all year.

Voice is not a tone slider

It is tempting to model brand voice as a single number on a friendliness-to-formality axis. We tried. It does not survive contact with reality. A retail brand that sounds warm in a Slack confirmation should sound concise in an SMS shipping update and slightly more careful in a refund decline. The interesting work is in the layers, not the slider.

Three layers we encode

  • Identity: a one-paragraph charter -- what the agent is, what it is not, what it cares about. Compiled into the system prompt.
  • Register: the channel-by-channel default formality, plus an in-the-moment override when the customer's own message tells us to dial it up or down.
  • Refusal style: how the agent says no, when it must escalate, and what it never speculates about. The most under-invested layer in every off-the-shelf bot we have audited.

Building in the open

If you run a CX, ops or brand function and the AI you ship has to sound like you — not like a generic LLM with your logo on top — we would love to hear from you. The Design Partner program includes free production access, weekly office hours with the team, and a named seat on the roadmap.

Want to see this work on your stack?

Most teams ship in under a week. Take the trial for a spin.