Defining how an AI product should talk
When Whatfix launched a family of AI features, nothing governed how they spoke. I designed the system that did.
Role
UX Designer & Writer, acting content lead. Content strategy, AI guidelines.
Context
Whatfix, 2024–25. B2B SaaS (digital adoption).
Outcome
A five-pillar system that shipped and governed the AI umbrella at launch
The product
Whatfix is a B2B digital adoption platform - in-app guidance, walkthroughs, and self-help layered onto other software so people can learn it without leaving the screen. As it rolled out a family of AI features, the open question wasn't what they'd do, but how they'd speak. For more information, see www.whatfix.com
The chair
I treat content as a system, not a set of strings, so when a product starts speaking, I ask the same question I bring to any interface: what is this for, and how should it behave when there's a human on the other side? Whatfix was rolling out an AI umbrella, a family of AI-driven features, and nothing defined how any of them should communicate. I designed that system.
Context
The AI features were shipping without a shared voice or any rules for the hard moments: when the AI is uncertain, when it fails, when it needs to hand a user to a person. Left alone, every feature would improvise its own tone and its own failure behavior, and the product would feel like several different assistants. As acting lead of the UX content team, I was asked to author the guidelines governing how the AI umbrella communicates, including its personality.
What I built: five pillars, weighted toward the two most teams skip
I researched how peer companies were approaching AI content, then authored a five-pillar system:
Clarity & transparency
States its certainty, limits, and capabilities.
Conversational tone
One consistent voice, aware of where the user is.
AI personality
A first-person voice aligned to the brand, not a novelty persona.
Error handling & guidance
How the AI behaves when it hits a failure or a limit.
AI–human handoff
When and how it escalates to a human.
The last two are what most AI content work underinvests in: everyone tunes the happy-path voice; far fewer design what the AI does when it's wrong or out of its depth. That's where trust is won or lost, so that's where I put the weight.
Outcome
The guidelines shipped and governed the AI features at launch, reviewed with my content team, the AI-features PM, and design with no material changes. Honest on measurement: governance work like this doesn't produce a clean metric; its effect is a product that speaks with one coherent, trustworthy voice, which shows up as an absence of confusion rather than a number. What I can point to is the artifact itself and the judgment in it.
What I carry forward
Designing how an AI speaks is a design problem, not a copy problem. The interesting decisions aren't the friendly greeting; they're the behavior at the edges: how it admits uncertainty, how it fails, when it gets out of the way and hands you to a person. Get the edges right and the personality takes care of itself.
"The interesting decisions aren't the friendly greeting. They're the behavior at the edges."