A $140M Series B, no AI product
A million businesses automate conversations on Instagram, WhatsApp and Messenger with Manychat. After a $140M Series B it needed AI inside them, and none of the patterns that requires existed.
Before: automation as a hand-built rule
Every reply was a branch drawn by hand. Predictable, inspectable, bounded. That was the bar.
Ship the first AI products. Build the system to scale them.
- Take AI Behaviour and AI Replies from zero to one, and own the quality bar for AI surfaces
- Create and lead the AI Design Playbook
- Own the design system roadmap and its foundations
- Own content design: team, operations and guidelines
- Manage seven designers across two teams
Nobody had done this here before
No prior AI feature to learn from, so discovery worked backward from the automations people already trusted.
The gap was not capability, it was configuration. Nobody setting up AI could say who it should sound like or what it must never say, and every designer touching an AI surface invented its patterns alone.
Teaching tone from real conversations, not a blank field
AI BehaviourGive it a starting point, not a blank form
Nobody can describe their brand voice into an empty box, so the setup never shows one. It learns tone from real conversations and asks the owner to approve examples.
Guardrails mattered more. When a question crosses one, the AI declines rather than improvises.
The preview where the AI declines to guess
Close the knowledge gap in the flow that found it
AI Replies answers only from what the business has taught it. Where it has nothing it says so, and the gap is fillable on the spot.
When it cannot, the conversation goes to a person. A wrong answer lands in a customer's inbox.
The playbook behind both of them
Every decision from shipping those two became the AI Design Playbook. AI features are not a category but a set of jobs, so it is indexed by job, not screen.
The public version, at aidesignplaybook.site
Foundations rebuilt for AI-product speed
The system held a stable library. AI products needed it to absorb patterns continuously, so I led the rebuild alongside the AI work.
I ran a workshop with the people who use the system daily, across design, engineering, content and brand, and had them cluster the problems.
The second cluster mattered more. Any team can rebuild a library. A system where nobody knows who decides decays within six months.
So what got built was not a library but one source of truth across design, engineering, content and brand, with a contribution and approval model: who proposes, who signs off, how it versions.
An AI product speaks twice
The AI talks to the customer. The product talks to the owner about the AI: what it will do, what it will not, what it did. Collapse them and the product impersonates what it explains.
So content design sat inside both teams rather than reviewing their output. Guidelines set which voice owns which moment, and the terminology went into the system.
What the AI reports back to the owner
Two calls that shaped the outcome
Ship the product and the pattern language together
Waiting for a finished design system would have meant shipping nothing. The Playbook came from building the real thing.
Trust is a design decision, not a legal disclaimer
Every AI surface had to show what it was about to do and make it easy to undo.
A validated product line, and a system for the next one
Both shipped as betas to a segment of users and made $200K before general availability. That turned AI from an experiment into a funded product line.
What this changed
An AI feature isn't trustworthy because it works. It's trustworthy because the person using it can see what it's about to do.