EnterpriseOne of the world's largest enterprise software platforms
Client under NDA
Agentic AI, made legible at the scale of an enterprise platform.
The name stays behind the contract, as these engagements usually do. The work does not: software that acts on its own is only adoptable if people can tell what it is about to do. We work at the platform layer on exactly that — the patterns governing how an agent introduces itself, asks permission, shows its reasoning, and hands control back.
Consistency at this altitude is not cosmetic. A pattern set once in the design system arrives in every product built on the platform — which is how thousands of teams ship intelligence that behaves the same way twice.
- 01
Introduce
An agent says what it is and what it can touch before it does anything at all.
- 02
Ask
Consequential actions wait for a person. Permission is specific, scoped, and easy to refuse.
- 03
Show
Reasoning is visible at the level a person needs to trust the outcome — not a transcript, not a black box.
- 04
Return
Control comes back cleanly. Whatever the agent did can be reviewed, undone, or taken over.
Try it — a fictional assistant following the four patterns
- 01 Introduce
- 02 Ask
- 03 Show
- 04 Return
Scheduling assistant — demo
I’m the scheduling assistant. I can read your calendar and suggest changes. I can’t change anything without asking you first.
- Design systems
- AI interaction patterns
- Platform governance
- Accessibility