The Design Office of Nick Sloggett
Back
EssaySep ’262 min read

A new surface inherits the rules

AI products are leaving the chat window for plugins, automations and other people's assistants. The trust built in the chat window doesn't go with them unless it was a rule of the platform.

Any assistant that can change something has to answer one question before it leaves the chat window: what has to be true before software nobody is watching gets to act?

On September 23, Rippling introduced plugins, skills and automations, a month after Rippling MCP, which lets assistants such as Claude, Codex and Cursor connect to Rippling and act in it. I led design for Rippling AI, so I’ve spent a long time on that question.

In a chat window, a person is right there. They asked, they see the answer, and nothing changes until they say yes. Rippling AI writes a proposed change down as the table payroll will receive, the current amount beside the new one, and waits.

Every surface past chat takes a person further from that moment. A plugin pulls in another company’s tool. A skill runs a team’s written-down way of doing something. An automation runs with no prompt at all, on a schedule or when data changes. And over MCP, the request comes from someone else’s assistant, in someone else’s interface.

If the confirmation lives in the chat window’s design, none of those surfaces gets it. Each one has to rebuild trust from scratch, and some won’t.

So trust has to be a rule of the platform. In Rippling’s case, every request over MCP runs as the person who made it, with the same permissions they have everywhere else. In Rippling’s own words, “the model proposes the work, and Rippling decides what is allowed.”

There’s a second-order effect worth designing for. A rule that lives under every surface also decides what the next surface can be. A new place for the assistant to show up is cheap to add when it inherits the permission model on its first day, and expensive when it needs its own. The platform rule turns out to be a roadmap decision as much as a safety one.

Rules like this cost something. Rippling AI won’t act without a person’s approval, so every change costs a read and a click, including the changes it would have got right. It asks “Which John do you mean?” before starting, which is one more turn. Teams that want the software to just do it will find that slower.

For anything that touches a paycheck, I’d pay it. A wrong change in payroll lands in someone’s bank account, so the receipt has to come before the change.

Any team taking an assistant past chat should list its promises and mark where each one lives: in the window, or in the platform. Only the second list goes everywhere the assistant goes.