I had our MCP and API work close to being merged, then pulled it back. I wasn't comfortable enough with how access would work.
Connecting an AI assistant to a business system opens up useful possibilities. It can ask questions about the data and eventually carry out work. But I wanted the admin to understand who was connected and what they could access.
The direction I described was to start with read access, within the permissions the person already had in the organisation. Write access would be a separate decision. I also wanted access limited to selected organisations, rather than opening everything at once.
Part of my concern was speed. A person takes time to create things in the system. An assistant can produce a lot of changes quickly. If it heads in the wrong direction, that can become a lot of unwanted work to untangle.
So I went back through the design instead of treating the connection itself as the finish line. We planned to trial it internally before a wider rollout. That was still the plan in the recording, rather than a public release announcement.
It's exciting to build something that lets people work with their data in new ways. I still want the access decision to be understandable to the person responsible for the business.
For me, delaying that merge was worthwhile. Getting the feature connected was only one part of the work. Being clear about its permissions, scope and rollout was part of deciding whether it was ready.
