Privacy as Product Infrastructure
Mike Knoop, co-founder of Zapier, on Zapier's LLM-powered future
This was the gating issue that turned LLMs from a fun demo into software a company like Zapier could actually ship. Zapier was sending natural language instructions and app data through its NLA stack to let models trigger Gmail, Slack, and thousands of other actions, so if model providers kept those prompts for training, Zapier risked exposing customer workflow data at the exact point where trust mattered most. The implication is simple, privacy terms were not legal fine print, they were product infrastructure.
-
Zapier designed NLA around explicit user control because LLMs could guess wrong. Users chose which actions ChatGPT could access, could lock specific fields instead of letting the model infer them, and could preview actions before committing them. That human in the loop design only works if the model layer is also contractually safe.
-
The operational risk was concrete. A prompt like find my latest email from Jan or send a Slack message passes through Zapier, touches connected app credentials, and returns trimmed payloads back into the model loop. In that setup, training on API traffic would mean business context, message content, and workflow metadata leaving the trusted automation boundary.
-
This also pushed Zapier toward a multi model posture early. In the same interview, Zapier said users wanted provider choice, and related research on Cohere framed privacy, private cloud, and on premises deployment as key enterprise differentiators versus OpenAI. Privacy was becoming a buying criterion for the model layer, not just the app layer.
The next step is a world where automation buyers expect model choice the same way they expect app integrations today. That favors orchestration layers like Zapier, because the winning product will not just connect tools, it will route sensitive work to the model and deployment setup that matches each company’s privacy bar.