Shipping copilots that earn trust
The product patterns that make internal AI copilots actually get used — beyond the launch demo.
Internal AI copilots have a predictable adoption curve: 60% of the org tries it in week one, 20% still use it in month three, and 5% are power users by month six. The distance between 20% and 60% is almost entirely trust.
Patterns that build trust: always cite the source; always show your work; always let the user drill from the answer back to the underlying data. Every answer without a citation is a withdrawal from the trust account.
Patterns that destroy trust: confident wrong answers, invisible model changes, and quietly rerouting to a different model without telling users. If a change moves the quality bar, tell people.
One anti-pattern we see often: adding a 'reasoning' step that shows off how much the model is thinking. Users don't care. They care whether the answer is right and cited.
Marcus builds production LLM systems. Formerly staff engineer at two AI-first startups and a hyperscaler.
More from MarcusRelated articles
Data contracts are boring — and that's the point
Why the least glamorous idea in the modern data stack is quietly the highest-leverage one for teams tired of 3AM pages.
How we evaluate LLM applications in production
A practical evaluation framework for teams shipping copilots and RAG systems — without inventing a research lab.
FinOps for the modern lakehouse
The five levers that actually move data platform cost, ranked by impact — and the three that don't.