Case: GovGPT
A government agency rolling out an internal LLM to 20,000 employees. Exploring risks around confidentiality, correctness, and ROI.

Disclaimer: GovGPT isn't the real name. The agency is part of the government but won't be named.
The situation was straightforward. Out of 20,000 employees, 3,000 were already using ChatGPT for work.
Three problems: no control over what data was leaving the organisation (confidentiality), no visibility into whether outputs were actually useful (correctness), and costs were spiralling (ROI).
The solution: GovGPT. Hosted internally. More control. Usage-based pricing instead of per-license costs.
The workshop
Within 2 hours, I introduced the RiskStorming format, walked them through the AI card deck, and wrapped up with clear action points.
I added some thought experiments: - If a product or marketing specialist was on your team, what would they push for? - Go through the risk mitigation activities. Turn sideways any you're already doing well. - Put a red dot on the highest priority risks.
What they committed to:
- Hallucination risk - Increase observability. First, define what that actually means. Find specific cases. Then run a Proof of Concept.
- Poor prompting - Start evaluating user prompts. Look at available tools. Run a Proof of Concept.
- User expectations - People expect ChatGPT but will get something different. More training videos, presentations, internal events.
My honest assessment:
I was worried about the "big monster LLM" approach. Kill ChatGPT, launch GovGPT, stuff it with internal data.
My advice: raise these risks with management. Get someone doing internal marketing. Talk to users. Start a newsletter. Run training courses. Create case studies showing ChatGPT vs GovGPT.
And for the love of everything: don't ingest any internal data until it's been verified and cleaned up. Start from there.

