When we talk about Harmonee's earliest deployments, the conversation often starts with healthcare or finance — the industries that most obviously can't move data to a cloud LLM. The actual leading edge of on-prem AI deployment turned out to be county government.
Counties have an unusual combination of constraints. The data sensitivity is real — citizen records, internal correspondence, sometimes records that touch active investigations. The procurement process is rigorous, but it's also fixed: hardware purchases follow a known path. State-level audits are routine. The IT teams are small but technically literate.
What counties also have, that larger institutions don't, is the freedom to deploy faster than the alternatives. A regional hospital takes six to twelve months to evaluate a new vendor. A county can move from initial conversation to live deployment in three weeks if the architecture matches the requirements. We've seen it happen.
The other thing counties have is operational pressure that's hard to absorb with headcount. Budgets aren't growing. Demand on operational departments is. The administrative layer of running a county — permits, citizen correspondence, internal coordination, records-request response — has been pulling staff time away from public-facing work for years. AI that handles the paperwork without exposing citizen data to a vendor is the rare proposal that makes both the operations director and the IT director say yes.
We expect healthcare and finance to be the larger market in the long term. But counties are the leading indicator. They show what an on-prem AI deployment looks like when the procurement, the security review, and the operational rollout are all moving on the timeline that the work actually demands. The pattern that's working in Klamath County is the pattern we're carrying into the next ones.