All case studiesProduction system
Service desk triage on local models, entirely on premises.
A national digitalisation agency needed AI triage for its service desk, but policy ruled out cloud processing of its data. Triage now runs on locally hosted models, 100% on premises and air-gapped, with no dependency on an external model provider.
- Client
- A national digitalisation agency
- Sector
- Public sector, IT operations
- Engagement
- Production system
- Measured result
- 100%on premises, air-gapped
100%on premises, air-gapped
Growth goal
Absorb growing service desk demand as more public services move online.
Operating constraint
Standard AI tooling depends on cloud APIs, and policy ruled out sending sensitive public-sector data outside the network.
Measured result
Tickets triaged and routed automatically, 100% on premises, with no vendor dependency for the models.
Summary
- Challenge
- Analysts read and routed every ticket by hand. Keyword rules had been tried and could not handle how people describe problems.
- What SIEL built
- Language models on the agency's own hardware classify each ticket and propose a queue. Models and routing logic ship as containers for the agency's platform, including air-gapped segments.
- Controls
- No outbound calls. Every routing decision is logged, uncertain or sensitive tickets wait for an analyst, and the desk can override any decision.
- Impact
- The agency holds the models and decides, category by category, how far automatic routing extends.
The full account
01More online services, more tickets
Every public service the agency moves online adds volume to its IT service desk. Analysts read each ticket, worked out what it concerned and assigned it to a team. Language models handle varied free text far better than keyword rules, but the standard tools send data to a cloud API, and the agency's policy on sensitive public-sector data ruled that out.
02Local models inside the agency's network
The triage service runs entirely on the agency's infrastructure. Models hosted on its hardware read each ticket from the existing ticketing system, classify it and propose a queue. Tickets below a confidence threshold, and categories the agency marks as sensitive, stay in a review queue for an analyst. Models, routing logic and the ticketing connector ship as containers and deploy through the agency's normal release process, including into air-gapped segments.
03What the agency has now
Tickets are triaged and routed automatically, 100% on premises, and sensitive data does not leave the network. Every routing decision is logged and can be overridden, and overrides feed model and configuration updates that the agency controls. There is no dependency on an external model provider: the agency can run, update or replace the models itself.
Discuss Your Operation With Our Engineers.
Describe one workflow and the systems it relies on. A senior engineer responds within two business days with an initial assessment: what we would build, what we would not, and why.