- 98
- Questions asked in live field use
- 6
- Deputies in the demonstration
- 3.7s
- Typical answer time
- $0.027
- Measured cost per question
Context
A sheriff's office holds its policy in general orders that deputies are accountable to and rarely have time to read. The questions that matter arrive mid-shift, in the minutes before someone acts, and the existing answer was to search a document system or call a supervisor. Both work. Neither is fast, and neither leaves a record of what deputies actually needed to know.
Constraints
The demonstration ran in live field conditions rather than a conference room, so it had to be usable one-handed by someone with something else going on. Adoption could not be ordered, only offered. And because a wrong answer about use of force is not a wrong answer about anything else, every response had to name the order it came from so a deputy could check it before acting.
RegainFlow's role
RegainFlow built the assistant, instrumented it, and ran the demonstration with the department. The corpus stayed the department's policy library, and the decisions stayed with the deputies. We did not touch case data, and we make no claim to have changed any outcome in the field.
What RegainFlow engineered
- Retrieval over the department's general orders, with every answer naming the specific order it came from so a deputy verifies against policy rather than trusting the tool.
- A question surface deputies could use from the field in plain language, without learning a query syntax or opening a second system.
- A shared question history, so one deputy's question becomes an answer the next deputy finds already written.
- End-to-end instrumentation on every question and answer, which is where each figure in this study comes from.
- A model-independent platform, so the assistant improves as the underlying models do without being rebuilt.
Selected artifacts
Reconstructed, not screenshotted.
How an answer gets made
- Deputy question
- Policy search
- Cited answer
What the client received
- A policy assistant grounded in the department's general orders
- Cited answers naming the order behind every response
- A shared, searchable record of every question asked
- Per-question cost and response-time measurement
- A readout of what deputies asked and where the questions clustered
- A recommended path from demonstration to department-wide operation
Outcome
Six deputies asked 98 questions in live field use. Every policy search returned an answer, typical response time was 3.7 seconds, and the questions concentrated where the stakes are highest: juvenile investigations, use of force, and domestic violence. Deputies kept asking questions off-shift, on their own time, which is the clearest signal a field tool can send.
From demonstration to operation
The demonstration proves the capability on public policy documents. Connecting the assistant to the department's private policy library, hardening it for criminal-justice deployment, and measuring adoption per deputy are the steps that turn a proven capability into departmental practice.