COMMERCIAL IN CONFIDENCE: please seek permission before sharing

Options for MyService Condition Classifier Change

Meeting: Nick Miller, Gwen Collins, Andrew Miller, Celia Perry, Les Jordan, 3 June 2025

Author: Nick Miller, Gov Law Tech

Talking Points:

  • Current status: existing Microsoft LUIS service to be retired in October
  • Options
  • Recommendation

Demo Input

Options

Option Number Description Demo Results
One

Turn off service. Change request to SvA. Settle a new user experience with SvA.

NA Impact depends on what the new user experience would be. Past analysis shows around 40% of conditions get matched to a decision ready condition with the current approach. So if that suddenly stopped, accepted conditions under the decision ready policies would fall by amount from MyService.
Two

Migrate to Microsoft's Conversational Language Understanding.

Steps: Gov Law Tech migrates existing config; DVA staff get familiar with new interface.

Cost: Minor project to upgrade, test, learn; Contract variation for Gov Law Tech - uncertain scope, T&M.

Pro: Probably we will be able to get close to the current status quo, so you can say you are not changing anything, or only making a minor change. No change request required with SvA.

Con: We're only using a small part of a large and fairly complex new tool - there will likely be minor hitches and unforeseen wrinkles in the transition.

Three

Semantic Search against SoP Condition Definitions

We use a SOTA sentence-embedding service to search the condition definitions directly from the SoPs.

Steps: Gov Law Tech can change the service; DVA needs to test and decide a threshold for a match, or decide to have no threshold; no SvA change request required.

Pro: Early testing indicates it is superior than current approach; has a principled, explicable foundation; no ongoing maintenance required from DVA, less work than current setup; simple and low risk to implement.

Con: Change management required from DVA: the outcomes will change from the current status quo - possibly more claims accepted, possibly less (depends on thresholds), likely better classification.

Four

Large Language Model

Here we send a list of candidate condition definitions and the user's description to an LLM, and request the best match.

Steps: It's a moderate complexity project, involving Gov Law Tech, DVA and SvA.

Pro: Significantly improved outcomes for veterans and DVA. Public service medals will rain form the sky.

Con: Manage expectations around hallucinations; requires new user experience; political issues around AI; needs high level executive champion.

An unhandled error has occurred. Reload 🗙