The case as recorded.
The agent reads cases, contacts, tasks and handover state through read-only tools. It sees only what the workspace exposes.
RSI Agent / Product
A case agent for
public-service handovers.
RSI Agent is a case-handover and CRM agent for Australian public-service and community-service teams. It reads the case record, proposes the next step and changes nothing until an officer approves.
One team.
One handover at a time.
Service teams pass cases between contact officers, case workers and reviewers. Each handover depends on what was said, what has been checked and what is still open. RSI Agent keeps those facts linked to their sources and puts every proposed change in front of the person responsible.
The public prototype is a service CRM with fictional cases, contacts, organisations and tasks. Its case agent uses Claude with tool use to read records and propose tasks, notes and case updates. Before any integration, a pilot would start with one service team and one recurring handover.
An exact change with a reason.
Each proposed task, note or field update is shown as the precise change and the reason for it.
An officer decides.
Nothing is written until the officer approves. A declined proposal goes back to the agent as a decision, not as an error.
Reviewed at an exact version.
Handover approval and record updates stay behind an exact-version human review. The agent cannot approve its own work.
What you can try.
What we plan to build.
Available / Public prototype
Service CRM with a case agent.
- Case queue, records, contacts, organisations, tasks and activity
- Case agent: read-only tools, proposals and officer approval
- Replay mode with a labelled, prepared plan when live mode is off
- CASE-1042: source-linked handover with exact-version review
- Fictional records kept only in your browser tab
Planned / Scoped pilot
One service team, measured.
- Persistent, tenant-isolated records with an append-only audit trail
- Organisational sign-in and staff, reviewer and administrator roles
- Server-side agent runs with stored proposals and approvals
- An evaluation set of synthetic cases run on every prompt change
- Privacy impact assessment and a documented data path
The public prototype uses fictional records only. It does not connect to an agency system or update an external record. Optional account sign-in sends a login email only when you request it. The case agent calls Claude only when live mode is enabled on the server; the panel always shows which mode is running. No customer deployment exists yet.
Replay today.
Live when it counts.
The public demo runs the case agent in replay mode: three prepared tasks follow a recorded plan through the same tools, proposals and approval steps as the live agent. No model is called, so the demo has no running cost, and the panel always says which mode is running.
In live mode, each step goes to Claude Opus 5.5 through the Claude Messages API with tool use. The instructions and nine tools are fixed on the server: five read-only tools and four that can only create proposals. The API key never reaches the browser, requests are size- and rate-limited, and the instructions are prompt-cached. Paid processing is locked off in this release. A future owner-approved release would also require invited sign-in, a shared AUD 20 monthly allowance, a reviewed cost envelope, provider limits and your live-processing permission. Signing in or changing environment flags does not enable paid calls.
Configured, with boundaries
Your process stays recognisable.
Useful customisation starts with qualification criteria, workflow stages, field mappings, source priority, reviewer roles, response style and escalation rules. Each change should have an owner, a version and an acceptance check.
- Customer-approved requirements
- Explicit review and escalation
- Agreed retention and budget settings
Implementation that compounds
Build what the next customer needs too.
The commercial thesis begins with a scoped implementation. Repeated needs can become reusable components, while customer documents and confidential instructions remain separate.
We will measure whether onboarding and support become more repeatable before expanding into a general platform.
MODELMake the assumptions visible
Calculate value
before claiming it.
Change the assumptions below. These are local planning illustrations, not observed customer outcomes or a price quote.
Handling includes review, correction and exceptions. A saved hour only has financial value to the extent you can put it to use.
GPT-6.1 Sol Standard: US$2 / million input and US$10 / million output tokens. Above 272K input: US$4 / US$15 for the full request. No caching. Checked 8 Oct 2026.Official API pricing
Illustrative net monthly valueA$490
Realised time valueA$990 / month
Simple setup payback6.1 months
This includes the example monthly fee. It excludes tax, financing and other costs. A positive result does not validate the assumptions.
The defaults are generic planning inputs, not RSI AI pricing or a customer result. No model call or enquiry is made by this calculator. Invalid inputs are corrected to the displayed allowed range.
Bring a specific workflow.
Start with one recurring enquiry or handover, the person responsible for it and the point where the next action becomes difficult to decide.
EVIDENCEA secondary workflow
Check the handover.
Keep the source visible.
Explore a synthetic maintenance job pack as a related evidence-review workflow. Inspect a source, change a sample record and see how the local check responds.
The sample evidence workbench
Enable JavaScript to inspect synthetic cases and try the local review controls. The whitepaper explains the development thesis in text.
Local rules check synthetic records. No model is called, no customer file is uploaded and no external message is sent.