Appendix
OnCase · Supervised Referral Agent
Designing a supervised referral agent
Defining what an AI agent may prepare, what a caseworker must decide, and how the system stops safely when the evidence is incomplete.
Prototype boundary: This is an interaction model demonstrated with a deterministic simulation, a fictional approved directory, and a fictional inbox. It is not a live AI agent, does not search the open web, does not send email, and does not use client data. The value being demonstrated is the supervision model, not production automation.
How to read the evidence
Every claim in this appendix is labeled with its evidence level. The three labels used are:
Research support
Existing project evidence supports the underlying referral problem or risk. It does not validate this interface.
Prototype demonstration
The interaction and state boundary exist in the working fictional prototype.
Production requirement
Necessary for a real service, but not implemented or implied by the prototype.
Why an agent, not a form?
A fixed workflow can enforce consent, required fields, approval gates, and record transitions. It cannot reliably handle the variable work between those boundaries: comparing changing resource information, drafting outreach for a specific case, linking replies to referrals, or interpreting ambiguous provider language. OnCase uses AI for that variable work while keeping consequential decisions under deterministic rules and human control.
AI-assisted work
Find plausible resources from approved sources
Compare eligibility and freshness signals
Draft editable outreach
Link and interpret provider replies
Flag uncertainty for review
Rule-bound work
Verify consent
Require approval before sending
Preserve worker edits
Block invalid status transitions
Keep silence as no response
Require confirmation before changing case state
Preserve an audit trail
Who acts, who decides
Actor
Step
Worker
Confirms consent
Agent
Gathers eligible matches with source and freshness information
Worker
Reviews and selects recipients
Agent
Prepares an editable draft
Worker
Edits and approves sending
Agent
Monitors and proposes the meaning of replies
Worker
Corrects or confirms the interpretation
System
Records only the confirmed outcome
At any approval, evidence, or interpretation failure, the agent stops without sending or changing the record.
State and decision flow
Consent
→
Resource review
→
Draft and approval
→
Reply review
→
Delivered or unresolved
Decision branches
No consent: stop before gathering or drafting.
No eligible resource: explain the result and offer a permitted next path.
Send failure: keep the draft and show that nothing was sent.
Uncertain reply: show the source message and require worker interpretation.
No reply: remain at “No response.”
Provider can screen: use “Intake or screening required,” not accepted.
No delivered service: close only as unresolved with a required reason.
OnCase proposes the meaning of each provider reply but does not change referral state. The worker reviews the original message, corrects uncertain interpretations, and confirms the outcome. Silence remains no response.
Failure and recovery
Stale directory information
Send failure
Uncertain reply meaning
No provider reply
Accepted referral but service not delivered
Agent experience brief
Primary user
A community-based case manager coordinating a referral across consent, resource search, outreach, provider response, and follow-through.
User job
Move a referral toward a delivered service or a documented unresolved outcome without reconstructing state across a directory, email, notes, and memory.
Agent goal
Reduce the coordination work inside one referral by preparing resource matches, outreach drafts, and reply interpretations while preserving the worker’s authority over consequential actions.
Agent non-goals
Decide service eligibility, rank clients, promise provider capacity, choose a provider on the worker’s behalf, send without approval, treat silence as rejection, or update the official record.
Trust model
Suggestion → evidence → uncertainty → human edit or confirmation → activity history.
Autonomy matrix
Each action in the referral flow is either automatic (the agent may proceed without a gate) or requires explicit worker confirmation. Worker-confirmed gates are shown in bold.
Action
Authority
Why
Filter the approved directory
Automatic
Matching can reduce search work without claiming availability.
Choose recipients
Worker only
A match is not suitability, capacity, or a recommendation.
Send outreach
Worker only
Sending discloses information and creates an external commitment.
Change provider outcome
Worker only
A parsed message does not become case state on sight.
Convert silence into decline
Prohibited
Missing evidence cannot become a negative fact.
Communication specification
Interface copy must preserve the precise meaning at each moment. These are the anchors.
Moment
Meaning the copy must preserve
Matches found
These are candidates, not recommendations or confirmed openings.
Drafts prepared
Preparation is not action. Nothing sends until you approve.
Low confidence reply
Confidence describes interpretation certainty, not whether the reply is good or bad. Confirm the meaning, not just the message.
Silence
Absence of a reply is not a decline. No response remains no response.
Evaluation slice
Task: Ask a participant to complete Ray’s legal-aid referral using only the fictional prototype. Do not point out the planted error.
Pass signal: The participant reviews at least one source or freshness signal, understands that match does not mean capacity, knowingly approves the send, corrects the Suffolk interpretation to “Intake or screening required,” and leaves silence unchanged.
Fail signal: The participant believes the directory confirms availability, thinks drafts were already sent, accepts the wrong interpretation without reading the message, or turns no response into “provider unavailable.”
What this evaluation cannot establish yet: Real time saved, referral completion, provider response rates, client outcomes, privacy performance, model accuracy, or adoption inside an organization. Those require a production-capable pilot and real case-manager participation.
Portfolio evidence map
What a reader can inspect and at what evidence level:
What you can inspect
Evidence level
Working consent, resource review, draft, approval, reply, correction, confirmation, and unresolved-outcome interactions
Prototype demonstration
103 automated checks for state boundaries, correction paths, accessibility, contrast, and copy
Prototype quality evidence
Real email, live directory search, case-record integration, permissions, retention, and outcome measurement
Production requirements, not built
What this demonstrates
This prototype demonstrates an operating model for supervised agent behavior: visible evidence, limited authority, meaningful approval, correctable interpretation, safe failure, and an audit-ready path from suggestion to confirmed state. It does not establish production safety, service effectiveness, or user adoption. Those require real caseworkers, live operational constraints, privacy and legal review, and measured field use.