How it works | InsightAdvocate

The architecture

From the claim record to the decision record

1

Claim record

Policies and clauses, carrier letters read into positions, the timeline, the loss schedule and requests.

2

Six engines

Notice, coverage, other insurance, SLA, settlement and loss trend. Plain data, every figure cited.

3AI

Ten agents

On the shared insight_agents pipeline, each reasoning over the engines’ numbered facts.

4

Grounding check

A number not in the facts throws the answer away. Fact markers are stripped before anyone reads it.

5

A person decides

Prepared decisions at L1 or L2 wait for the advocate or the practice leader.

6

Decision record

insight_decisions: the run, every claim control that fired, every answer, signed.

Alongside: the engines’ fields feed the claim controls in Rules Studio, which write to the same record; the copilot reads the record through read-only tools.

1 · The engines

Every figure and every assessment, computed and cited

The engines read the claim’s rows and return plain data, each figure with a citation to the wording, the letter, the timeline entry or the loss schedule it rests on. They are tested on known answers.

notice

Per policy: the notice condition, the deadline from the day the client knew, sent on time, late or overdue.

coverage

Each carrier position mapped to the wording: supported, weak, not in the policy, needs facts; grants not addressed.

other insurance

Other lines the facts bring in, excess layers the exposure reaches, the prior claims-made year.

sla

Service-agreement clocks and illustrative regulatory timelines from dated events; the escalation ladder.

settlement

Claimed, accepted, disputed; each disputed amount weighted by its assessment; net of retention; the floor.

loss trend & scorecard

A client’s loss runs by year for the renewal; carriers’ claims handling across the book.

2 · The agents and the grounding check

The model explains what the engines computed. It never makes a number

Analyze runs ten agents in order on the shared pipeline, each as itself at its autonomy level. Each reasons over numbered facts with Claude Haiku 5.5 on Amazon Bedrock; an answer carrying a number not in the facts is thrown away and the engine’s findings stand. Fact markers never reach a reader.

sequenceDiagram
  participant U as Advocate
  participant P as Pipeline
  participant E as Engines
  participant M as Model
  participant D as Decision record
  U->>P: Analyze
  loop ten agents in order
    P->>E: Facts for this agent
    E-->>P: Numbered facts with citations
    P->>M: Reason over these facts
    M-->>P: Reading
    P->>P: Grounding check, strip markers
    P-->>U: Prepared decision at L1 or L2
  end
  P->>D: Control hits, the run and its steps
  U->>D: Accept, reject or ask, signed

The last Analyze on an illustrative claim: each of the ten agents with its finding, its level and whether it prepared a decision, all from the engines with the model off — **One run.** Each agent’s finding, level and outcome. With the model off (as here), every agent falls back to the engines and says so.

3 · The autonomy dial

Capped at ask first

The level for an action is the lowest of the agent’s level, the action’s cap and the tenant ceiling (ADVOCATE_AGENTS_MAX_LEVEL). Every action in InsightAdvocate is capped at L2.

L1

Draft

The agent drafts; a person does the work. Where every agent starts.

CeilingL2

Ask first

The agent prepares the decision; a person accepts, rejects or asks.

L3

Act and tell

Locked. A letter to a carrier or client is never sent by an agent.

L4

Act

Locked.

Promotion is earned: 90% agreement over 50 answers recommends a notch, which a practice leader accepts; below 70% over 20 moves it down at once.

4 · Rules Studio and the decision record

Practice standards as claim controls, on the record

Eight claim controls ship: notice overdue (block), notice due in seven days, unsupported positions, excess reach (warn), large denials, SLA breaches, settlement gaps and large losses (refer to a lead). Each is a sentence over the engines’ fields, tested against open claims before it is published, versioned, and written to the decision record when it fires.

Rules Studio on the illustrative book: the claim controls that fired most in 30 days, each a readable sentence with its outcome — **Rules Studio.** The claim controls that fired most in 30 days, each a sentence over the engines’ fields with what it does: block, warn or refer to a lead.

5 · The platform

Built on the shared Insight platform

insight_agents

The agent pipeline, registry, autonomy service, run store and traces.

insight_decisions

The decision record: control hits, answers and sign-offs.

insight_rules · insight_broker

Rules Studio and claim controls; the broker roster.

insight_diligence

The agents’ model kit: the findings schema, citation and grounding helpers.

insight_assist

The claims copilot (Claude Sonnet 5.5): read-only tools, figure checks.

insight_home · insight_identity

Home per persona; sign-in, roles and tenancy.

Regulatory timelines in the SLA engine are illustrative model timelines (configurable), not any state’s law and not legal advice.

Engines compute. Agents prepare. People decide.