Ten claims-advocacy agents, inside authority

A claim control blocking because a notice deadline passed unsent, above the notice checker's prepared notices — A kitchen fire at intake (illustrative): the block banner for a passed notice deadline, and the notice checker’s prepared decision — send notice to two carriers, one overdue — with Accept, Reject and Ask.

InsightAdvocate is AI-first: ten agents work every client claim. They are not chatbots and they are not autonomous. Each reads the facts the engines computed, reasons over them with the model, and prepares one kind of decision for a person.

The agent catalogue's coverage and carrier step: four agents, each with its action and what it never invents — The agent catalogue, coverage & carrier step: the coverage position analyst, proof-of-loss coordinator, rebuttal drafter and carrier SLA monitor, each with the action it may prepare.

The ten, in the order they run

Agent	What it reads	What it prepares
Notice checker	Each policy’s notice condition and deadline from the wording; what is late or due	The notices to the carriers
Coverage position analyst	Each position in the carrier’s letters against the wording; grants not addressed	The assessments (supported, weak, not in the policy, needs facts)
Other-insurance finder	Other lines the facts bring in, excess layers the exposure reaches, the prior claims-made year	Putting the claim to those policies
Carrier SLA monitor	Service-agreement clocks and illustrative regulatory timelines	The escalation up the carrier’s ladder
Proof-of-loss coordinator	The proof-of-loss checklist and document requests	The request to the client
Rebuttal drafter	The weak and unsupported positions and the engine’s reasons	The rebuttal letter, quoting the wording
Settlement evaluator	Claimed, accepted, disputed, the weights, the retention	The counter, the floor and the recommendation
Client update writer	The claim’s stage and figures	The client’s status update
Loss-trend analyst	The client’s loss runs for the renewal	The brief for the placement team
Supervisor agent	The claim controls; the decision record	The practice leader’s sign-off when a control fires

How one agent step works

flowchart LR F["Engine facts,
numbered and cited"] --> M["Model reasons
over the facts"] M --> G{"Every number
in the facts?"} G -- "yes" --> R["Agent's reading
markers stripped"] G -- "no" --> E["Discarded:
engine findings stand"] R --> P["Prepared decision
at its level"] E --> P P --> H{"A person:
Accept, Reject, Ask"}

Every step is recorded with the model it used, tokens, cost and a trace. With the model switched off — as on this site’s screenshots — each agent falls back to its engine findings and says so on the card (“From the engines”); cards reasoned with the model say “Reasoned with the model”. The agents run on Claude Haiku 5.5 by default through Amazon Bedrock.

The last Analyze: ten agents, each with its finding, level and outcome, on illustrative data — The last Analyze on the illustrative claim: each of the ten agents with what it found, its level (L1), that the engines produced it, and whether it prepared a decision.

What none of them may do

Invent a figure. Deadlines, assessments, floors and loss ratios are the engines’.
Invent a clause or an argument. The rebuttal quotes wording; each point is the engine’s reason.
Send anything. Every action is capped at ask first (L2). What goes to a carrier or a client is a person’s call, and today it goes as a dry run.