Skip to content
HONESTASDecision-Evidence Operating System

For the Chief Underwriting Officer

A submission went straight through without a human. Who decided it did not need one, and where is that written down?

Routing is the first decision anyone makes about a case, and in most operations it is the decision nobody keeps. The score that chose the automated lane, the room the case would otherwise have been argued in, the level of autonomy it was granted — all of that happens before the case is ever worked, and by the time an auditor asks, none of it survives.

The routing decision, rendered as the event it already is

Convening decisionLoggedconvening.tier.assigned
  1. T1InlineA model answers inside the request.in production(shipped)
  2. T2Callback with your recordsWork that needs your data runs against it and calls back.Assigned to this casein production(shipped)
  3. T3AcceleratedSpecialised compute — the same path, a different machine.designed, not built(designed, not built)
  4. T4Long-runningScenario runs and large ingests: durable, resumable, watchable.designed, not built(designed, not built)
  5. T5VerifiedNothing releases until a second model agrees and the record seals.designed, not built(designed, not built)
basis
per-field confidence · coverage, liability and litigation scores · live appetite state
room
claims.coverage-committee
chair
a named human, assigned at convening time
scope
tenant · owner · business unit
ladder
tier chosen, and the choice recorded as its own event
on failure
the lane degrades to a human queue, loudly — it never drops the case

[O] Illustrative composition, built from the components the product ships. No customer data appears anywhere on this site. The dispatch path, the human-approval waitpoint and the interaction log are observed at NX/services/nexus-orchestrator/src/routes/dispatch-routes.ts and NX/services/nexus-workflows/src/services/governance.ts.

The mechanism, in three lines

A meta-decision agent scores each submission or claim and convenes the right room, the right chair and the right autonomy tier — and the routing decision is itself a logged event.

First

It scores before it routes

Per-field confidence on the submission, coverage and liability and litigation scores on the claim, and the live accumulation state of your portfolio are read before any lane is chosen.

Then

It convenes, rather than queues

The room, the seats, the named chair and the autonomy tier are chosen together. A case that needs argument gets a room; a case that does not still passes the gate.

Finally

The routing is an event

The score, the tier, the chair and the reason are written to the same governed thread as the decision they precede — so the question “why did nobody look at this?” has an answer with a timestamp.

shipped(shipped)Tiers one and two carry live traffic today. The rest of the ladder is drawn, not running.

2 of 5rungs of the dispatch ladder running in production today[O]◆
3axes every routed case is scoped on: tenant, owner, business unit[O]◆
~2/3of surveyed European undertakings already use generative AI, most still at proof of concept[V]
50%of non-life insurers already use AI somewhere in the value chain[V]

The honest limit

Change Control & A2ABook a 20-min demoDeliberation Chamber