rotascale
Sign in See it live

Solutions — sector

Two agents, one ceiling, one record.

Claims is where multi-agent pipelines arrive first and where delegation gets dangerous quietly: an orchestrator hands work to a specialist, the specialist hands work onward, and nobody is tracking what the tree can spend in total.

EU AI Act · GDPR · sector conduct rules National insurance supervisors clause map
Status
High-risk under Annex III where used in pricing or eligibility
Applies to
Claims, underwriting, fraud detection

The delegation problem, made concrete

An orchestrator holds authority for a claim. It delegates to an adjuster, which delegates to a payments step. Each is inside its own bound. The question nobody asks until afterwards is what the tree could spend, and the answer in most systems is: as much as you like, one grant at a time.

Attenuation is enforced
A delegated grant cannot exceed its parent on scope, window, budget, conditions or enforcement mode. Not by convention — the API refuses to issue one that does.
Spend debits every ancestor
So the root's ceiling is the tree's ceiling, and an exhausted ancestor refuses a child that still has room of its own.
Revocation reaches down
Revoking the orchestrator's grant revokes the adjuster's, and whether anything was allowed in between is answered rather than assumed.

Fairness, where the stakes are individual

A screen over declared attributes is not proof of fairness, and a flagged cohort is not proof of discrimination. Both are grounds for a controlled analysis somebody else performs. What the platform will not do is infer a protected attribute in order to produce a chart — a decision carrying no declaration is reported as undisclosed, and a cohort too small to measure is suppressed rather than given a rate it cannot support.

The agents your teams are about to ship

Claims is where multi-agent pipelines land first, because the work decomposes so neatly. That is also what makes it the place delegation goes wrong quietly.

  1. 1 authority Is there any authority for this?
  2. 2 status Is the grant active?
  3. 3 window Is now inside the grant's window?
  4. 4 scope Does the grant cover this action?
  5. 5 clean_context Was the context clean, where that is required?
  6. 6 bounds Do the per-action limits hold?
  7. 7 policy Does the policy on the grant permit it?
  8. 8 budget Is there room under the ceiling?
  9. 9 review Does this need a person?
Nine gates, evaluated in this order before the action. The lit ones are the gates that hold the four agents below — which is what differs between one industry and the next. The order does not.

One of these is already running, and here is what it recorded

claims-orch:CLM-4003:p2:20260813 claims-orchestrator · prod · dispatch → claims-adjuster allowed
  1. authoritypassed
  2. statuspassed
  3. windowpassed
  4. scopepassed
  5. contextpassed
  6. boundspassed
  7. policypassed
  8. budgetpassed
  9. reviewpassed

Not a refusal — a delegation, which is the thing this sector's pipelines do constantly and record almost never. The orchestrator handed claim CLM-4003 to claims-adjuster under a subdivision of its own authority, and both halves are in the ledger as separate governed trajectories that cite the same root grant.

The adjuster cannot exceed what the orchestrator held, and spend on the child debits the parent. The tree has a ceiling, not just the calls in it. observed →

Claims triage orchestrator budget

Takes first notification of loss, splits the claim across an adjuster, a fraud check and a payments step, and closes what it can.

Consequential action
Authorises a settlement payment
What goes wrong
Each delegate is inside its own limit and the tree as a whole has none. The question nobody asks until afterwards is what the whole pipeline could pay out in an afternoon.
What is recorded
Every debit against every ancestor, so the root ceiling is the tree ceiling and an exhausted parent refuses a child with room to spare.
Underwriting agent authority

Prices and binds standard risks up to a threshold, referring anything unusual to a human underwriter.

Consequential action
Binds cover — a contractual commitment
What goes wrong
Pricing and eligibility are Annex III high-risk. Binding cover is a legal act, and an agent that can bind without a traceable delegation has created an obligation nobody authorised.
What is recorded
The grant it bound under, its expiry, and the human in whose name it acted.
Fraud detection agent policy

Scores claims against behavioural patterns and refers suspected fraud for investigation.

Consequential action
Refers a policyholder for investigation
What goes wrong
A referral is consequential for the person referred. A model that leans on a proxy for a protected attribute produces a discriminatory outcome nobody chose and nobody can see.
What is recorded
Cohort outcomes where an attribute was declared, undisclosed where it was not, and suppression rather than a rate a small cohort cannot support.
Renewal and retention agent window

Reprices books at renewal and applies retention offers inside a mandate.

Consequential action
Changes a premium on a live policy
What goes wrong
A mandate that outlives the campaign it was written for. Six months on, the agent is still repricing to a rule nobody remembers approving.
What is recorded
The window the authority was valid for, and the refusal of every call outside it.

Ambition is the point of these: none of them is a chatbot. Each is an agent taking an action with a consequence somebody has to answer for — which is exactly the moment a bounded authority stops being paperwork and starts being the reason the project is allowed to ship.

What your conduct regulator will ask

Not hypothetical questions. These are the ones that arrive in writing, with a deadline, and the honest answer to most of them is a query rather than a project.

Can you show the customer outcome, not the model metric?
The record is per decision and per policyholder, not per model run. What a claimant was told, and when, is in the trajectory rather than reconstructed from logs.
How is delegated authority controlled?
Attenuation is enforced by the API, not by convention: a delegated grant cannot exceed its parent on scope, window, budget, conditions or enforcement mode. The API refuses to issue one that does.
If you revoke authority, what happens to work in flight?
Revoking a parent revokes its children, and whether anything was allowed in the interval is answered rather than assumed.
How do you evidence fairness without inferring protected attributes?
We do not infer them. A decision carrying no declaration is reported as undisclosed, and a cohort too small to measure is suppressed rather than given a rate it cannot support. Both are grounds for an analysis somebody qualified performs.

Identifiers this deployment verifies here

Every one of these is a checksum, not a shape. The claim is “this is a valid national identifier and the check digits agree”, not “this looks like one” — which is why the count on a screen can be cited rather than caveated.

FamilyHow it is verified
Bank account (IBAN) — mod-97 over the rearranged string
National identifier (PL, BE, NO) — scheme-specific weighted checksum
US Social Security Number — structural rules only — asserted, not observed
Australian Medicare number — weighted check digit
EU VAT number — per-state check digit where the state publishes one

And what it does not look for. Names, addresses and dates of birth in free text carry no checksum, and finding them needs a model this deployment does not run. Neither are card numbers, medical record numbers, or national schemes beyond the three above. A clean screen means no verified identifier was found — never that the payload holds no personal data.

Written for this sector

20 posts touch this sector — the five most recent are above.

See it on your own problem

Bring your claims pipeline, orchestrator and all. We will delegate authority down the tree in front of you, exhaust the root ceiling, and show you the child refused while it still had room of its own.

A walkthrough is a working deployment with your claims and conduct leads in the room, not a slide deck — bring the agent you are most nervous about and we will put it behind a grant while you watch. If you would rather look first, the demo needs a work address and one click.

Insurance walkthrough See the live demoOr talk to sales →