The agent that audits the agents

Allianz Project Nemo

In July 2025, Allianz Australia put the group's first agentic AI to work on storm-season claims: seven agents that verify, screen, and calculate in minutes. The seventh exists only to audit the other six. And none of them may pay a cent. By design, payout decisions are never automated.

Worked case · public-source reconstruction Autonomy: prepare

Why this case matters

The boundary came before the build.

A weaker version of the story would be "insurer automates claims." The stronger reading is the order of decisions: a global insurer that could have automated payouts chose a deliberately tiny scope, placed an auditor inside the agent system, and put its never-do in the press release. Allianz says the bounded case settles claims up to 80 percent faster. The design is what made that speed safe to buy.

Live
July 2025, Allianz Australia
Architecture
Seven specialised agents
Scope
Food-spoilage claims under AUD 500
The never-do
Payouts are never automated

The case

Six agents do the work. The seventh checks it.

A storm knocks out the power, a freezer full of food spoils, and a small claim lands in a queue behind thousands of others. It is exactly the kind of claim that used to take days: low value, high volume, and arriving in bulk at the worst moment. Project Nemo hands it to a pipeline. A planner agent orchestrates the work; a cyber agent checks data security; a coverage agent verifies the policy; a weather agent confirms, against external meteorological data, that the storm actually happened; a fraud agent screens for suspicious patterns; a payout agent calculates the settlement.

Then comes the part that makes this a design story. A seventh agent, the audit agent, reviews every step the other six took, writes a summary of all their decisions, and passes it to a human claims professional, who makes the payment decision. "The ultimate responsibility always rests with a claims professional," says Maria Janssen, Chief Transformation Officer at Allianz Services. The AI phase completes in under five minutes; Allianz says settlement times fell from days to hours, and that the whole system was built in under one hundred days.

The scope is the other half of the design. One claim type, one peril, one ceiling: food spoilage, severe-weather outages, under five hundred Australian dollars. Expansion to travel delays, simple motor claims, and property assessments comes only after the bounded case proves out. The model is not the story. The boundary is.

Why it matters

Oversight you can draw on an architecture diagram.

Most agent governance lives in policy documents that the workflow can quietly skip under deadline. Nemo's governance lives in the system itself: one agent's entire job is reviewing the others, and the one action that touches money is withheld from every agent. The never-do is not a line in a handbook; it is an absence in the architecture.

Read through the Agent Operating Model, the design is unusually legible. Cognition comes from the models inside each agent. Control comes from the audit agent and the human payout gate. Reach stops, deliberately, one step short of the money. Its power comes from the combination, and so does its safety.

The filled canvas

Turning the story into nine design decisions.

The entries below are intentionally compact. A canvas should not read like a requirements document; it should make the load-bearing choices visible enough for a team to argue about them.

Worked example

Allianz Project Nemo

Agentic AI Design Canvas
designtheagent.com

North Star

What business outcome should this agent's work ultimately contribute to?

Settle simple storm-season claims in hours instead of days for customers in distress, while keeping every payout decision in human hands.

Target Workflow & Agent Role

Which workflow is it part of, and what contribution is the agent responsible for making within it?

Workflow: food-spoilage claims under AUD 500 arising from severe-weather power outages (Allianz Australia). Role: verify, screen and cost each claim (a seven-agent pipeline) so a human assessor can settle it the same day.

01

Users & Stakeholders

Who uses it directly, and who else is affected by what it does?

Direct users: claims professionals who receive the assembled file and decide. Affected: storm-hit customers awaiting payment, claims operations, fraud teams, regulators, and reinsurers.

02

Success Measures & Standards

What measures will show that the workflow improved, and what performance standards must the agent meet?

Measures: settlement time (days to hours) and claims closed within a day. Standards: AI phase under five minutes, every claim weather-corroborated and fully auditable.

03

Autonomy by Action

For each consequential action in its role, how independently may the agent act?

Autonomy: prepare. It verifies, screens and costs the claim, then hands the assessor a settlement recommendation. Paying it is never automated, by design.

04

Tools & Action Channels

Which tools, systems or channels may it use to get information, make changes, communicate or take action?

Reads policy and claims systems; writes a structured audit summary for the human decision-maker. The one action that touches money, paying the claim, is withheld from every agent.

05

Rules & Boundaries

What must it always do, what must it never do, and when must it stop, escalate or hand off?

Never automate a payout. Stay inside the bounded scope: food spoilage, severe-weather outages, under AUD 500. A cyber agent enforces data-security guardrails on every claim.

06

Context & Knowledge

What information may it use, and which sources should take priority when they disagree?

Policy records and coverage terms, the customer's claim and documentation, transaction patterns for fraud screening, and external meteorological data confirming the storm actually happened.

07

Memory & Learning

What should it remember across interactions, what must it forget, and how should it learn and improve over time?

Learning happens by widening scope, not by silent drift: prove the bounded case, then extend to travel delay, motor, and property claims. Retention is not publicly specified.

08

Ownership & Oversight

Who owns the workflow outcome, who is accountable for how the agent operates, and how will it be reviewed?

An audit agent reviews every step the other six take and writes the summary; a claims professional makes every payment decision. Ultimate responsibility rests with that person, not the pipeline.

09
Job · what it's for Authority · what it may do Knowledge · what it may know Accountability · who answers

What to notice

Four lessons travel beyond Allianz.

Cell 06

The never-do came first.

"By design, payout decisions are never automated" is a sentence from the company's own announcement, not a compliance gloss added later. The boundary preceded the build.

Cell 09

Oversight is architecture, not policy.

The seventh agent exists only to audit the other six and package their work for the person who decides. A control that is part of the system cannot be skipped under deadline.

Cell 01

Scope is a design tool.

Under AUD 500, one claim type, one peril. Small enough to prove safely, frequent enough to matter: storm season delivers these claims in bulk, to customers in distress.

Cells 04 & 08

Reach is earned, not assumed.

Travel delays, simple motor claims, and property assessments come only after the bounded case proves out. The expansion path is the learning loop, stated in advance.

Coach audit

We ran this canvas through the coach.

Not as an outside fact-checker, but as a claims-transformation lead asking the question that matters before the scope widens: does the design that made the bounded case safe survive its own expansion? Here is what came back.

Coach review · round one

Minor gaps

What holds together

  • One consistent story at prepare autonomy: a narrow claim type, a hard human payout gate, and the money-moving action withheld from every agent.
  • The scope is deliberately bounded: food spoilage, severe-weather outages, under AUD 500.
  • Every claim is corroborated against external weather data, and every step is traceable enough to support an audit summary a human can act on.

What pulls apart

  • Success Measures & Standards (Cell 03). The measures are speed and volume only. Nothing shows the recommendations are correct, so there is no bar for the quality the North Star depends on.
  • Ownership & Oversight (Cell 09). The check on how the pipeline operates is another agent, not a named person who can change its standards or stop it. That is a weak point precisely where the scope is due to expand.
  • Memory & Learning (Cell 08). What claim data is kept, and for how long, is unspecified, which becomes a live exposure as the pipeline moves into new countries with different privacy rules.
  • Rules & Boundaries (Cell 06). There is a firm never-automate-payout rule but no clear point at which the agent must stop and flag a claim for a person, for example when fraud patterns look off.
“Who is the person who can inspect the pipeline's work, tighten its standards, or switch it off, and is that a different person from the assessor deciding each claim?”The coach, on Ownership & Oversight (Cell 09)

Sources

What we can say publicly.

Only public-source claims are treated as facts. The speed and build-time figures are Allianz's own; the underlying platform is undisclosed. Internal thresholds and routing rules should be read as design implications unless Allianz confirms them.

More worked cases