SauronIDRocket Money published by Anthropic
Scale of that deploymentPersonal finance app, phased rollout
Accord explicite required before any financial action
- The agent cancels subscriptions and negotiates bills, so it takes acts with consequences
- A classifier routes each request to a specialised subagent
- Routine tasks are frozen into deterministic code, not left to the model
What stays human, thereThe call on every financial action, and what happens when one fails.
Read the sourceWhat it costs today100 %of lines checked
Assumptions behind this figureSample-based checking at 8% today
SetupQuick entryTwo to three weeks of scoping, a single connection, a scope that fits on one page.
What it costs, three ways
Today€21,300
A person opens every file
With a copilot€17,000
A person still opens every file, faster
With a bounded agent€4,000
A person opens one file in four
- 75% handled alone
- 10% prepared, decided by a person
- 15% refused by the rule, picked up by a person
Annual gap between today and the agent€17,200
Assumptions behind this figureSample-based checking at 8% todayCopilot gain taken at +20%, the top of the range in the controlled studies.
Modelled from the assumptions shown, not a client result. The 75 / 10 / 15 split is a working assumption: at scoping it gets measured on your own files.
What the agent does, on one case
- 01Opens every claim and every attached receipt
- 02Tests each line against internal policy
- 03Approves what passes, under the agreed threshold
- 04Escalates the rest to the manager, with the reason
The indicators we measure before and after
- Share of lines actually checked
- Policy breaches caught
- Reimbursement delay
What we usually connect to
- The expense tool
- The internal policy, as written
- Payroll or accounting
The act that commits
A reimbursement is approved.
The rule that bounds itauto-approved below the threshold, above it the manager decides
What stays with your team
The justified exception, the conversation with the employee, changing the policy itself.
Your figures are not these ones.
Give us yours in three minutes. We redo the arithmetic on your volume and your loaded cost.