Case Study: Governing Robin AI, Octupus's ERP Copilot for Odoo
Octupus Technologies builds Robin AI, a copilot that works inside its clients' Odoo ERP: it answers questions about customers, drafts quotes, reads invoices and summarises sales performance. Every one of those tasks means sending business records to a language model.
The problem: ERP data is not just personal data
The first analysis of Robin AI traffic showed the classic identifiers — names, emails, tax IDs, IBANs — but also something most PII filters ignore: commercially sensitive figures. Revenue, margins, contract values and price lists appear in almost every ERP prompt, and for a client they are as confidential as a customer's address.
What changed in Privaro
- Money detection: amounts with symbols (€, $, £) and currency codes (EUR, USD, GBP) are now detected and handled by policy, in the proxy and in the public demo
- Agent pipelines: each Robin AI deployment runs on its own pipeline with policies per client, so one ERP tenant never shares configuration or tokens with another
- Output scanning: answers are checked before they return to Odoo, catching data the agent pulled via tools rather than from the prompt
How the flow works
Robin AI sends the request to Privaro instead of the provider. Privaro detects entities, tokenizes them against the client's vault, forwards the sanitized prompt, and restores tokens in the response only for the authorised user. Every step is written to the audit log with the agent run it belongs to.
The result
"Governing what those agents send to a model is not optional — it is the first requirement of any deployment," says Octupus's CEO. Instead of rebuilding masking, key management and audit in every client project, Octupus consumes it as infrastructure, and each deployment ships with the evidence its client's DPO will ask for.
Next step
Building agents on top of business systems? See Privaro for AI agents or try detection yourself in the live demo.