Agent projects rarely stall on the model

Agent projects rarely stall on the model. They stall the moment the agent has to write into the ERP the business runs on.

Reading is the easy half — reports, exports, a read-only view. No side effects, demo-ready in a week, and usually the half that got funded.

Writing is where the constraints live:

• referential integrity the agent does not own
• validation that fires server-side, after the call left
• postings you cannot amend, only reverse with another accounting event
• batch windows where writes queue instead of failing
• a retry against a non-idempotent endpoint = a duplicate order

Two things people skip:

Identity. The agent needs its own account with narrow permissions, not the shared service user whose password lives in three config files. "The integration posted it" is not an answer to an auditor.

The sandbox. Most ERP test environments do not mirror production master data. Your agent passes there, then meets duplicate customer records and half-migrated item codes on day one live.

None of this is model work. It is mappings, idempotency, permissions, error handling and reconciliation — and it is most of the budget. Put it in the estimate.

Full write-up: https://www.omarbaruzzo.it/en/blog/agente-ai-dentro-il-gestionale