An agentic ERP is an AI layer built on top of a company’s ERP: it reads the system’s data, shows what is missing or inconsistent, and proposes actions. It starts read-only because a wrong write to the books is expensive and hard to undo, while a slow answer costs almost nothing. On an existing ERP, writing comes later and goes through a person’s approval.
Start by knowing what the ERP actually holds
In a real estate firm’s case, the whole operation lived in a property-management ERP with thousands of undocumented API services. Before any automation, KINTO inventoried them, and the work was done read-only: leases and properties in the ERP were cross-checked against documents kept across email, shared folders and scans, and against the old website’s catalog. The result was a verifiable inventory of leases, properties and their documents. Thousands of endpoints were inventoried and curated for agents (measured).
What the agent layer does
- Reads by default. Operations are curated for agents, and error messages suggest the next step.
- Shows the status. A close-the-books dashboard shows, lease by lease, the status of each document.
- Writes only through approval. ERP writes go through a queue a person approves, with a dry run, a read-back and duplicate protection.
- Answers in plain language. Management can ask the ERP in plain language instead of navigating menus.
In that case, the close runs through a dashboard with an approval queue, with the ERP and the ledger matching document by document.
When the old system has to go
Sometimes the right move is to replace the system. In a logistics operator’s case, KINTO replaced a two-decade-old application and a macro-driven spreadsheet. The condition for turning the old one off was proof: six months of billing periods matched the old system to the cent on 100% of shared waybills, and the remaining differences were explained one by one (measured). The new system also made customer-reactivation opportunities visible. The guiding principle was “autonomous by default, human by exception”: the system handles the routine and calls a person only when something falls outside what is expected.
A read-only layer is useful even when a project stops
In a food manufacturer’s project, a read-only layer let agents query real ERP data without being able to write. The project did not move forward because of access restrictions, and the layer remains a reusable asset.
What we do not know yet
The accounting hours saved in the real estate case have not been timed. In the logistics case, a pay-for-results pricing model was analyzed, with no agreement closed to date.
To see how this is scoped, read Agentic ERP and Migrations and rescue.
