If your ERP already works, keep it.
The useful question is not “how do we add AI to the ERP?” It is: which recurring job around this ERP takes significant staff time?
Information might arrive by email, requiring somebody to find the right account, check several fields, prepare an update and request anything that is missing. The ERP already knows how to store the result. The repeated handling around it is the part an agent may be able to support.
Use the door that already exists
The cleanest route is usually an API: a documented way for software to read or change specific information. The quality and coverage of these APIs varies between systems.
Other routes exist:
- controlled browser operation
- scheduled imports and exports
- database access with strict permissions
- email and document tools around the edges
- a small connector built for one job
Whatever the route, the boundary matters. The agent should receive the minimum access required for the process rather than broad permissions added for convenience.
Read and prepare before write and send
A sensible first version might read the request, collect the relevant information and prepare the ERP update for a person to approve.
Once that works reliably, some low-risk actions may happen automatically. Exceptions should still stop for review, and expensive or irreversible actions should require explicit approval.
This is less ambitious than promising an autonomous company, but it avoids uncontrolled changes to important records.
The goal is to reduce repetitive handling around the existing system without making the wider business process less controlled.