An agent is only as useful as the systems it can reach. That's also where the enterprise value appears: working with ERP, CRM and warehouse data.
01. Not Straight to the Database
Don't connect an agent directly to the ERP database. Put a service layer in between that defines which operations are possible. That layer enforces both permissions and business rules.
02. Define Tools Narrowly
Instead of a "run SQL" tool, define narrow purposeful tools like "get the customer's open orders". A narrow tool is both safer and easier for the model to use correctly.
03. Start with Reads
In the first version give only read tools: order status, stock lookup, account balance. Value appears quickly and risk stays near zero.
04. The Write Step
When writes open up, every operation should pass approval and be idempotent: running the same request twice must not create two records. That protection matters during network interruptions.
05. Data Freshness
How current the agent's data is must be clear. An agent fed by a reporting database copied overnight answers stock questions wrongly.
06. A Test Environment
Don't develop an agent against the live ERP. A test environment full of realistic data reduces the cost of the mistaken calls that are inevitable in the first weeks to zero.