Add AI.
Keep the tools that work.

Start with the flow of a case

Follow a request from its arrival to its closure. Who reads it? Where is it copied? Which identifier links it to a customer? At what point does a person rule on an exception? This observation shows where a model can be useful and where a simple rule is enough.

Fictional example: a request received by email must produce a record in the CRM. The AI extracts the need and prepares a category. The software remains responsible for the customer identity and the case number. A similarity between names is not enough to merge two companies.

Define the exchange contract

• Input: message format, usable documents, size and language.
• Output: expected fields, accepted values and the origin of the information.
• Access: read, prepare or write in each tool.
• Recovery: what to do after an error, a timeout or a dropped connection.
• Duplicates: recognize an operation already processed before replaying it.

A documented API, an existing connector or a validated export are different options. Their availability and limits must be examined in the software actually used. There is no universal connector that erases these constraints.

Get a complete slice working

A useful first workflow runs from the request to a record that can be viewed, with its source and status. It does not stop at a successful text in a chat window. Then test a missing document, an unknown customer, a case already received and a refused access.

Going live also depends on people: who handles the exceptions queue? Who can interrupt the integration? How is an ERP update prepared? These answers are part of the connection, just as much as the code.

What Saphan can prepare

A diagram of the exchanges, a prototype on an agreed environment, the test cases and a recovery procedure. The final scope depends on the access and the capabilities of your software. The tools already in place remain the reference wherever they do their job.