Play 2: Connect the Systems of Record

Play 2: Connect the Systems of Record

Goal: the facts that matter flow onto the model from the systems that already hold them.

Connect, don't migrate

The headless enterprise doesn't replace your systems of record. The ERP stays the source for the ledger, the HR system for people, the work tools for the work, the cloud bills for cloud spend. The model connects them and adds what none of them hold: how they relate, and what they mean for the business.

How to run it

• Inventory first. List the systems that hold the facts your first question needs. For each, note who owns it, how it can be reached, and whether it already exposes itself to AI through MCP.
• Map to the model, not to each other. Point each system at the model's objects and fields rather than building system-to-system integrations. Every new connection then benefits every agent and every surface.
• Connect in conversation where you can. AI can now find a system's API, draft and test the requests, and propose where the data belongs in the model. Keep sign-in and credentials out of the conversation, and test on live data before anything is saved.
• Keep the source visible. Every value on the model should link back to where it came from, so anyone — or any AI — can check it.
• Reconcile money to the ledger. Wherever cost appears on the model, make sure it ties back to what finance has booked. Numbers that don't reconcile are worse than no numbers.

Outputs

• The first domain's key systems connected to the model.
• A map of which system is the source for which facts.
• Reconciliation of any money on the model to the books.

Signs it's working

• Nobody is re-keying data from one system into another to prepare a review.
• When someone questions a number, the trace back to its source takes one step.

Pitfalls

• Building point-to-point integrations. They multiply, break and never add meaning.
• Treating freshness as binary. Some facts need to be current to the hour, many to the week. Decide per fact.
• Hiding gaps. Missing or unmapped data should be shown and named, not quietly spread across everything else.