Most enterprise AI has to coexist, not replace
A lot of AI discourse assumes a green-field starting point — a new system, built cleanly, with AI native to the architecture. Most real deployments don't look like that. They have to work alongside an ERP that's been customized for a decade, a CRM full of institutional workarounds, and internal tools nobody wants to touch because nobody fully understands them anymore.
Why "just integrate it" undersells the work
Connecting to a legacy system is rarely just an API call. It usually means understanding data that was never structured with this use case in mind, working around systems with limited or fragile integration points, and being careful not to introduce a new failure mode into a system that, whatever its flaws, currently works.
What tends to work
The integrations that hold up long-term usually start narrow — one well-scoped read or write path into the legacy system, rather than a broad integration touching everything at once. That narrow scope makes it possible to validate the connection is reliable before expanding it, and it limits the blast radius if something needs to be rolled back.
It also helps to treat the legacy system as the source of truth it already is, rather than trying to migrate its role to the new AI layer prematurely. The AI system earns a larger role over time, as it proves it can be trusted with one.
The real constraint
In most engagements, the technical integration turns out to be more tractable than the organizational question of who owns the legacy system and how change requests to it get approved. Planning for that process early tends to save more time than optimizing the integration architecture itself.