The Multi-AI Model Stack Is Here. Now Someone Has to Manage It
InformationWeek, Monday, September 14th, 2026
Put a gateway between applications and models before you scale, and separate standing routing policy from task-level choice.
At payments and data company Deluxe, no request goes straight to a model: a gateway routes it, choosing among GPT-5.6, Claude Opus, Claude Sonnet or models under evaluation depending on which of 50-plus agents is working.
CTDO Yogaraj Jayaprakasam's advice is to put that gateway between applications and models before you scale, not after, to avoid building governance and security controls model by model.
Grant Thornton's Sumeet Mahajan describes two decisions masquerading as one: the standing policy of which models are approved, how requests route and who pays, versus the local design decision of which model handles a given task.
In organizations doing this well a central platform team owns the routing layer while the business unit running the workflow makes the task-level call.