70% of Tech Leaders Say Business Users Are Outpacing IT on AI Agents. BCG Says Stop Trying to Stop Them.

Sellers are building AI agents directly inside their CRM, and IT often finds out after the fact. BCG’s Sakshit Teotia, Rajeev Singh, Matt Emara, Karan Gupta, and Srikanth Srinivasan report that 70% of technology leaders say their teams deploy AI solutions faster than IT can track, over half of sellers have already used AI agents on the job, and 94% of sales leaders using agents now consider them critical to hitting business targets. BCG’s read is unusual for a governance piece: don’t try to centralize or block this. Shut it down and the shadow-AI activity doesn’t stop — it just stops being visible.

The four risks BCG names are concrete rather than abstract: customer record fragmentation as agents write back to the CRM inconsistently, uncontrolled agent decision-making, deepening vendor lock-in as business teams commit to whatever platform ships an agent-builder first, and unpredictable consumption costs under new SaaS pricing models built around agent usage rather than seats.

That vendor-lock-in risk is already playing out at scale. Salesforce and Anthropic’s Claudeforce announcement shows Salesforce committing roughly $300 million to Anthropic tokens this year, on top of an existing $5 billion stake — a single-vendor pairing BCG’s “controlled orchestration layer” recommendation is implicitly written to prevent enterprises from getting locked into by default. And the operational shape of the problem BCG describes — agents from multiple vendors, governance logic buried inside individual agents instead of a shared layer — is exactly what Forrester calls “agent sprawl”, with Forrester warning flatly: “those that choose a runtime architecture for their agents will win this one” — this decade, the way API-first architecture won enterprises the last one.

BCG’s prescription — unified data architecture, graduated-autonomy governance, an owned orchestration layer, transparent cost models — is a CIO job description for the next two years, not a one-time project.