In the two weeks spanning late August and early September, enterprise AI infrastructure hardened on two fronts. Five major vendors independently shipped nearly identical three-layer governance architectures for AI agents. In the same window, McKinsey’s annual survey revealed that a third of organizations had declined a software purchase because coding agents could build the functionality in-house.

Standardizing agent oversight

Broadcom, Citrix, CrowdStrike, ServiceNow, and Genesys each shipped agent governance stacks built on the same three-layer architecture: connectivity and routing through the Model Context Protocol, a security and governance layer, and observability. Gartner found that 60% of generative AI proofs of concept were abandoned in 2024, with governance gaps as the primary cause, and all five vendors responded by making governance their competitive layer. Each built on existing positions in identity management, endpoint security, application delivery, and platform orchestration.

Each arrived from a different legacy position. Broadcom and CrowdStrike came through security, with AgentMinder binding agent identity at runtime and Falcon Guardian fusing agent monitoring with endpoint telemetry. CrowdStrike claims 99% detection efficacy on prompt attacks at 100-millisecond latency, a performance threshold that reflects its bet on the endpoint as the primary enforcement point. Citrix extended its NetScaler application-delivery platform with AI and MCP gateways for authentication and rate limiting. ServiceNow moved its AI Control Tower to general availability with more than thirty enterprise connectors, and Genesys anchored a four-product stack around a centralized AI Control Plane. Competitive differentiation comes from which enterprise data paths each vendor already controls.

None of the five are selling agent governance as a standalone product. Citrix includes its MCP Gateway in the existing platform license with no separate SKU. Broadcom ships AgentMinder only as part of VMware Private AI Cloud. ServiceNow bundled AI Control Tower into every tier of its restructured AI-native licensing, which replaced five legacy packages in April. Governance is becoming a default configuration of infrastructure that enterprises already own.

Moving in-house

McKinsey’s State of AI 2026, published August 25 and drawing 1,719 respondents from 97 countries, found that 32% of organizations had declined at least one software purchase because coding agents could build the functionality internally. The figure measures a procurement decision. McKinsey asked whether organizations had decided against a purchase, without tracking whether the internally built replacement had shipped or entered production. Among the 6% that McKinsey classifies as high performers, defined as organizations attributing 5% or more of EBIT to AI, the figure approached half. By sector, technology led at 41%, followed by healthcare at 39% and financial services at 36%.

The same survey found that the share of organizations attributing any EBIT impact to AI held at 37%, unchanged from 2025. Eighty percent reported individual productivity gains, but organizational earnings have not followed at the same rate. Forty percent of large enterprises now scale agents, up from 27% in 2025. One in five organizations cited AI operating costs, including token costs, as a constraint on further deployment, and high performers reported cost pressure on coding agents at roughly three times the rate of everyone else.

Shifting priorities

The vendor bundling and the build-vs-buy shift respond to the same condition. Agents now touch enough enterprise infrastructure to require governance, and they have grown capable enough to replace some of the software that runs on it. Vendors are embedding governance into existing platforms to hold the infrastructure through which agents operate, even as some of the software running on that infrastructure gets rebuilt in-house.

Governed agent infrastructure makes enterprise AI deployment viable at scale, which accelerates the coding-agent capability that erodes vendor software revenue elsewhere in the stack. The large enterprises scaling agents will run them through the governance layers that these five vendors just shipped. Some of those agents will build the tools that would otherwise have been a vendor sale.

Just the beginning

Governance and procurement are now reinforcing each other. The governance stacks shipped over the past two weeks make enterprise agent deployment viable, and the agents deployed through them generate the coding capability that lets organizations skip the next vendor purchase. That dynamic pushes vendors deeper into governance bundling to hold the customer relationship.

The 37% EBIT figure, flat for a full year, suggests that both fronts have hardened faster than the profit case beneath them. Enterprise AI’s plumbing and procurement model have each found their shape, but the organizational returns that would justify the pace of change still reach only a third of the market.

Subscribe now

AI Central

A Systemic Shift