Coverage of artificial intelligence tends to focus on what has just been launched: a new model, a new agent, a new capability. A closer look at recent enterprise announcements suggests the more consequential shift is not about what AI can do next, but about how organisations govern and control it. As adoption accelerates, administration, infrastructure, and workforce readiness are becoming the limiting factors, and enterprise AI governance is moving from an afterthought to a prerequisite for scaling.
Governance moves to the centre
The clearest signal is the growing investment in tooling to manage AI agents. Salesforce and Databricks have both introduced agentic governance features aimed at controlling how agents operate across an enterprise, as reported by CIO Dive. Databricks’ Unity AI Gateway, for example, governs which agents can access which systems and APIs, applies policies to large language model usage, and logs requests with associated costs. Those releases follow Amazon Web Services, whose Agent Registry (introduced in preview) provides a central place to discover, share, and govern agents across frameworks and environments.
The motivation is straightforward. As agents proliferate across systems, they introduce new layers of complexity affecting security, accountability, and oversight. Without a way to register, monitor, and constrain them, that sprawl becomes difficult to audit, which is why governance is increasingly treated as a precondition for putting AI into production rather than a box to tick afterward.
Reliability built in earlier
A related shift is visible at the development stage, where updates to agent-development frameworks increasingly emphasise secure, reliable deployment rather than raw capability alone. At the architectural level, the idea of a “context layer” is gaining traction as a way to capture business rules and decision logic, the elements that make an AI system not merely technically capable but usable in a specific enterprise setting.
Infrastructure and power expand to match
The infrastructure underpinning all of this is expanding rapidly. Amazon has signalled capital expenditure of roughly $200 billion for 2026, a substantial year-over-year increase, reflecting a broader pattern of building capacity ahead of demand; collectively, the largest hyperscalers are on track to spend well over half a trillion dollars on AI infrastructure in a single year. Energy is becoming a binding constraint, prompting deals such as Oracle’s expanded partnership with Bloom Energy, under which Oracle plans to procure up to 2.8 gigawatts of on-site fuel-cell power for its data centres. The shift toward dedicated, on-site generation underscores how acute power limitations have become. Related coverage of where value in the AI stack is shifting appears in this analysis of cross-datacenter model serving.
Adoption deepens across industry and government
On the ground, deployments are becoming more structured. Stellantis expanded its collaboration with Microsoft into a five-year partnership covering more than 100 AI initiatives across customer care, product development, and operations, an example of a large manufacturer embedding AI into core parts of the business. In the public sector, Digital Dubai launched an “AI+” programme to train 50,000 government employees, with tracks tailored to leaders, managers, and frontline staff. At a certain scale, AI adoption stops being purely a technical problem and becomes a workforce and change-management challenge.
From experiment to operations
Taken together, these developments point to AI moving out of an experimental phase and into an operational one, where success depends less on access to the newest model and more on governance, infrastructure, and the ability to use these systems reliably and at scale. The competitive edge increasingly lies in control: knowing what agents are running, on what data, under what rules, and with what oversight.
What to watch
This is an emerging pattern drawn from a cluster of recent announcements, and vendor messaging naturally emphasises the upside of new governance products; their real value will be proven by independent adoption and measurable reductions in risk and cost. Several of the infrastructure and investment figures cited are company projections rather than realised spending, so they may change. The broader thesis, that operationalising AI is now as much about administration, energy, and skills as about model quality, is consistent across multiple independent reports, but the specifics will keep evolving as the market matures.
