From Pilot to Production: The Year the Control Plane Ate the Model
[NAVIGATION LINKS]
Between April and June 2026, Amazon, Microsoft, and Google each shipped the same thing: a control plane for AI agents. Not a smarter model, a governance layer. Three of the four largest cloud providers concluded independently that the model is the commodity and the harness around it is the prize. For enterprise leaders whose pilots are not becoming production systems, that convergence is the answer to a question they have been asking for two years.
Three years ago, the AI competition had one scoreboard: whose model was smarter.
Benchmark after benchmark. Release after release. The industry stacked parameters and tokens in a race that felt like it had no finish line. Chief AI officers briefed boards on benchmark numbers most executives could not interpret, and vendors made claims that no independent test had yet verified.
In April 2026, the scoreboard changed.
On 22 April, at Google Cloud Next 26, Google announced it was discontinuing the Vertex AI brand. The replacement: the Gemini Enterprise Agent Platform. That rebrand was not cosmetic. Google was signalling, in the most direct way a major enterprise vendor can signal, that the model, even one of the best in its class, is no longer the unit of strategic value. The platform governing the agents that run the model is.
Nine days later, on 1 May, Microsoft announced that Agent 365 had reached general availability. Not a model. A control plane: a single administrative surface to observe, govern, and secure every AI agent in the organisation, whether built by Microsoft, AWS, Google, or a third party.
AWS had moved earlier. Amazon Bedrock AgentCore reached general availability in October 2025. At AWS Summit New York in June 2026, the company reported that the number of tasks performed by agents on AgentCore had grown fifteen times in six months, and announced the Managed Harness as a new generally available abstraction: a production-grade agent in three API calls, with identity, memory, sandboxing, and observability built in. That fifteen-times figure is a vendor-reported metric, not an independently verified one.
Three of the four largest cloud providers, within ninety days, reached the same conclusion. The model is a commodity. The control plane is the prize.
That conclusion matters enormously for enterprise leaders still asking why their AI pilots are not becoming production systems.
The governance bottleneck, measured
There is now enough data to characterise the pilot-to-production problem precisely.
According to McKinsey's State of AI 2026 (February 2026), 88 per cent of organisations report regular AI use in at least one business function. Seventy-two per cent use generative AI, up from 33 per cent in 2024. And yet nearly two-thirds have not begun scaling AI across the enterprise. Fifty-one per cent have experienced at least one negative consequence from AI use.
Separately, according to Proofpoint's 2026 AI and Human Risk Landscape report (April 2026), 87 per cent of organisations have deployed AI assistants beyond pilot stage. Seventy-six per cent are piloting or actively rolling out autonomous agents. But 52 per cent are not fully confident that their security controls would detect a compromised AI agent. The adoption is real. The governance behind it is not.
Gartner, projecting in June 2025, predicted that 40 per cent of agentic AI projects would be cancelled by the end of 2027. Governance, observability, and unclear ownership were named as the primary failure modes, not model quality. (This is the Gartner June 2025 cancellation projection, analyst Verma. It is a separate study, from a separate analyst, from the Gartner May 2026 decommission prediction discussed below.)
The failure pattern is consistent across every major research source. Organisations are not stalling because they chose the wrong model. They are stalling because nobody can answer three questions about their agents: who owns the identity? Who approved this permission? Who sees the audit trail? The model cannot answer those questions. The control plane can.
That is what Gartner measured when it estimated that spending on AI governance platforms would reach US$492 million in 2026, rising to more than US$1 billion by 2030 (Gartner, 17 February 2026). The market did not create that spending category out of curiosity. It created it because the governance bottleneck is now a measurable cost.
What "control plane" actually means
The term sounds like infrastructure jargon. It is worth being precise about what it means in practice, because the architecture has direct consequences for every AI decision an enterprise makes in the next two years.
A control plane is the governance layer that a model does not provide. Six capabilities define it.
Identity: every AI agent needs a non-human identity, with credentials that can be granted, scoped, and revoked. Agent 365 does this through Microsoft Entra conditional access policies applied per agent. AgentCore does it through AWS Identity and Access Management integration. Gemini Enterprise uses a per-agent cryptographic identity layer called Agent Identity.
Permissions: agents need to operate with least privilege, accessing only the data and tools their task requires. The control plane enforces this boundary; the model does not.
Policy: what is the agent allowed to do? What is it prohibited from doing? The answer must be auditable and consistently applied. Each control plane provides this through different mechanisms: Purview data loss prevention in Agent 365, Guardrails integration in AgentCore, Agent Gateway policy routing in Gemini Enterprise.
Memory: short-term and long-term agent state management, so workflows can persist and resume without losing context. This is infrastructure, not model capability.
Observability: session-level traceability across every agent action, covering latency, error rates, tool selection, and goal success. Without this, production incidents cannot be diagnosed.
Audit trail: the governance record of who approved what, when, and with what outcome. This is the capability that distinguishes a production system from a sophisticated pilot.
An AI agent running without a control plane has none of these properties. It has capability. It does not have accountability. And accountability, as Gartner's May 2026 guidance noted, is precisely what distinguishes the 60 per cent of enterprises heading for stalled AI from the ones that scale.
The proportional governance requirement
Gartner's May 2026 guidance goes further than the headline prediction. It makes an architectural point that every enterprise architect should carry into their control plane evaluation.
Applying uniform governance across all AI agents, regardless of their autonomy level, leads to enterprise AI agent failure (Gartner analyst Shiva Varma, 26 May 2026). The failure mode is not choosing the wrong platform. It is misconfiguring the one you chose. An organisation that applies the same identity, permission, and audit requirements to a low-autonomy summarisation agent and a high-autonomy code-modification agent will either over-restrict the first (reducing value) or under-restrict the second (creating risk). Gartner estimates this design error will cause 40 per cent of enterprises to demote or decommission autonomous agents after production governance failures.
Each of the three control planes supports proportional governance by design. Agent 365 provides Entra conditional access policies scoped per agent type. AgentCore allows fine-grained IAM policy binding per agent role. Gemini Enterprise uses Agent Identity and Agent Gateway to enforce differentiated policy per agent. The architecture is there. Whether enterprise teams configure it proportionally, or apply uniform governance by default, is the decision that will separate AI scaling successes from failures over the next eighteen months.
Organisations with a dedicated governance platform are 3.4 times more likely to reach high effectiveness in AI deployment than those without (Gartner, February 2026). That multiplier is not about the platform. It is about what the platform forces you to think about: who owns the agent, what it is allowed to do, and what happens when it does something unexpected.
A simple classification exercise illustrates the design requirement. Consider three agents that a mid-sized New Zealand financial services organisation might deploy. A summarisation agent reads customer correspondence and surfaces key themes for a relationship manager to review. A research agent retrieves and synthesises information from internal databases and external sources on behalf of a senior analyst. An approval-workflow agent acts on customer data to initiate, modify, or flag financial transactions based on policy rules. These three agents have materially different autonomy levels and materially different risk profiles. Applying identical governance to all three fails in both directions: it over-governs the summarisation agent (reducing throughput and value) and under-governs the approval-workflow agent (creating financial and regulatory exposure). Proportional governance requires classifying each agent by its autonomy tier before selecting its identity scope, permission boundary, and escalation rules. The control plane enables that classification. It does not automatically perform it.
The open-standard hedge: A2A and the portability question
The three hyperscalers did not converge on the same control plane. They converged on the same market layer. The control planes are proprietary. The interoperability between them, at the agent-communication level, is not.
The Agent-to-Agent (A2A) protocol version 1.0, launched at Google Cloud Next 26 in April 2026, is an open specification governed by the Linux Foundation's Agentic AI Foundation. It allows AI agents from different vendors to discover each other, negotiate capabilities, and delegate tasks securely without human intervention. As of April 2026, A2A version 1.0 is in production at approximately 150 enterprise organisations.
The relevance for enterprise architecture is not A2A's current scale; it is its function. A2A enables agent-to-agent communication across vendor boundaries. It does not govern identity, permissions, or audit trails. That governance remains proprietary within each control plane.
The MCP (Model Context Protocol) precedent is instructive. MCP, now under open-foundation governance, reached approximately 10,000 public servers and 97 million monthly SDK downloads by March 2026. Every major hyperscaler adopted MCP for tool connectivity. None of them stopped competing on proprietary control planes. The open standard lowered switching costs at the communication layer. It did not commoditise the governance layer.
A2A is likely to follow the same pattern: adoption as a commodity interoperability layer, alongside continued proprietary competition on the governance infrastructure that actually determines whether an enterprise can govern, audit, and if necessary terminate every agent action in its stack.
The New Zealand practitioner's decision
For New Zealand enterprise architects, the practical question is not which control plane is best. That question has no answer that is independent of your existing stack, your data residency requirements, and your Sovereignty Agility posture.
A more useful question: when your AI agent takes an action today, can you answer the three accountability questions? Who owns the identity? Who approved the permission? Who sees the audit trail?
If the answer is "our hyperscaler provider," you have a control plane. That is a significant step. If the answer is "we are not sure," your pilot-to-production gap is not behind you.
New Zealand's context sharpens this. Deloitte NZ's 2026 agentic adoption data finds that 66 per cent of New Zealand organisations have human identity management in place, while only 36 per cent have extended those identity management strategies to non-human AI agents. (These are single-source figures from Deloitte NZ and are carried here as that firm's research, not as independently corroborated data.) That thirty-point gap is the New Zealand form of the global control-plane problem: organisations that understand identity governance for people have not yet extended it to agents that act on their behalf.
TUANZ's 2026 Digital Priorities Report (April 2026) describes what that gap costs. A surge in Shadow AI (unsanctioned tools operating outside formal governance) has contributed to recent data breaches. TUANZ chief executive Craig Young characterised New Zealand's technology posture as having "plateaued", with "strategic drift" the risk for organisations that do not actively govern their AI deployments. That "strategic drift" framing is Mr Young's characterisation, attributed to TUANZ, not a conclusion this article adopts as its own. TUANZ recommends treating AI systems as "digital employees" governed by zero-trust security frameworks, which is precisely the architecture that non-human identity management in all three control planes provides.
New Zealand's AI Strategy ("Investing with Confidence", MBIE, July 2025) is principles-based and voluntary. There is no binding AI transparency or governance obligation on New Zealand private-sector organisations comparable to Australia's framework. In June 2026, Australia's Digital Transformation Agency activated the first mandatory AI governance obligation in the Australasian region: all 94 non-corporate Commonwealth entities now maintain an internal AI use-case register with a named accountable owner. December 2026 obligations add impact assessments, oversight processes, and incident reporting. Australia's mandatory register fields (description, AI technology type, named accountable owner, scope criteria, inherent risk rating, residual risk rating, and date of last impact assessment update) are structurally similar to the metadata that all three hyperscaler control planes already enforce for agent registration.
The voluntary posture in New Zealand means the choice of control plane determines the governance architecture by default. The platform becomes the de facto governance standard in the absence of a regulatory one. That is not an argument for or against mandatory regulation. It is a practical observation: the architecture decision is unavoidable, whether it is made deliberately or by drift.
For New Zealand enterprise architects evaluating that decision in the second half of 2026, three questions are worth structuring before any vendor conversation. First: what is your existing hyperscaler footprint? If your workloads are predominantly AWS, AgentCore extends naturally; if predominantly Microsoft 365, Agent 365 is already in your licensing conversation. Procurement gravity is real and should be named, not ignored. Second: what are your data residency requirements? AgentCore and Gemini Enterprise have broader regional availability than Agent 365 at this point; for data classified under New Zealand government data sovereignty requirements, residency options matter. Third: does your architecture need to support agents from multiple providers, or is single-vendor sufficient? If multi-vendor, A2A version 1.0 interoperability in your chosen control plane should be an evaluation criterion. If single-vendor, it is less material but still worth noting for future optionality. None of these questions has a universal answer. The point is to ask them before the platform decision defaults to whichever vendor sends the first proposal.
The Sovereignty Agility dimension
Part 23 of this series described the Fable 5 and Mythos 5 suspension: a United States Commerce Department export-control directive on 12 June 2026 pulled two frontier AI models from global access within hours, with no warning and no transition window. Whether or not access has since been restored, the episode revealed a structural dependency that does not depend on its outcome.
The control-plane decision has a direct bearing on how exposed an enterprise is to that kind of access disruption.
Each hyperscaler control plane governs identity, permissions, and audit trails within its own stack. If A2A version 1.0 reaches adoption scale comparable to MCP, it creates a portable agent-communication layer: switching the underlying model or agent framework requires reconfiguration rather than platform migration. But the governance layer (who owns the identity, who sets the permission, who holds the audit trail) does not move with the agent if you migrate. It stays with the platform.
This is the Sovereignty Agility trade-off as a procurement decision. A control plane that supports A2A interoperability provides governance plus portability at the agent-communication layer. It does not provide portability at the governance layer. An organisation whose identity, permissions, and audit infrastructure are deeply embedded in a single vendor's control plane is harder to migrate than one whose agent communication runs on an open standard with governance managed independently.
For an enterprise in a small open economy dependent on US-domiciled frontier AI providers, the June 2026 event is the test case: how quickly could your production AI workflows reroute to a different model? The answer to that question is partly architectural. It depends on whether your workflows are pinned to a specific model or abstracted behind a governance layer that can reroute. The control plane is the abstraction layer. But it can also be the new pin, if the control plane itself is single-vendor.
The United States government's National Security Presidential Memorandum 11 (NSPM-11, 5 June 2026) is instructive as a benchmark rather than a compliance obligation. It codifies four federal AI governance requirements: reliability, robustness, steerability, and controllability. Each maps to a control-plane capability. Reliability maps to goal-success-rate evaluation. Robustness maps to policy enforcement via conditional access. Steerability maps to interrupt gates and human-in-the-loop controls. Controllability maps to administrator override. The United States government and the commercial market have independently converged on the same four requirements. That convergence is a governance maturity signal: the vocabulary of AI governance is stabilising.
The harness is the moat
The Hamberger Report identified this structural transition in its analysis of the agentic enterprise: the orchestration and governance layer wrapping an AI agent is the sustainable competitive advantage, not the model embedded within it. In 2026, the three largest cloud providers validated that observation at the product-market level. They built the harness. They built it into their own stack. They are pricing it at the enterprise level.
The infrastructure consolidation pattern is not new. When virtualisation became the dominant server architecture between 2006 and 2010, VMware, Microsoft Hyper-V, and Citrix converged on the same layer simultaneously. When identity and access management became a standalone market between 2010 and 2015, Microsoft Active Directory, Okta, and Ping Identity competed for the same governance role. In both cases, the open question was not whether the layer would consolidate, but who would own it and what lock-in would look like. The agent control-plane market is following the same structural pattern. Three hyperscalers concluded independently that the governance harness is the market position worth owning. The consolidation is already underway.
The practical implication for enterprise leaders is not which control plane to choose; that is a workstream for the next six months. The immediate implication is recognising that the pilot-to-production gap was never a model problem. It was a governance infrastructure problem. The infrastructure now exists. The question is whether enterprise architects configure it with the intentionality the moment requires, or inherit a governance architecture by default.
The implication on the sovereignty side is worth stating directly. A control plane consolidates the identity, the permission set, and the audit trail into one provider's stack, and the three that matter are all domiciled in the United States. For defence and intelligence procurement, that governance layer is the property classified and operational systems are built to hold under sovereign control. The United States codified the requirement for its own systems in National Security Presidential Memorandum 11 (5 June 2026), whose four control properties include controllability: the override that lets an operator stop an agent mid-action. Controllability is a sovereign-capacity question before it is a product feature. For a small open economy with no domestic hyperscaler, adopting an allied control plane means the override, and the audit trail behind it, sit with the platform rather than the nation. That is foundation-model concentration in its procurement form.
When your AI agents take actions, in your systems and in the systems your organisation depends on, can you name who owns the identity, the permission, and the audit trail?
If the answer involves "we would need to check with the vendor," the control plane has been chosen. The governance posture has not. I would like to hear from enterprise architects, CIOs, and risk leaders who have started this evaluation. Which of the three accountability questions is hardest to answer in your stack right now, and why?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger is a New Zealand leader in Architecture and Security and Associate Member of the Institute of Directors. The Hamberger Report: Generative AI 2026 provides enterprise leaders with evidence-based analysis of the AI landscape.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] Google Cloud. "Google Cloud Next 26: The Gemini Enterprise Agent Platform." 22 April 2026. https://cloud.google.com/blog/
[2] The Next Web. "Google Retires Vertex AI Brand for Gemini Enterprise Agent Platform." April 2026. https://thenextweb.com/
[3] Futurum Research. "Microsoft Agent 365 Reaches General Availability: Turning Shadow AI Into a Governed Asset Class." May 2026. https://futurumresearch.com/
[4] Amazon Web Services. "Amazon Bedrock AgentCore and the Managed Harness: General Availability." AWS News Blog. June 2026. https://aws.amazon.com/blogs/machine-learning/
[5] McKinsey and Company. "The State of AI in 2026." February 2026. https://www.mckinsey.com/
[6] Proofpoint. "2026 AI and Human Risk Landscape." 28 April 2026. https://www.proofpoint.com/
[7] Gartner. "Gartner Predicts 40 Percent of Agentic AI Projects Will Be Cancelled by End of 2027." 25 June 2025. https://www.gartner.com/en/newsroom
[8] Gartner. "Gartner Forecasts Worldwide AI Governance Platform Spending." 17 February 2026. https://www.gartner.com/en/newsroom
[9] Gartner. "Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure." 26 May 2026. https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure
[10] Linux Foundation, Agentic AI Foundation. "Agent-to-Agent (A2A) Protocol Version 1.0." April 2026. https://www.linuxfoundation.org/
[11] Model Context Protocol. "MCP Adoption Metrics: Public Servers and SDK Downloads." March 2026. https://modelcontextprotocol.io/
[12] Deloitte New Zealand. "Agentic AI Enterprise Adoption 2026." 2026. https://www.deloitte.com/nz/
[13] TUANZ. "2026 Digital Priorities Report." April 2026. https://tuanz.org.nz/
[14] Ministry of Business, Innovation and Employment. "Investing with Confidence: New Zealand's Strategy for Artificial Intelligence." July 2025. https://www.mbie.govt.nz/
[15] Digital Transformation Agency (Australia). "Policy for the Responsible Use of AI in Government: Mandatory AI Use-Case Register." June 2026. https://www.digital.gov.au/
[16] Crowell and Moring. "NSPM-11: Four Pillars of Federal AI Governance." June 2026. https://www.crowell.com/
[17] Anthropic. "An Update on Fable 5 and Mythos 5 Access." 13 June 2026. https://www.anthropic.com/news/fable-mythos-access