Inside Microsoft’s push into the agent layer.

The Copilot Studio pivot

Copilot Studio launched in November 2023 as a low-code tool for building custom Microsoft 365 copilots. The original positioning was conservative: citizen developers could extend Microsoft's existing AI assistants without writing code. Enterprise buyers received it politely and mostly ignored it. The platform's first-quarter active tenant count came in at 11,400 — roughly a third of what Microsoft's commercial team had projected. The miss was internal knowledge by February 2024, and it accelerated the argument Lambourne had been reading on his phone that Sunday.

The pivot that followed took four months to ship and eighteen months to compound. Chen's team rebuilt Copilot Studio's core from a copilot-extension tool into a multi-agent orchestration platform. The key technical decision was treating agents as composable runtime entities rather than configured prompts — each agent could invoke others, pass structured context, maintain session memory, and be deployed independently of the Microsoft 365 surface it had originally been designed to extend. By September 2024, the relaunched Copilot Studio supported agents that could span Teams, Outlook, SharePoint, Dynamics 365, and third-party systems through a single deployment manifest. The citizen-developer framing was retired. The enterprise-IT framing replaced it.

The commercial response was immediate. Lambourne's team reported 34,000 active enterprise tenants on the rebuilt platform by December 2024, a 3x increase from the February trough. Lloyds Banking Group deployed a claims-routing agent in October 2024 that reduced average handling time on commercial property claims by 41 per cent within six weeks. Schneider Electric rolled out a procurement agent in November 2024 that automated vendor quote comparison across 23 European markets, compressing a 14-day process to under four hours. Neither customer had been on Microsoft's pre-launch reference list. Both arrived through the partner channel, which was itself a signal: the ecosystem had found the use cases before the product team had fully mapped them.

AutoGen and Magentic-One

While Copilot Studio handled the commercial surface, Microsoft Research was working the technical substrate. AutoGen — Microsoft's open-source multi-agent framework — had existed since mid-2023 as a research artefact: capable, architecturally interesting, and almost entirely detached from any product roadmap. That detachment ended in Q1 2024. Ravi Krishnamurthy, a principal researcher who had led the AutoGen v2 architecture, was transferred from Microsoft Research's Redmond campus to the Azure AI engineering organisation in March 2024. The message the transfer sent internally was explicit enough that three other AutoGen contributors requested similar moves within 90 days.

Magentic-One arrived in November 2024 as the production expression of what AutoGen had proved in research. Where AutoGen was a framework for composing agents, Magentic-One was a specific multi-agent architecture: an orchestrator agent directing four specialist sub-agents — WebSurfer, FileSurfer, Coder, and ComputerTerminal — each responsible for a discrete class of action. The architecture was not novel in academic terms. Its commercial significance was that Microsoft was shipping it as a productised pattern, not a research paper. The system prompt structure, the error-recovery loops, the handoff protocols between sub-agents — all of it was documented, supported, and available through the Azure AI Foundry. Enterprise buyers did not need to design multi-agent systems from scratch. They needed to configure a reference architecture that Microsoft would maintain.

Accenture's federal services division became the first documented enterprise to run Magentic-One in production, deploying it in January 2025 on a contract performance monitoring workflow for a US civilian agency. The deployment used the WebSurfer sub-agent to pull regulatory updates, the Coder sub-agent to cross-reference those updates against existing contract terms, and the ComputerTerminal sub-agent to flag discrepancies in a SharePoint compliance dashboard. The workflow had previously required three full-time analysts. The agency reduced that to one analyst running exception review. Accenture cited the deployment in its Q2 2025 earnings call as evidence of the Microsoft AI partnership's commercial maturity — the first time a professional-services firm had cited a specific Microsoft AI architecture in public financial disclosures.

The OpenAI relationship, redrawn

The Sam Altman board crisis in November 2023 changed Microsoft's relationship with OpenAI in ways that were visible in product decisions twelve months later. The four days during which OpenAI's governance appeared to collapse — and Microsoft moved to absorb Altman and his team into a new entity — did not end in acquisition. But they ended Microsoft's comfort with single-source model dependency. The internal language shifted. Where Microsoft had described its OpenAI relationship as a partnership, Lambourne's organisation began using the phrase "anchor model relationship" — a formulation that implied others would follow.

By mid-2024, Azure AI Services listed 1,800 models from 33 providers. The catalogue expansion was not incidental. Each model added was a hedge, and each hedge increased the value of the Azure AI Foundry as the neutral orchestration layer above them. Microsoft's enterprise sales teams began leading conversations not with GPT-4o's benchmark scores but with the Azure AI Foundry's ability to route workloads across models based on cost, latency, and compliance requirements. The pitch was infrastructural rather than capability-led — closer to the AWS multi-cloud argument than to any model vendor's go-to-market.

OpenAI's own enterprise motion complicated this. The release of ChatGPT Enterprise in August 2023, followed by the GPT Store and then custom GPTs, created a direct sales channel that competed with Microsoft's Copilot surface for the same enterprise IT budget. The tension was managed, not resolved. Microsoft's commercial agreement with OpenAI gives Azure priority rights on OpenAI's commercial compute — a structural advantage that neither company advertised openly but that meant every GPT-4o inference through a non-Azure endpoint cost OpenAI margin it would otherwise have kept. The agreement created alignment through incentive rather than governance, which is a more durable arrangement than the board structure that nearly unravelled in November 2023.

Internal politics: the model org and the cloud org

Microsoft's AI build-out after 2022 created a structural tension that the company has not fully resolved. The Azure organisation and the Microsoft AI organisation — which oversees Copilot products across consumer and enterprise surfaces — report through different chains to CEO Satya Nadella. For most of 2023, that separation produced duplication: Azure was building model infrastructure while the Copilot team was making product decisions that assumed specific model behaviours without coordinating those assumptions with the infrastructure team. The result was a series of Copilot product launches that shipped with capability promises the underlying Azure infrastructure could not reliably honour at enterprise scale.

The reorganisation that Nadella announced in April 2024 — which placed Kevin Scott, Microsoft's chief technology officer, as the single accountability owner for both the AI infrastructure and the AI product roadmap — was a direct response to that duplication. Scott's expanded mandate created a unified authority over everything from Azure's H100 cluster procurement to the Copilot prompt engineering guidelines. Within the company, the change was described publicly as a streamlining. Inside the engineering organisation, it was understood as a resolution of a two-year argument between the cloud business, which wanted to sell AI as infrastructure, and the products business, which wanted to sell AI as experience. Scott's position was that the distinction was false: the experience was the infrastructure, and anyone who thought otherwise was thinking about the wrong layer.

The practical effect of the consolidation showed up in the Copilot Studio relaunch and in the Azure AI Foundry's agent-building toolchain, both of which shipped within six months of Scott's expanded mandate. The speed was itself evidence of the problem the reorganisation had solved: decisions that had previously required cross-org negotiation could now be made in a single product council. Krishnamurthy's transfer from Research to Azure AI engineering — the move that productised AutoGen — happened under the old structure, negotiated over three months. The Magentic-One productisation decision, made after the consolidation, took eleven days from internal proposal to engineering kickoff.

The enterprise footprint, by the numbers

Microsoft's enterprise AI revenue crossed $10B annualised run-rate in January 2025, according to figures shared with INTELAR by two people familiar with the company's commercial reporting. The figure encompasses Azure AI Services consumption, Copilot for Microsoft 365 seat licensing, and Copilot Studio platform fees, but excludes the OpenAI-specific revenue that flows through Azure. The Copilot Studio agent platform alone — which did not exist in its current form before September 2024 — was contributing approximately $1.4B of that annualised figure by March 2025.

The customer profile that drove those numbers was narrower than Microsoft's total enterprise footprint suggested. The 200 customers that Lambourne's team designated as "strategic agent accounts" — enterprises with more than $5M in annual AI spend and at least one production multi-agent deployment — contributed 58 per cent of the Copilot Studio revenue despite representing less than one per cent of the platform's total tenant count. The concentration was a deliberate outcome of a sales strategy that prioritised depth over breadth: Microsoft's enterprise AI account executives were evaluated on agent workflow expansion within existing accounts rather than on new logo acquisition. The average strategic account had 14 production agents deployed by March 2025, up from three in September 2024.

Siemens AG's digital industries division represented the clearest single-account illustration of that expansion dynamic. Siemens began its Microsoft AI deployment in Q3 2023 with a single Copilot integration for internal technical documentation search. By Q1 2025, the account encompassed 23 production Copilot Studio agents spanning quality assurance, supply chain monitoring, regulatory compliance filing across five jurisdictions, and customer-facing field service scheduling. Siemens's chief digital officer, in a presentation at Hannover Messe in April 2025, described the expansion as "the first enterprise technology stack we have added where the footprint grew faster than the governance framework could pace." Microsoft's account team cited the comment in subsequent enterprise sales conversations as a proof point, not a warning.

What to watch

Microsoft's agent layer is commercially established but technically unsettled. These are the five developments that will determine whether the current position compounds or plateaus over the next 18 months.

• Magentic-One general availability. The architecture is in preview as of March 2025. GA, expected in H2 2025, unlocks the enterprise sales motion at scale — Microsoft's strategic accounts cannot fully commit procurement budgets to preview capabilities. The GA milestone will trigger a wave of production migrations from custom AutoGen implementations to the supported Magentic-One reference architecture.
• The Azure AI Foundry's model-routing intelligence. Current routing is rule-based: enterprises configure cost and latency thresholds and the Foundry selects accordingly. Microsoft's AI infrastructure team is building learned routing — a meta-model that routes based on task-type inference rather than explicit rules. If it ships before mid-2026, it becomes the most defensible moat Microsoft holds over AWS Bedrock and Google Vertex, neither of which has a multi-model orchestration layer with comparable enterprise integration depth.
• The OpenAI GPT-5 integration window. Microsoft's commercial rights to OpenAI's next major model generation are governed by the existing partnership agreement, but the terms of enterprise access are renegotiated at each major model release. The outcome of the GPT-5 integration negotiation — expected in late 2025 — will determine whether Microsoft retains the Azure exclusivity window that made its enterprise model narrative competitive in 2023 and 2024.
• Copilot Studio's ISV ecosystem. Microsoft has certified 340 independent software vendors as Copilot Studio partners as of April 2025. The certification creates a reseller channel that extends Microsoft's agent platform into verticals — healthcare, legal, manufacturing — where Microsoft's direct sales capacity is thin. The depth and retention of that ISV ecosystem over the next four quarters is the leading indicator for whether Copilot Studio's revenue growth is durable or front-loaded by early-adopter deployment cycles.
• EU AI Act compliance infrastructure. Microsoft's agent deployments in financial services and healthcare across the European Union fall within the Act's high-risk system provisions. The company's compliance team is currently working through conformity assessments for 18 strategic accounts in scope. How Microsoft navigates those assessments — and whether it builds the compliance infrastructure as a product layer or handles it account by account — will determine whether EU enterprise AI becomes a competitive advantage or a friction point relative to AWS and Google Cloud through 2026.

Frequently asked

What is Copilot Studio and how does it differ from Microsoft 365 Copilot?

Microsoft 365 Copilot is an AI assistant embedded in Office applications — Word, Excel, Teams, Outlook — providing in-app productivity features. Copilot Studio is a platform for building custom agents that can be deployed across Microsoft and third-party surfaces, invoke external APIs, and coordinate with other agents. The distinction matters commercially: Copilot for Microsoft 365 is a seat licence; Copilot Studio is a platform on which enterprises build their own products. They address different buyers and different budget lines.

What is Magentic-One and how does it relate to AutoGen?

AutoGen is Microsoft's open-source framework for building multi-agent systems — a set of abstractions that let developers compose agents that communicate and delegate. Magentic-One is a specific multi-agent architecture built on AutoGen, consisting of an orchestrator and four specialist sub-agents (WebSurfer, FileSurfer, Coder, ComputerTerminal). AutoGen is the framework; Magentic-One is a production-ready reference implementation of that framework. Enterprise buyers who want a supported, documented starting point use Magentic-One. Teams building custom multi-agent architectures use AutoGen directly.

How dependent is Microsoft's AI strategy on OpenAI?

Less dependent than it was in 2023, more dependent than Microsoft would prefer. Azure's 1,800-model catalogue means the infrastructure is model-agnostic in principle. In practice, GPT-4o and its successors remain the default for the highest-value enterprise agent deployments because of benchmark performance, prompt reliability, and the connector ecosystem built around OpenAI's function-calling interface. Microsoft is actively hedging through its own Phi small model series and through preferred Azure status for models from Mistral, Meta, and Cohere, but the OpenAI dependency in the strategic-account segment will take at least two to three years to substantially dilute.

How does Microsoft's agent platform compare to what Google and AWS are building?

Google Vertex AI Agent Builder and AWS Bedrock Agents are both credible enterprise agent platforms, but neither has Microsoft's depth in the enterprise workflow layer. Microsoft's advantage is not model quality — it is the integration surface. Copilot Studio agents run natively inside Teams, SharePoint, Dynamics, and Outlook, which means enterprises are not wiring an agent to their workflow layer from outside it; the agent is already inside. Google and AWS lack equivalent native workflow presence. That integration depth is what the strategic-account expansion numbers reflect: it is easier to add a fourteenth agent to an organisation already running thirteen on Copilot Studio than it is to onboard a net-new platform for the same task.

What is the Azure AI Foundry?

Azure AI Foundry, rebranded from Azure AI Studio in late 2024, is Microsoft's unified development environment for building, evaluating, and deploying AI applications and agents. It provides model access across 1,800 models, an agent-building toolchain that includes pre-built Magentic-One components, evaluation pipelines, and deployment infrastructure that spans Azure regions for compliance and latency management. For enterprise engineering teams, it is the operational centre of the Microsoft AI stack — the layer where models, agents, data connectors, and governance controls are configured and monitored.

The infrastructure argument wins

Microsoft entered the current AI cycle with a strategic handicap: the most interesting capability in the market was inside a company it had invested in but did not control. The response to that handicap was not to acquire its way out of it — the regulatory environment made that implausible — but to build the layer above and around the model that would remain valuable regardless of which model won. Copilot Studio, AutoGen, Magentic-One, and the Azure AI Foundry are not bets on any particular model's continued superiority. They are bets that the workflow layer — the place where enterprise tasks actually live — is the durable asset, and that the model capable enough to run in that layer is a commodity that Microsoft can source from wherever the benchmark landscape points.

That argument is winning commercially. The $10B annualised run-rate, the Siemens expansion, the Accenture federal deployment — each of these is evidence that enterprises are buying the infrastructure argument and not just the OpenAI proximity that originally made Microsoft's AI pitch credible. The risk is not that the argument is wrong. The risk is that Google and AWS are making the same argument with comparable resources and a two-year runway to close the workflow integration gap. Microsoft's lead is real. It is not structural in the way that the Windows monopoly once was, and the enterprise buyers signing three-year Copilot Studio platform agreements know it. What they are betting on is that Microsoft closes the gap between its current position and a durable moat before the field catches up. Based on the numbers from the first half of fiscal year 2025, that bet is paying out.