Microsoft ships agents that read screens, and quietly puts Claude next to GPT

Microsoft moved two pieces this week that are, I think, the same story. Computer-using agents in Copilot Studio went to general availability on 26 May, meaning enterprise buyers can now deploy agents that navigate desktop and web applications by reading the screen and clicking, no API required. The same announcement window confirmed Anthropic's Claude is being added to Azure AI Foundry, Microsoft's platform layer that has until now centred on OpenAI models. Microsoft Build, 2–3 June, is the formal venue.¹

Neither move is, on its own, surprising. Together they describe what Microsoft is building: a procurement surface where the model underneath is increasingly interchangeable, and where the agent on top can reach software that was never designed to be reached by anything other than a human.

What was actually announced. The Copilot Studio May update graduates computer-using agents (CUA — agents that interact with software by reading the screen rather than calling an API) from preview to general availability. Microsoft positions this against the long tail of enterprise software: legacy desktop apps, on-premise systems, vendor portals with no modern integration surface. The Claude addition to Azure AI Foundry was confirmed in the same news cycle without a specific version, pricing tier, or rev-share disclosure.¹ Build is where the commercial terms are expected to land.

I'd read the primary Copilot Studio release note for the exact language, and the absence of executive quotes in the secondary coverage is itself a tell — this is being rolled out as a product update rather than a positioning announcement. The positioning will happen at Build.

The structural frame. Two lenses apply here and they point in compatible directions. The first is the FDE (forward-deployed engineering — the model where vendor or consultancy engineers embed in customer teams to ship AI) market structure question: who captures the integration work as enterprises deploy agents? The second is the SaaS apocalypse frame: what happens to per-seat software pricing when agents become the user?

CUA at GA is a structural move against the systems integrator economy. The work of wiring a legacy claims-handling system, an old ERP screen, or a vendor portal into an enterprise workflow has historically been billed by Capgemini, Accenture, and the RPA (robotic process automation — software bots that automate clicks and form-fills) specialists, by the project. The pitch is that you build a bespoke connector, or an RPA bot trained on the screens. CUA does not eliminate this work — it relocates it. The integration becomes a prompt, a policy, and a set of guardrails inside Copilot Studio, configured by an enterprise builder rather than a consultant. The platform absorbs the billable hour.

That second-order effect is the SaaS apocalypse beat. Vendors who spent the last eighteen months racing to ship "AI-ready" APIs and MCP servers to defend seat counts were defending against the wrong attack. The agent reads the screen. It does not care whether you built the integration.

Where the frame holds. CUA explicitly targets the segment, apps with no API, where SI fees are highest and where SaaS vendors have the least integration leverage. The FDE displacement reading fits this cleanly. Microsoft is not pricing CUA as a premium consulting product; it is pricing it as a Copilot Studio capability, which means the value capture moves to seats of Copilot Studio and to Azure consumption, not to project-based services revenue. That is a structurally different P&L for the integration work.

Where the frame breaks, or at least bends. RPA vendors, UiPath, Automation Anywhere, have offered screen-interaction automation for a decade, with mature audit logging, governance, and enterprise compliance tooling that Copilot Studio CUA does not yet match. The SI displacement thesis has been made for RPA before. The incumbents adapted; the consultancies built RPA practices on top. There is no obvious reason that pattern does not repeat with CUA, at least for the first cycle. What is different is the reasoning layer, not the screen-reading.

There is also a reliability gap. GA designates availability, not parity with API integrations. Multi-step workflows where every click matters, and where one misread screen propagates errors downstream, remain the hard case. Enterprise buyers in regulated industries will move slowly, and the audit-trail question is not solved by shipping the capability.

The Claude-on-Foundry piece, structurally. Microsoft adding Claude to Azure AI Foundry alongside OpenAI does one specific thing: it formalises Microsoft as the procurement surface, and removes "which model" from the locked-in part of the relationship. The original Azure–OpenAI arrangement, with Microsoft's roughly $13bn cumulative commitment, included exclusivity provisions for OpenAI hosting on Azure.² Claude's presence means that exclusivity has either expired, been renegotiated, or was always narrower than reported.

The reading I'd hold lightly until Build: Microsoft is moving its moat from model exclusivity to governance, identity, compliance, and the enterprise contract. Which weights run underneath becomes a routing decision the customer makes inside a single procurement surface. This is good for Microsoft, neutral-to-mildly-bad for OpenAI's distribution economics, and structurally interesting for Anthropic, which gets enterprise reach without having to build the procurement layer itself.

The risk on this reading is that the commercial terms still heavily favour OpenAI, priority placement, rev-share, default routing in Copilot products, and the multi-model posture is marketing rather than parity. That is the specific thing to watch at Build: disclosed pricing differentials, default-model behaviour in Copilot, and whether Claude in Foundry comes with a comparable Azure-native SLA or is positioned as an additive option.

The thing nobody is pricing yet. Every computer-using agent generates a complete record of what it saw and did across applications. In a regulated enterprise, that record contains screens, fields, and cross-system activity that was never intended to be logged in a single corpus. Current data-loss-prevention tooling was designed for queries and API calls, not for agents that see. The compliance debt is structural, and it will be discovered by whichever bank, hospital, or law firm first has a CUA deployment audited.

What to watch.

• Build (2–3 June): disclosed Claude pricing in Foundry, default-model behaviour in Copilot products, any explicit rev-share or routing language.
• Copilot Studio CUA error-rate disclosures and any enterprise reference customers in regulated sectors.
• Whether Capgemini, Accenture and the larger SI firms announce Copilot Studio practices in the next two quarters — the tell that the billable hour has moved.
• UiPath and Automation Anywhere positioning in the next earnings cycle. If they reposition as governance and audit layers over Copilot Studio CUA, that is the incumbents adapting on schedule.
• Anthropic's enterprise revenue disclosures in any subsequent funding materials — Foundry distribution should show up if it is commercially meaningful.

Footnotes

¶