Every DACH Mittelstand company with Microsoft 365 licences has had the same conversation in the last twelve months. Someone — the IT lead, an innovation manager, a board member who attended a Microsoft event — has said: "We should build AI agents in Copilot Studio. It is part of our existing stack, it is low-code, and Microsoft says it can do multi-agent orchestration now." The statement is not wrong. But it is incomplete in ways that determine whether the investment produces a functioning agent system or an expensive proof of concept that cannot scale to the workflows where AI creates real enterprise value.
This is a practitioner assessment based on hands-on implementation experience and the state of the platform as of mid-2026. It is not a product review. It is an architectural evaluation — what Copilot Studio genuinely delivers, where it hits hard limits, and how those limits map to the Three Levels of AI integration that determine whether your AI investment produces tool-level productivity or workflow-level transformation.
What Copilot Studio actually delivers in 2026
Copilot Studio has matured considerably since Microsoft folded Power Virtual Agents into it in November 2023 and built it out into a full agent platform. The product has received aggressive, near-monthly updates, and over the course of 2026 its multi-agent capabilities — Agent-to-Agent communication, Microsoft 365 Agents SDK orchestration, Fabric data-agent integration — have rolled to general availability. This is genuine capability, not vapourware. Understanding what it can actually do is essential before discussing what it cannot.
Built-in RAG that eliminates the retrieval pipeline. The Knowledge tab lets you connect SharePoint files, Dataverse and SQL data, and public websites as grounding sources. The platform handles retrieval and grounding for you — no custom vector database, no embedding pipeline, no retrieval tuning to stand up. For a company policy bot, an HR FAQ agent, or an IT helpdesk assistant, that collapses time-to-value from months to a working pilot in days. A knowledge agent that answers questions about internal policies, product documentation, or compliance procedures is the kind of thing a competent maker can have running inside a week. That speed is the platform's single strongest argument.
Multi-agent orchestration via a master-child pattern. The standard architecture is hub-and-spoke. A master agent receives the request, interprets intent through its instruction set, and routes to the appropriate child agent. Each child has its own instructions, knowledge sources, and tools; it completes its task, the turn ends, and control returns to the master. On top of that, Copilot Studio now supports Agent-to-Agent (A2A) communication over an open protocol, so an agent can delegate work to other agents — first-party, second-party, or third-party — using shared organisational context rather than a brittle point-to-point integration.
Two orchestration modes with different trade-offs. Generative orchestration lets the agent decide autonomously where to route based on the prompt — faster to configure, but the routing logic is opaque. Classic orchestration requires you to specify the routing paths explicitly — more work to set up, but deterministic and auditable. For enterprise deployments where you have to explain why a request was routed to a particular agent, the deterministic path is usually the right choice, even though the generative mode is the more seductive default.
Model choice — within the Microsoft perimeter. This is one area where the 2025 conventional wisdom is now wrong. As of early 2026, Anthropic's Claude models (Claude Sonnet 4 and Claude Opus 4.1) are selectable in Copilot Studio alongside OpenAI's GPT models, with OpenAI remaining the default for new agents. The catch matters for this audience: for organisations in the EU, UK, and EFTA — which is most of DACH — an administrator must explicitly opt in before Anthropic models become available, and if they are later disabled, affected agents silently fall back to the default OpenAI model. So model choice is real, but it is governed choice: you are selecting from the models your tenant policy permits, not freely picking the best model for the task. For anything beyond that curated list — Gemini, Llama, your own fine-tuned models — Microsoft points you to its pro-code platform, Foundry, which is a separate environment with a separate skill set.
Governance infrastructure that enterprise IT genuinely values. Data loss prevention policies, admin controls, connector permissions, and audit trails are built in. For an IT department that has to govern what data agents can touch and what actions they can take, Copilot Studio supplies controls a hand-rolled framework would take months to reproduce. This advantage is real and routinely undersold in the "low-code versus pro-code" debate — it is one of the main reasons enterprise IT teams advocate for the platform in the first place.
Where Copilot Studio hits its architectural ceiling
The capabilities above are real, and they cover a meaningful share of what most organisations attempt in their first year of agent work. But the most valuable applications tend to sit in the part that the platform does not cover — and there the architectural ceiling becomes visible.
No shared memory across agents. This is the most consequential limitation. Agents in Copilot Studio do not natively share learnings, state, or context with each other beyond the orchestration hand-off. When the master routes to a billing agent, that agent has no knowledge of what the support agent discovered about the same customer an hour earlier. When a compliance agent flags a risk, the operations agent does not absorb that finding to prevent the next occurrence. Each agent runs in its own context, and the contexts do not compound.
That matters because the core thesis of an AI Operating System — and a recurring conclusion across the enterprise AI research from McKinsey and BCG — is that AI value compounds when systems learn across interactions, share findings across functions, and accumulate organisational knowledge over time. A system where each agent starts fresh on every turn produces linear value. A system where agents build on each other's learnings produces compounding value. Copilot Studio, by design, produces the former.
No autonomous decision-making. Copilot Studio agents are conversation-triggered. A user asks, the agent answers. The platform's native model is not built for an agent that independently watches a data source, detects an anomaly, and acts within governance boundaries with no human prompt to set it off. The agent that monitors inventory and raises purchase orders when reorder points are hit — the kind of autonomous behaviour that produces workflow-level transformation — sits outside what the platform is designed to do natively.
No iterative reasoning loops. Agents cannot natively engage in multi-turn negotiation, debate, or self-correction with one another. A research agent that drafts findings, passes them to a validation agent that checks them against source data, takes the corrections, and iterates until the findings hold — a standard pattern in pro-code multi-agent frameworks — is not how master-child orchestration works. The pattern is route, execute, return. There is no native mechanism for agents to challenge, refine, or build on each other's output in a loop.
Black-box orchestration. When generative orchestration misfires — routes to the wrong agent, misreads intent, drops context — debugging is hard, because the routing logic is opaque and the official documentation does not always track the live behaviour of a platform that changes monthly. That is a governance risk as much as an engineering one. If you cannot explain why your agent reached a particular decision, you will struggle against the logging and traceability obligations the EU AI Act places on high-risk systems: under Article 12, high-risk AI systems must automatically record events over their lifetime so that operation is traceable and post-market monitoring is possible. Opaque routing and Article 12 are not natural allies.
Platform coupling. Copilot Studio is tied closely to Azure, Microsoft Entra identity, and Power Platform licensing. Cross-system orchestration — a single workflow whose agents span Microsoft, another cloud, self-hosted models, and third-party APIs — pushes you towards leaving the platform. For DACH Mittelstand companies with hybrid estates or data-sovereignty requirements that reach beyond what an Azure region offers, that coupling is a strategic constraint, not just a technical inconvenience.
Consumption licensing with fair-usage limits. Copilot Studio is metered. Since September 2025 the unit is the Copilot Credit, available either in prepaid packs or pay-as-you-go through an Azure subscription, billed on what your agents actually consume. Crucially, even a proactive, agent-initiated message is billed, and Microsoft applies fair-usage limits it reserves the right to revise as usage patterns evolve. For high-volume, system-driven scenarios processing large transaction counts, credits add up — and the economics push you precisely away from the autonomous, machine-triggered workflows that tend to generate the most value.
How the limitations map to the Three Levels
The Three Levels of AI Integration give the clearest frame for what the ceiling means strategically.
Level 1 — Assistance. Copilot Studio fully covers Level 1. Individual tools that make individual employees more productive — chatbots, knowledge assistants, FAQ bots, document-search agents — are its sweet spot. If your AI ambition is Level 1, Copilot Studio is a defensible, efficient, and governable choice. You will ship faster and spend less than with any pro-code alternative, inside the Microsoft estate you already run.
Level 2 — Augmentation. Copilot Studio partially covers Level 2. Simple workflow augmentation — a master that routes customer requests to specialists by intent, a procurement agent that assists with purchase-order creation, a compliance agent that flags risks in a draft — works within the platform. But advanced Level 2 — agents that act autonomously inside governance boundaries, share memory and learnings across interactions, and run multi-step workflows with iterative refinement — exceeds what it supports. This is exactly the gap where so many adopters stall: deployment without the workflow redesign that actually creates value. McKinsey's 2025 State of AI makes the scale of that gap concrete — roughly four in five organisations now use AI, yet only a small minority (its "high performers," about six percent of respondents) attribute material EBIT impact to it, and the differentiator is workflow redesign, not the model.
Level 3 — Autonomy. Copilot Studio cannot reach Level 3. Multiple autonomous systems coordinating across functions, sharing learnings, and optimising an end-to-end process is beyond the platform's design. And this is where the value is heading: BCG's 2025 research puts the share of AI value attributable to agentic AI at 17 percent in 2025, rising to a projected 29 percent by 2028. Capturing that requires shared memory, autonomous monitoring, cross-system coordination, and iterative reasoning — structural absences in Copilot Studio, not features waiting on the next release.
Microsoft Foundry: Microsoft's own escape hatch
Microsoft itself acknowledges the ceiling. Foundry — the platform that began as Azure AI Studio, became Azure AI Foundry at Ignite 2024, and was renamed Microsoft Foundry at Ignite 2025 — is the pro-code companion. Its Agent Service is built for complex, multi-agent orchestrations with a broad model catalogue, custom models, and enterprise networking, explicitly aimed at developers and ML engineers rather than citizen makers.
Foundry bridges the gap between Copilot Studio's simplicity and a full custom build. It gives you more architectural control without forcing you to stand up and maintain agent infrastructure entirely from scratch. For organisations that need more than Copilot Studio offers but are not ready to commit to a bespoke framework, it is a legitimate intermediate step — and, increasingly, a unifying layer across Azure, Microsoft 365, and Fabric.
But Foundry does not dissolve the harder questions. It buys you flexibility within the Microsoft ecosystem; it does not, on its own, hand you the cross-platform orchestration and data-sovereignty options that a hybrid DACH estate often demands. It is a better starting point, not automatically a different destination.
The right question to ask
The question is not whether Copilot Studio is good. It is good — genuinely capable, rapidly improving, well-governed. The question is whether it is enough for where your organisation needs to go. If your AI strategy is Level 1 — tools that make individual employees more productive inside existing workflows — Copilot Studio is very likely the right answer. It ships fast, governs well, and lives inside the Microsoft estate most DACH organisations already run.
If your strategy carries Level 2 or Level 3 ambition — redesigning workflows around AI, building agents that learn and compound knowledge over time, deploying autonomous systems that act within governance boundaries — then Copilot Studio is a starting point, not the platform. From there you will either build the integration layer in pro-code or accept that your investment plateaus at the level where most companies stall: tool-level productivity without workflow-level transformation.
The mistake most organisations make is not starting with Copilot Studio. It is having no plan for what comes after it. They deploy Level 1 agents, declare victory, and discover a year or two later that they have automated conversations while their more disciplined competitors have rebuilt the workflows underneath them. The platform choice is downstream of that strategy — which is the conversation worth having before you commit months of build to a ceiling you have not yet measured.
A Fit Call maps where your agent architecture sits in the Three Levels framework and whether your current platform choice can carry you to where you need to go — before you invest months in a platform that cannot get you there.
References: Microsoft Copilot Blog, "Microsoft Power Virtual Agents, now part of Microsoft Copilot Studio," 2023; Microsoft Copilot Blog, "New and improved multi-agent orchestration, connected experiences, and faster prompt iteration," 2026; Microsoft Copilot Blog, "Anthropic joins the multi-model lineup in Microsoft Copilot Studio," 2026; Microsoft Learn, "Billing rates and management — Microsoft Copilot Studio," 2026; Microsoft Learn, "What is Microsoft Foundry Agent Service?," 2026; EU Artificial Intelligence Act, Article 12 (Record-keeping), 2024; McKinsey & Company, "The State of AI: How Organizations Are Rewiring to Capture Value," 2025; BCG, "The Widening AI Value Gap," 2025.