Power Virtual Agents Is Gone. Here's What Replaced It and Why It Matters.
If someone on your team still calls it “Power Virtual Agents,” they’re working from an outdated map. Microsoft retired the product on November 15, 2023 — not just the name, but the underlying architecture, the pricing model, and the entire design philosophy behind it.
What replaced it is Microsoft Copilot Studio. And while Microsoft handled the migration automatically, “migrated” doesn’t mean “modernized.” There’s a meaningful difference between your bots running inside Copilot Studio and your bots actually using what Copilot Studio is capable of.
This post explains what changed, what it means, and what your team should be doing about it.
What Power Virtual Agents Actually Was
PVA launched in December 2019 as the fourth pillar of Microsoft’s Power Platform, sitting alongside Power BI, Power Apps, and Power Automate. The pitch was simple: let business users — HR, IT, customer service, marketing — build chatbots without writing code.
It delivered on that promise within clear limits. You built bots using a visual topic-and-flow canvas. The bot would recognize a trigger phrase, walk through a scripted conversation tree, collect inputs, and hand off to a Power Automate flow if something needed to happen in a back-end system.
That model worked well for predictable, structured scenarios: IT helpdesk triage, leave request forms, FAQ lists, password reset flows. By 2023, PVA had grown to 1,000+ certified connectors and support for 15+ languages. It was genuinely useful.
But it had a hard ceiling. If a user asked something outside the scripted paths, the bot either deflected or failed. It couldn’t reason over unstructured content, couldn’t answer open-ended questions without significant manual extension, and it had no concept of your organization’s context beyond what you explicitly programmed into it.
That constraint became increasingly difficult to defend once GPT-class models arrived.
Why Microsoft Moved On
Three things converged to make PVA’s architecture obsolete.
Generative AI made scripted topic trees feel like a workaround. Once large language models could read a SharePoint site and answer questions about it in natural language, the idea of manually authoring every conversation path looked like unnecessary work. The ceiling PVA bumped into wasn’t a product limitation Microsoft could patch — it was a fundamental design constraint.
Copilot became Microsoft’s umbrella AI brand. Microsoft 365 Copilot, GitHub Copilot, Security Copilot, Copilot in Azure — by late 2023, everything Microsoft was building around AI carried the Copilot name. More importantly, these Copilot products needed a shared surface for customization and extension. PVA was never architected to serve that role.
Dynamics 365 and Microsoft 365 needed to connect. As Microsoft built Copilot capabilities into its enterprise apps, it needed a single platform where organizations could extend those Copilots with custom agents and plugins. Copilot Studio was designed for that from day one. PVA wasn’t.
What Copilot Studio Is Today
Copilot Studio is not a chatbot builder. That framing undersells it significantly.
Microsoft’s own description: a platform for building, extending, and governing AI agents that act across systems. In practice, that breaks into three types of things you can build:
Conversational agents — What PVA did, but with generative AI as the default reasoning engine instead of scripted topic trees. These agents can pull answers from SharePoint, Dataverse, websites, and uploaded files without you manually authoring every response.
Extension agents — Custom agents that plug into Microsoft 365 Copilot and surface as skills or knowledge inside Word, Teams, Outlook, and Copilot Chat. Instead of building a standalone bot, you’re extending the AI your users are already working with.
Autonomous agents — This is the capability that has no PVA equivalent. These agents are triggered by events — a new email arriving, a Dataverse record changing, a scheduled time — and execute multi-step workflows without a human initiating each one. Autonomous agents reached general availability in March 2025.
Copilot Studio also now supports WhatsApp (GA September 2025), full voice and telephony, Computer-Using Agents that operate apps via virtual mouse and keyboard (US preview), and multi-agent systems where a top-level agent orchestrates specialized sub-agents.
None of that existed in PVA.
The Three-Layer Stack, Explained Simply
To understand why Copilot Studio behaves differently from PVA, you need a basic picture of how Microsoft structured its AI architecture. There are three layers.
Layer 1 — Foundation models. These are the AI models doing the reasoning: GPT-4.1 (the current default for Copilot Studio agents), GPT-5, and as of September 2025, Anthropic Claude Sonnet 4 and Opus 4.1 are also available within the stack. There are over 11,000 models accessible through Microsoft Foundry.
These models are powerful but context-blind. They know nothing about your organization, your employees, your policies, or your data. Out of the box, they’re general-purpose — useful, but not organization-aware.
Layer 2 — Work IQ. This is the layer that makes Copilot genuinely useful at work, and it became an official Microsoft product name at Ignite 2025 (not just marketing language). Work IQ connects to your Microsoft 365 tenant data — SharePoint, OneDrive, Teams chats, meeting transcripts, emails, Dataverse — and builds a continuously updated semantic understanding of your organization.
When a user asks Copilot something, Work IQ retrieves the relevant organizational context and passes it to the foundation model. The model reasons over real, current, company-specific information instead of guessing.
Critically: Work IQ inherits your existing Entra ID permissions. Users can only retrieve content they’re already authorized to see. Copilot doesn’t create new access — it works within the access model you already have.
Layer 3 — The Copilot interface. This is what users interact with: Copilot in Word, Excel, Teams, Outlook, custom agents published to your intranet or WhatsApp, the Copilot Chat hub. The interface is the visible surface; Work IQ and the foundation models are doing the work underneath.
A simple way to hold these three layers:
- Foundation model = the reasoning engine
- Work IQ = your organization’s memory and context
- Copilot = the interface you talk to
Without Work IQ, you have a smart but generic assistant. With Work IQ, you have one that actually understands your business.
What Happened to Your Existing PVA Bots
Microsoft was clear about this in its announcement: existing bots would migrate automatically with no interruption. Topics, flows, authentication settings, channel configurations — all carried over.
So if you haven’t touched them, they’re still running.
The problem is they’re running in what Microsoft calls “classic mode” — the original topic-based engine from PVA. They are not using generative answers. They are not grounded in Work IQ. They cannot run as autonomous agents. To unlock any of that, someone needs to explicitly enable generative orchestration in the agent settings.
The automatic migration moved your bots into the new house. It didn’t renovate them.
The Thing Most Teams Get Wrong
The most common mistake is treating Copilot Studio as a rebrand. It isn’t.
Three things changed in ways that actually affect how you operate:
The pricing model is completely different. PVA ran on message-based billing. In September 2025, Microsoft replaced that with Copilot Credits. A scripted FAQ response costs roughly 1 credit. A single deep GPT-5 reasoning response can cost up to 100 credits. If your finance team is modeling AI costs based on the old message-pack pricing, those models are wrong.
The AI has fundamentally changed. Classic PVA bots used intent recognition against manually authored topic trees. Copilot Studio’s generative orchestration uses large language models to reason over knowledge sources, decide which tools to call, and synthesize responses. The behavior, the quality, and the failure modes are all different.
Governance is now a first-class concern. PVA governance was relatively simple. Copilot Studio operates at a different scale — agents with autonomous triggers, access to sensitive organizational data, and the ability to take actions across systems. Microsoft has built a governance stack to match: Entra Agent ID (managed identity for each agent), Microsoft Purview DSPM for AI, Defender for AI runtime protection, and Agent 365 (GA May 1, 2026) as a centralized control plane. These tools exist because the stakes are higher. Not using them is a real risk.
What to Do Now
If you haven’t audited your Copilot Studio environment yet, that’s the starting point. Inventory every bot, its traffic volume, its business owner, and what it’s connected to. Even a simple spreadsheet with those four columns tells you a lot.
From there, triage by volume. Under 500 conversations per month: don’t rush a refactor. Over 5,000: that bot is doing real work and deserves modern architecture.
For the high-traffic bots, the refactor path is: enable generative orchestration, connect named SharePoint sites or Dataverse tables as knowledge sources, and test against a set of real user queries before re-publishing. Industry benchmarks put refactoring time at 8 to 12 hours per agent depending on complexity.
Before you refactor anything, sort out billing. Confirm your tenant is on Copilot Credits — not legacy message packs. Use Microsoft’s agent usage estimator to model expected consumption under generative orchestration. Credit consumption will likely run 3 to 10 times higher than classic-mode bots.
And before publishing any refactored agent to external channels, run Purview DSPM for AI in audit-only mode first. Know what data your agents are reaching before you expose them to users.
The Bottom Line
PVA served its purpose well. The topic-tree model made chatbot creation accessible to non-developers and produced useful internal tooling across the Microsoft ecosystem for several years.
But that model hit a ceiling the moment generative AI made it possible for agents to reason over unstructured content, operate autonomously, and respond to questions they were never explicitly programmed to answer. Microsoft didn’t retire PVA because it failed — they retired it because the bar moved.
Copilot Studio is where Microsoft is investing. The organizations that treat their existing PVA bots as a starting point — rather than a finished product — will end up with agents that are more capable, more maintainable, and better positioned for where the platform is headed.
The ones that assume the automatic migration handled everything will find that out the hard way when they need the new capabilities and their bots can’t deliver them.
Is your organization still running PVA-era bots?
WME helps Microsoft-first organizations audit, modernize, and govern their Copilot Studio agent estate — from legacy bot refactoring to Work IQ grounding and Agent 365 governance setup.
Talk to a Microsoft Expert
