Copilot Studio vs Power Virtual Agents - What Actually Changed
If you've been running Power Virtual Agents for the last few years, you've probably noticed it disappear from Microsoft's product pages. The product wasn't deprecated. It was rebranded and substantially upgraded into Copilot Studio. But the rebrand happened during the biggest reshuffle in Microsoft's history, and a lot of people I talk to are still trying to work out what's actually changed under the hood versus what's just marketing.
This is the post I keep meaning to send to clients when they ask me "should we upgrade?" or "do we need to migrate?"
The short version: Copilot Studio is Power Virtual Agents plus generative AI plus an actual agent framework. If you're already on PVA, your bots still work. But you're using maybe 30% of what the product can now do. If you're evaluating a chatbot platform for the first time in 2026, ignore everything you read about PVA. It's a different product now.
A quick history for context
Power Virtual Agents launched around 2019 as Microsoft's no-code chatbot builder. It was a topic-based system. You'd define trigger phrases, build conversation flows in a visual designer, and connect to backend systems via Power Automate. It was reasonable. Not exciting. Solid for FAQ bots and basic service desk scenarios.
In late 2023 Microsoft started bolting generative AI features onto PVA. Then in early 2024 they renamed the whole product to Copilot Studio and pulled in a bunch of capabilities from the broader Copilot family. By 2026 it's become Microsoft's headline product for agentic AI, not just chatbots.
So when someone asks "what's the difference," the honest answer is: Copilot Studio is several generations ahead of what PVA was, even if you've been on the same SKU the whole time.
The five things that are genuinely different
1. Generative answers, not just authored topics
In PVA, you had to author every topic. Question comes in, the system matches it to a trigger phrase, runs your flow, returns your scripted response. If a user asked something you hadn't authored, the bot would say it didn't understand.
In Copilot Studio, you can point the bot at a knowledge source (a website, a SharePoint, a set of uploaded documents, a Dataverse table) and it generates answers using GPT-4 class models. The topics still exist for structured flows where determinism matters, but the default behaviour is now "answer from knowledge."
This is the single biggest change. It moves Copilot Studio from "FAQ tool" to "actual question-answering assistant." In our experience this also moves the success/failure question from "did you author enough topics" to "is your underlying content any good," which is harder to fix but more honest.
2. AI-powered authoring
PVA authoring was clicky. You drew flows on a canvas. Copilot Studio lets you describe what you want in natural language and it scaffolds the topic for you. "Create a topic that asks for the user's order number, looks it up in our system, and returns the shipping status." Done in 30 seconds.
The generated topics aren't always production-ready. You'll still need to clean them up. But it's roughly 5x faster than building from scratch, especially for simple flows. For a mid-sized rollout (say 20-50 topics) this saves real weeks.
3. Actions and tools - the agent layer
This is where Copilot Studio moves beyond chatbot territory. You can give your copilot a set of "actions" (basically tools the LLM can choose to call). The model decides when to invoke them based on the conversation.
Actions can be:
- Power Automate flows
- REST APIs (with auth, including OAuth and managed identity)
- Connector actions from the 1,000+ Power Platform connectors
- Custom Microsoft Graph queries
- Other agents (yes, agents can call agents now)
In PVA, you had to author the decision logic. The bot calls this flow when this topic is triggered. In Copilot Studio, you describe what each action does and the model figures out when to use it. This is what makes it an "agent" rather than a chatbot. It's also what makes it harder to test, because the behaviour isn't fully deterministic.
For a deeper look at where this fits into Microsoft's broader agent story, see our work on Microsoft AI Agent Framework consulting.
4. The autonomous agent tier
In early 2025 Microsoft added autonomous agents to Copilot Studio. These don't need a user to chat with them. They sit there, listen for triggers (a new email, a Dataverse record change, a scheduled time), and act on their own.
This is genuinely new ground. We've built autonomous agents for:
- A Brisbane logistics firm that triages inbound carrier emails and updates shipment records
- A Sydney legal practice that takes new client intake forms and creates a draft engagement letter
- An insurance broker that monitors policy renewal dates and drafts client communications
PVA could not do any of this. The autonomous tier needs its own licensing (it consumes messages much faster than interactive chat) and is where the cost story gets serious. Pricing in Australia generally lands around AUD $0.18 per message for interactive use and AUD $0.45 per message for autonomous use, but message definitions are nuanced and you should model your specific use case.
5. Channels and deployment have expanded
PVA published mainly to Microsoft Teams, web chat, and a handful of channels via the bot framework.
Copilot Studio publishes to all of those plus:
- Microsoft 365 Copilot (your custom agent shows up inside the user's main Copilot)
- SharePoint sites as embedded agents
- Outlook as a side panel
- Standalone Copilot Studio mobile apps
- WhatsApp, SMS, and voice via the right connectors
- Custom apps via the Direct Line API (still works)
The Microsoft 365 Copilot integration is particularly important for organisations that have already standardised on M365 Copilot. Your custom agent shows up alongside the built-in Copilot experience without users needing to learn a new tool.
What hasn't changed
Some things are essentially the same.
The visual topic designer still exists and looks familiar. Your existing PVA topics still work without changes. The Power Automate integration is the same plumbing. The Dataverse storage is the same. Authentication options are the same (Azure AD, generic OAuth, no auth, depending on channel).
If you've got a working PVA bot, you don't have to rebuild it. You can keep it as-is, then progressively add generative answers and actions where they help.
Pricing - and this is where people get confused
PVA pricing was tenant-based. Copilot Studio is consumption-based. Two SKUs in 2026:
- Per-user: AUD $40-50/user/month, included with most Microsoft 365 Copilot licenses. Gives that user access to build and use agents.
- Per-message: Pay-as-you-go message packs. Around AUD $300-400 for 25,000 messages. Different message types consume different "amounts" of your pack.
For a typical mid-market deployment with one customer-facing copilot doing 10,000 conversations a month, expect total Copilot Studio costs of AUD $1,200-2,500/month depending on how many messages each conversation consumes. Authored topics consume one message. Generative answers from knowledge consume two. Autonomous agent actions can consume five or more.
This is more expensive than PVA was. It's also more capable. We've broken down the full pricing model in detail in our Copilot Studio pricing guide for Australia.
When to use Copilot Studio in 2026
It's the right tool when:
- You're already a Microsoft 365 shop
- You need an agent that integrates with Dataverse, SharePoint, Teams, or Outlook
- Your IT team is more comfortable with low-code than custom development
- Your use case fits within the "answer questions and take routine actions" envelope
- You need fast time-to-value (4-8 weeks to a production agent is realistic)
It's the wrong tool when:
- You need fully custom UI or complex conversational flows
- You're handling very high-volume customer service (the message economics get rough above 100,000/month)
- You need deterministic behaviour for compliance reasons (insurance quotes, financial advice, medical triage)
- Your data is mostly outside the Microsoft ecosystem
- You need on-premises deployment
For those cases we usually recommend a custom build using the Microsoft AI Agent Framework, LangChain, or Semantic Kernel. See our breakdown of building custom AI agents vs Copilot Studio for the decision framework.
Decision table
| Scenario | Copilot Studio | Custom build |
|---|---|---|
| Internal HR/IT FAQ bot | Yes | No |
| Customer service for M365 customers | Yes | Maybe |
| Insurance claims triage | No | Yes |
| Public website chatbot under 50k conversations/month | Yes | Maybe |
| Public website chatbot over 200k conversations/month | Maybe | Yes |
| Voice agent for inbound calls | Maybe | Yes |
| Agent that drives a custom application UI | No | Yes |
| Document-heavy RAG with custom retrieval | No | Yes |
| Sales enablement agent inside Teams | Yes | No |
| Autonomous agent for email triage | Yes | Maybe |
Common misconceptions we hear
"It's just a rebrand." No. The generative answers, agent actions, and autonomous tier are substantially different capabilities. The visual designer looks similar, which leads people to underestimate what's been added underneath.
"It'll replace our customer service team." Almost never. Done well, it deflects 20-40% of routine queries and frees agents for complex work. Done poorly, it makes customers angry and increases call volume to the live team.
"We can just turn on generative answers and we're done." You can, and the results will be mediocre. Generative answers are only as good as the content you point them at. If your SharePoint is messy, your bot will be messy. Content curation is the unglamorous work that determines whether your project succeeds.
"PVA users need to migrate." No. Your existing bots keep working. You can incrementally adopt new features. There's no forced migration.
A practical migration approach for PVA users
If you've got an existing PVA bot, here's the path we usually recommend:
- Audit your existing topics. How many fire monthly? Which ones have low confidence scores? Which ones get escalated to human agents most often?
- Identify generative answer candidates. Topics that are essentially FAQ retrieval are prime candidates to be replaced by generative answers over your knowledge sources.
- Add knowledge sources first. Point Copilot Studio at your existing FAQ pages, policy documents, or product documentation. Test the responses against your top 20 user queries.
- Introduce actions for transactional flows. Where your bot needs to look something up or update a record, convert these to actions and let the model decide when to invoke them.
- Consider autonomous agents for high-value workflows. Once you trust the interactive experience, look for back-office workflows where an autonomous agent could replace or augment manual work.
Most clients we work with go from "PVA chatbot" to "Copilot Studio agent" over 8-12 weeks. The work is mostly content curation and testing, not technical rebuild.
When to bring in help
Honestly, if you're a confident Power Platform shop with someone who's already built bots in PVA, the basics of Copilot Studio you can probably handle in-house. Where consultants like us add value:
- Setting up the right governance model before you have 50 shadow IT copilots in your tenant
- Architecting agents that combine authored topics, generative answers, and custom actions cleanly
- Building autonomous agents where the production risks are higher
- Connecting Copilot Studio to non-Microsoft systems via custom connectors and on-premises gateways
- Evaluation frameworks so you can measure whether your agent is getting better or worse over time
We're a Sydney-based AI consultancy that does a lot of work in the Microsoft AI stack. If you've got a PVA bot you're trying to modernise, or you're planning a Copilot Studio rollout and want a sanity check, get in touch via the contact page. For straight Copilot Studio implementation, our Copilot Studio consultants handle everything from strategy to deployment.
The rebrand happened. The product is much better than it was. Whether it's right for your specific use case is the question worth asking, and that's the conversation we'd rather have than one about whether to migrate at all.