In the SharePoint-to-RAG article I argued that grounding an LLM on your own content is the natural next step for anyone who's done enterprise search. Microsoft Copilot Studio is where that becomes low-code: a place to build conversational AI agents that answer from your data and take action through the Power Platform — no orchestration framework to hand-roll. Here's how the pieces fit, and where a builder's judgment still matters.
From Power Virtual Agents to Copilot Studio
If you remember Power Virtual Agents, this is its evolution — rebranded Copilot Studio and reoriented around generative AI. The classic, fully-authored chatbot model still exists (and matters for deterministic flows), but it now sits alongside two AI capabilities that change everything: generative answers and generative orchestration.
Topics: the deterministic backbone
A topic is an authored conversation path, matched by trigger phrases — example utterances that teach the agent when to enter it. Topics give you precise, predictable control where you need it (collecting a ticket, running a specific process):
Trigger phrases
An authored topic: trigger phrases route the user in; nodes ask questions, capture entities, and respond.
Inside a topic you chain nodes: Ask a question (capturing the answer into a variable, with built-in or custom entities doing slot-filling), Condition, Message, and Call an action. It's the same conditional-logic muscle as a Power Automate flow, applied to a conversation.
Generative answers: grounding on your data
You can't author a topic for every question. Generative answers let the agent respond from knowledge sources you connect — public websites, SharePoint and OneDrive, Dataverse, uploaded documents, or any indexed content — using retrieval-augmented generation under the hood. The agent retrieves relevant passages, the model composes an answer, and (critically) it cites the source.
Actions: turning answers into outcomes
An agent that only talks is half a solution. Actions let it do things by calling:
- Power Automate cloud flows — run a process, write to a system, kick off an approval.
- Connectors — the same 1,000+ connectors the rest of the platform uses.
- Prompts & AI Builder — invoke a custom GPT prompt for summarisation, extraction, or classification.
So "check my contract status" can call a flow that queries Dataverse and returns the live record, and "request an extension" can start the approval flow from the previous article — the agent becomes a front door to your existing automation.
Generative orchestration: let the model choose
The biggest shift in the modern Copilot Studio is generative orchestration. With it enabled, you don't wire every path by hand. You describe each topic and action in natural language, and at runtime the model reads the user's intent, picks the right actions (even chaining several), gathers the inputs it needs, and composes the response:
Generative orchestration: the model plans which actions to call from one free-form request, instead of a hand-wired tree.
Don't skip security and testing
An AI agent reaches into real data and real actions, so governance is not optional:
- Authentication — require sign-in so the agent acts as the user and respects their data permissions, rather than answering as an anonymous service.
- DLP & environments — agents live in environments and obey the same DLP policies as apps and flows; build them in a governed environment, not the default one.
- Test & measure — use the test pane while authoring, then the analytics dashboards (resolution rate, escalations, abandoned sessions) to see whether it actually helps users.
The takeaway
Copilot Studio is the low-code front end to the AI patterns this whole series has been building toward. Authored topics give determinism; generative answers ground responses on your SharePoint and Dataverse content; actions connect to the flows and connectors you already use; and generative orchestration lets the model plan across them. The platform handles the RAG plumbing — your job is curating trusted knowledge, writing crisp action descriptions, and governing what the agent can touch. That's the same enterprise discipline, pointed at a brand-new surface.