The short version: Microsoft spent the last year turning Dataverse from a low-code database into the governed place AI agents read and write business data. The MCP server went GA in December 2025, a Python SDK followed in June, agents are getting their own audited identities in Entra, and Work IQ is putting Dataverse tables inside Microsoft 365 Copilot conversations. What each one is, dated from Microsoft's release plans, and what, if anything, a small NZ firm should do about it.
Getting existing data into Dataverse in the first place is on the Power Platform migration hub.
How to read Microsoft release waves, in one paragraph
Microsoft ships Power Platform in six-month waves: wave 1 covers April to September, wave 2 October to March. The 2026 wave 1 plan ran April to September, so some of what follows is generally available and some is still in public preview, and Microsoft's own release-plan pages carry the standing caveat that timelines can move and planned features can vanish. Everything below is dated from those pages as they read in August 2026. Treat any GA date as stated intent and check the current status before you plan around one. Microsoft has also said release plans stop being published from September 2026, with new capability going to the AI at Work roadmap instead, so that is where to look for anything newer than this post.
The four changes that matter
| Change | Status, per Microsoft's release plan | Plain English |
|---|---|---|
| MCP server | GA since 15 December 2025 | AI agents query and act on your business tables through a standard protocol, not a custom integration |
| Python SDK | GA since 1 June 2026 | Developers script against Dataverse natively in Python, the language most AI tooling already speaks |
| Agent identities in Entra | Public preview since June 2026, GA not yet announced | Agents become their own least-privilege users; their actions are audited as the agent, not a borrowed human account |
| Dataverse in Work IQ | Rolling out in preview through mid-2026 | Business tables become something you can chat and reason over inside Microsoft 365 Copilot, with workflows encoded as reusable skills |
Why the MCP server is the load-bearing one
MCP, the Model Context Protocol, is the standard that emerged from the AI world for connecting agents to tools and data, and Microsoft adopting it for Dataverse is the practical headline of the year. Until now, wiring an AI agent to business data meant custom integration work per system. With a GA MCP server, the AI tooling a developer already runs can reach Dataverse through one supported, governed front door: the agent asks for the tables, the platform enforces who's allowed what. I build with MCP servers weekly in my own tooling, and the difference between "custom integration per system" and "standard protocol with governance attached" is the difference between AI pilots and AI in production.
Pair it with agent identities and the shape is obvious: Microsoft is betting that the next wave of business software is agents acting on operational data, and that the winner is whoever makes that auditable. An agent with its own Entra identity, its own security role and its own trail in the audit log is one your accountant and your insurer can live with. That's the precondition for the whole category, and the piece I'd watch hardest if you're deciding this year where new operational data should live.
Dataverse's MCP server reached general availability on 15 December 2025, so this is now a thing you can switch on rather than a thing to watch. The setup, the tool list and the billing change are in the full Dataverse MCP walkthrough.
The rest of the wave, quickly
Storage gets a proper admin story, a Data Hub and Storage Advisor in the admin centre with recommendations, which matters because Dataverse capacity is the line item that creeps. Governance gets granular Copilot credit reporting with pay-as-you-go caps, closing the "surprise agent bill" gap I flagged in the credit billing post. Power Apps standardises modern theming and keeps expanding generative pages, and Power Automate is growing self-healing desktop flows and agent authoring. And for completeness: none of this reaches the free Teams tier, whose walls are exactly where they were.
The through-line, if you want one sentence for a strategy doc: Dataverse is being repositioned from "the database under Power Apps" to "the audited system of record AI agents operate on", and the 2026 investments, protocol access, agent identity, Copilot integration, all point the same direction.
What a small NZ firm should do
For most 5 to 20 person firms: nothing this quarter, and that's a fine answer. Treat the changes as direction rather than homework. Two exceptions. If you're commissioning any agent that touches business data this year, ask whoever builds it how the agent authenticates and what its audit trail looks like, because "it uses my login" is now the wrong answer. And if you're choosing where a new system's data lives, weight Dataverse's direction of travel. Structure and security used to settle that call. Now so does whether the AI layer arriving in your Microsoft stack can see the data at all.
Questions about the 2026 Dataverse changes
What is the Dataverse MCP server?
MCP, the Model Context Protocol, is the emerging standard way for AI agents to talk to tools and data. Dataverse's MCP server, generally available since 15 December 2025, means an AI agent can query and act on your business tables through a standard protocol instead of a custom-built integration. Practically: the AI tooling your developers already use can now reach your business data through a supported, governed front door.
What are Dataverse agent identities?
AI agents get provisioned as their own users in Dataverse with Microsoft Entra agent identities and least-privilege security roles, in public preview since June 2026 with no general availability date announced. The point is auditability: an agent's actions in your data are logged as the agent, not as whichever human's account it borrowed. If agents are going to touch business data, this is the plumbing that makes it defensible.
Does any of this matter for a small NZ firm right now?
Mostly as direction, not as a to-do list. Nothing here needs action this quarter for a typical 5 to 20 person firm. What it signals is that Microsoft is making Dataverse the governed place AI agents read and write business data, so the choice of where new operational data lives is quietly becoming an AI decision too. Release plans can also slip, so treat GA dates as Microsoft's stated intent, not gospel.
Want this mapped onto your own stack?
If you're weighing what any of this means for a system you're planning, book a free 30-minute call. I work in this stack daily, agents included, and the useful version of this post is the one applied to your data, not the general one.