Summary
Microsoft briefly published, then hastily retracted, details about a new product called Agent 365 (A365) in its admin centre on 6 November. The announcement, pulled back within hours, revealed plans for AI agents that would function as pseudo-employees with their own accounts in Microsoft Entra ID, positions in organisational charts, and the ability to send emails and join meetings. What could have been a straightforward product launch has instead exposed a tangled mess of definitional confusion, pricing complexity, and unanswered questions about how enterprises will actually deploy and pay for these digital workers.
This article is based on a conversation between Alexander Golev, Partner and Head of Consulting at SAMexpert, and Jukka Niiranen, a Microsoft licensing consultant and blogger who has been tracking the Agent 365 development since August.
The Leak and the Scoop
The admin centre announcement wasn't entirely unexpected. Tom Warren from The Verge, who writes the Microsoft Notepad newsletter, had reported in August that Microsoft was working on a product initiative called Agent 365. Niiranen wrote speculative analysis at the time based on Warren's scoop, trying to piece together what the product might entail.
"The announcement from Microsoft in the admin centre was pulled back after a few hours," Niiranen explains. "Obviously it contained a bit too much information that was supposed to be kept under the blanket before the Ignite keynote probably."
Microsoft's Ignite conference runs from 18 to 21 November, where the company is expected to provide official details. For now, the industry is left parsing the fragments that escaped before Microsoft's communications team hit the delete button.
What is known: Agent 365 will require an A365 licence assigned by an administrator. These agents will have accounts in Microsoft Entra ID (formerly Azure AD), be positioned in the organisational chart like human employees, and function as what Microsoft calls "agentic users". The leaked material described them as "coachable, adaptive, and context aware" agents that "collaborate effortlessly, becoming an intuitive part of your team from day one."
The screenshots that briefly appeared included examples like a "Support agent from ServiceNow, authored by ServiceNow" and references to "AI efficiency gains" at organisations like Microsoft wellness centre hospitals. The marketplace model appears baked in from the beginning.
What Even Is an Agent?
The fundamental problem facing Microsoft, and by extension its customers, is that nobody can agree on what constitutes an "agent" in the Microsoft ecosystem, which makes it impossible to budget accurately when you're spending at enterprise scale.
"The real problem is that even before the announcement, there was already such a broad scope for what an agent can be in Microsoft products," Niiranen notes. "SharePoint has agents, but really they are just glorified search engines. You can search a site and then you'll get an LLM citing information back to you based on your prompt, and that doesn't even fill the criteria of what generally the industry or developers or what Anthropic uses as the definition of an agent."
The industry standard definition, used by companies like Anthropic, describes an agent as an LLM model that calls tools in a loop. That loop, that autonomy, is the defining characteristic. You give an assignment to an agent and it completes it, calling various tools and resources as needed without requiring step-by-step human intervention.
Microsoft's recently published MS Agent Framework documentation, released about a month ago to the developer audience, does embrace this definition. The framework is likely to become the foundation for building A365 agents, and it does include the crucial loop mechanism that queries MCP servers and other sources and actually performs work autonomously.
Agent 365 introduces AI users with licenses, identities, and autonomy — but without a clear definition or cost model.
But that's not how Microsoft has been using the term "agent" across its product portfolio.
"Microsoft themselves have muddied the waters by their previous rush to call everything an agent," Niiranen says. "They started that roughly a year ago. In Copilot Studio, for example, they changed the terminology so you now create agents, not bots. So everything went into the agent lingo at that point in time."
The problem is that not all of what Microsoft has been calling agents can possibly fall under the A365 licence immediately.
"There's no FAQ that would say, okay, what is an agent that would cover all their products," Niiranen observes. "The different product teams don't agree on that themselves."
Will the existing agents that sit inside Microsoft 365 Copilot (like Project Coordinator, Facilitator, and others announced recently) migrate to this new A365 licensing model? Or will they remain bundled with Copilot, whilst true agentic workers require the new licence?
Nobody knows. Not yet.
The Small Matter of Cost
The licensing uncertainty is already stagnating adoption of Microsoft's AI products because enterprise buyers have no reliable way to forecast costs.
Golev is blunt about the problem: "I talk to commercial people of other vendors at least monthly, and one of the biggest problems all of us have, and that includes Microsoft, is that the complexity of the Copilot licensing is not understood by anybody on the buyer side. That's why they are all flocking to us experts, because they simply don't trust each other anymore within the licensing community. It's too complex, it's too dangerous, and people are used to understanding what they buy. And they just can't."
The current working theory, still unconfirmed, is that the A365 licence functions almost as a substitute for a standard M365 licence. The agent would effectively be a user of M365 with bundled Copilot functionality. It would need access to email, Teams, Office applications, and other M365 services to function as part of the organisational workforce.
But that won't be where the costs end.
"Although that will probably have a fixed price, I don't think it's going to be the end of the costs," Golev argues. "If you look at how users behave with Copilot, you have add-ons, you have premium functionality, you have capacities, you have pay-as-you-go plans and all that complexity of the current Copilot AI functionality."
Then there's the third-party question. If an organisation deploys a ServiceNow support agent, will ServiceNow charge separately for access to that agent? The leaked screenshots suggested third-party agents will be distributed through a marketplace, with no-code tools available for publishers.
Niiranen confirms this is on the roadmap: "There's going to be a no-code publishing channel for third parties to get their agents into this store."
🖐 Navigating Microsoft AI licensing complexity requires clarity and precision. Learn how we support enterprise decisions: Microsoft Licensing Services for Enterprises.
Instance Models and Consumption Chaos
Another unresolved question: will each user need their own instance of an agent, burning one A365 licence per instance, or will agents be shared resources that multiple users can access concurrently?
"Are we going to have multiple instances of that support agent for each user who then burns one A365 licence for that instance, or are we going to have shared agent instances where everyone can access them?" Niiranen asks. "I don't remember there being anything in those definitions that they briefly published before pulling it back."
The current Copilot pricing already offers two models. There's the $30 per user per month for Microsoft 365 Copilot, which covers a bundle of user experiences and a usage allowance. Critically, it's not unlimited. There are limits for custom agent performance, and there's shared capacity across all users in that licence purchase. A single user doesn't have a quota they can max out, but they can max out everybody's capacity.
Then there's the pay-as-you-go model. Microsoft Azure OpenAI services charge for API credits. Copilot Studio uses a pay-as-you-go model of $2 per 1,000 qualifying messages (a session and a response from the LLM). Microsoft has also talked about Copilot credits, which are in preview but haven't been rolled out yet.
"All of this is still in flux," Niiranen says. "We know the high-level model, it's pay-per-use, and how will the model change then into pay-per-agent?"
The fundamental problem is token consumption. When an agent autonomously joins meetings, sends emails, processes documents, and makes decisions, it's consuming computational resources. Those tokens need to be counted and paid for somehow. But if an organisation purchases both A365 licences for agents and M365 licences for human users, how does the consumption get allocated? Are they competing for the same pool of resources, or are they separate?
"If the user is using some other Copilot themselves, just in the products and chatting with that and getting responses, or automation sending messages from multiple different sources like Power Automate and whatnot, then that's competing for the consumption," Niiranen points out. "But maybe they're purchasing two licences now, the A365 and the M365 licence."
The Price of Confusion
The confusion is holding back the entire market, and Microsoft is paying for it as much as their customers, Golev argues.
"After talking to three different suppliers of the same component or the same parts of M365 Copilot licensing, you get three different takes sometimes," Golev says. "And it's not because they're stupid, it's just because the complexity is so high and it's so new that there is literally nobody who understands everything."
When an executive asks how much Copilot will cost over the next 12 months, the honest answer is: we don't know. There are too many unknown factors.
"I understand why there are so many unknown factors," Golev acknowledges. "I understand why the credit mechanism is kind of a good idea on one side because of the fluctuating demands and all of that. But on the other side of the equation you have people who just want to understand what they are paying for, and there's this disconnect."
Niiranen has encountered the same problem from the technical side. "I talked to some partners who work on behalf of customers doing analytics and reporting. They need to combine multiple data sources just to see what their organisation has consumed today. So the telemetry and the reporting layer on top of these different pockets of consumption doesn't give even the transparency right now. It's kind of a mess in its current state."
With fragmented telemetry and no unified reporting, enterprises struggle to understand AI consumption and predict future costs.
The lack of unified reporting means organisations can't even see what they're spending in real time, much less forecast future costs. Enterprise software can't function this way.
"My hope is that Microsoft stops confusing people more," Golev says. "I don't know whether they can get the word agent back from their marketing machine. But I think that if there's a possibility to clarify that simplified pricing, I think they should pursue that."
The Marketplace Wild Card
There's another wrinkle in the Agent 365 model: the Microsoft marketplace itself. Third-party agents will presumably be distributed through this channel, but marketplace dynamics in enterprise organisations don't follow the traditional procurement process.
"The average buyer of a marketplace, even in an organisation, is a 27-year-old person," Golev notes. "Much less responsibility, much more enthusiasm. I'm not saying that's a bad thing. It's just a fact we need to deal with and be prepared for."
If junior staff members can spin up agents from a marketplace, who's tracking the proliferation of licences? Who's monitoring consumption? Who's responsible when an agent starts burning through credits autonomously? Questions that create shadow IT problems and surprise bills.
Putting it to the Test
Niiranen plans to be among the first to test Agent 365 when it becomes available, following his usual approach of finding where theory diverges from practice.
"As long as there's going to be some pay-as-you-go option that doesn't have an enterprise barrier, then yeah, I'm going to be there trying to see where the theory does not meet practice," he says. "Start hitting the walls and seeing when it breaks. That's what I usually do with these things."
But he's cautiously optimistic that Microsoft might finally be delivering something more mature. "I'm kind of hoping that they would now roll something that can be adopted by customers without avoiding addressing those obvious flaws and gaps that have existed there."
The industry has been through two years of what Niiranen calls the "happy marketing phase" around Copilot, where the messaging was relentlessly positive whilst the reality was often disappointing.
"We've had too much of that," he says. "It's not sustainable. We need to get to a place where people are realistic. People are not expecting for AI to go away, they are more waiting: if only it would work a bit better, if only we would know more about it and have guardrails in place, then we could take it into use."
Perhaps Microsoft has compressed five years of typical product evolution into two years of rapid AI development, he suggests. Perhaps Agent 365 represents a more mature, more realistic approach. Or perhaps it's another layer of complexity added to an already incomprehensible pricing model.
"Is it going to be now, two years after the first Copilot wave? Well, it could be," Niiranen says. "Maybe we've arrived at a place where agents could actually be really useful for customers."
The Rogue Agent Problem
Independent licensing consultant Rich Gibbons raised a question in his blog post that emerged during the conversation: if these agents can join meetings and send emails to people, what happens if they go rogue?
The laughter this question provoked is telling.
"Same as with people," both Golev and Niiranen responded. "Same as with people, honestly."
Microsoft is already pushing lead qualification agents in Dynamics 365 Sales that send emails to customers and collect data from them. Is it worse if agents do this internally within an organisation?
"You create some team manager or whatever and it can send an email to anyone in the org but not outside the tenant, or governance guardrails are in place," Niiranen says. "Is it going to be that much worse that people then use Copilot or ChatGPT to write an email that they're not going to read and send it with your name there?"
The point is that if you didn't write the email, or if the recipient isn't going to read it anyway, we've already lost the battle for validation that communications come from humans and contain genuine creative thought.
"People write emails that nobody reads," Golev says. "It's just crazy."
Will Machines Talk to Machines?
The conversation touched on a future that sounds almost absurd but is probably inevitable: agents writing sales proposals that other agents review, all running through traditional procurement systems that were designed for human interaction.
"Are we going to have agents writing for us sales proposals and then other agents reviewing the sales proposal?" Golev recounts from another discussion. "By the way, it's already happening."
But Golev argues this misses the bigger picture. The future won't be agents mimicking human workflows. It will be AI talking to AI in AI language, cutting out the theatrical performance of documents that nobody reads.
"It's not going to be like this," Golev predicts. "Agents are not going to write full-scale proposals and other agents read them and run through traditional procurement systems. It's going to be AI talking to AI in AI language, much shorter, saving tokens, cutting the rubbish out or introducing more rubbish, but it's going to be completely different."
He's been working with data exchange systems since 1998 in e-commerce. Automate the interface, cut out the middleman, radically transform the process.
"I hope we can cut the middleman out and that's probably what's going to happen, and the world will be a completely different place," he says. "I wouldn't be afraid of it. I think we need to embrace it."
An Incentive to Optimise
There's a silver lining: the A365 licensing model might actually push Microsoft to optimise AI performance.
If Agent 365 licences abstract away from the consumption layer so customers don't have to count each credit, but there's still modelling underneath that determines how much the agents can use, then Microsoft would benefit financially from making the agents more efficient.
"That would give Microsoft the incentive: okay, we can now get more money if we optimise the usage of how these agents perform there in the background without making all the messages, calls, and API calls so transparently visible to the customers in the Copilot credit way," Niiranen explains.
The current model, where customers see every API call and credit consumed, creates perverse incentives. Nobody wants to use a service where they have to worry whether their task will take 10 API calls or 100. That's not how enterprise software should work.
"If I give a task to my agent, it's not up to me to figure out is it going to do 100 API calls or 10," Niiranen says. "If I'm an app maker or automation maker today I can affect that, but once I give it to a prompt to an LLM, then it's out of my hands. Someone else then has to manage that, optimise that."
A per-agent licensing model could put that responsibility squarely on Microsoft's shoulders, where it belongs. The company would need to ensure agents complete tasks efficiently to preserve margins on the licence fees. That's a healthier dynamic than passing the optimisation burden to customers who have no control over it.
"There needs to be a way to leave that in the engine room and make it something that could be covered by a new licence type like per-agent licence," Niiranen says.
Ignite and After
Microsoft Ignite runs from 18 to 21 November, and the company is expected to provide official details about Agent 365 during the conference. The industry will be watching for answers to fundamental questions:
What exactly qualifies as an agent under the A365 licence?
How much will the licences cost?
Will existing Copilot agents migrate to this licensing model?
How will consumption be metered and billed?
Can agents be shared across users, or does each user need their own instance?
How do third-party marketplace agents fit into the pricing?
What reporting and telemetry will be available?
What governance controls will exist?
The answers will determine whether Agent 365 works or just adds more confusion.
Golev and Niiranen plan to host another conversation after Ignite, once more tangible information becomes available. The focus will remain on the economics of agents, because that's what ultimately determines whether enterprises can adopt this technology or continue to sit on the sidelines, waiting for clarity that may never come.
"That's what ultimately worries people in business," Golev notes. And he's right. The technology might be revolutionary. The vision might be compelling. The agents might actually work. But if nobody can understand what they'll cost, it may all stay on the shelf.