Notes from ESPC25

ESPC25 in Dublin has always had a bit of a “family reunion” feel for me as I’ve been attending every year since they started in 2011 (Berlin). But this year it also felt like a checkpoint. Two weeks after Ignite 2025 in San Francisco, the Day 1 keynote pulled a lot of those Ignite announcements into a single, coherent story: Microsoft 365 is becoming a place where people, data, and AI agents work together by default, and the company is trying hard to make that future feel both useful and governable.

ESPC’s Kevin Monaghan opened the week with the kind of welcome that reminds you why ESPC has lasted this long. The stats alone show the scale: 1,750 attendees from 52 countries, with almost 50 Microsoft staff traveling in from Redmond and elsewhere. But the tone wasn’t “big conference brag.” It was more like: you’ve all invested time and money to be here, so let’s make it worth it.

He used a metaphor that landed better than most tech metaphors do: the Irish uilleann pipes. They evolved from being played by mouth to being played by elbow, a big technical shift that didn’t change the truth underneath: mastery still takes time, apprenticeship still matters, and community is how skills actually spread. It was a nice way to frame the week: tools change fast, but the work of learning them (and helping other people learn them) stays stubbornly human.

Then Microsoft’s Jeff Teper took the stage, and the keynote shifted into what felt like the central theme of ESPC25: the platform is moving quickly, but Microsoft is trying to keep the foundation stable enough that customers can bet their work on it.

The part nobody expects: most effort isn’t “AI”

One of Teper’s most interesting points was also one of the least flashy. He said their resource split is roughly one-third on AI and two-thirds on the “core.” In other words: security, reliability, performance, accessibility, and the everyday workflows people depend on.

He put it in reliability terms: moving from “three nines” (99.9%) to “four nines” (99.99%). That tiny decimal shift sounds boring until you translate it into real life: fewer outages, fewer weird glitches, fewer moments where the tools you need for a meeting just don’t cooperate. It was a reminder that AI can’t save a platform that users don’t trust. If the basement is shaky, nobody wants a fancy new penthouse.

He also described a tension Microsoft is clearly hearing from customers: leaders want productivity gains, but employees are already maxed out. Organizations want speed; people want relief. The goal, as he framed it, is to “work smarter, not harder,” and to use AI to reduce overload rather than add to it.

Three leaps in AI that change what’s possible

Teper’s overview of the AI “why now” story was refreshingly simple. He described three big jumps:

1. Pre-training scale: models growing from millions to billions to trillions of parameters, unlocking natural language capability that’s now almost taken for granted.
2. Reinforcement learning: teaching models to get better at specific tasks, like coding, building spreadsheets, or producing structured plans.
3. Reasoning models: giving models time to plan and evaluate their own work, which turns “answer this” into “run a multi-step workflow.”

That last point matters for Microsoft 365 because it’s where “Copilot as a chat tool” becomes “Copilot as an assistant that can actually do work while you do other things.” It also tees up the larger story they kept returning to: the move from one-on-one AI help to teams of humans plus teams of agents.

Copilot’s differentiator isn’t just the model

A lot of AI conversations still get stuck in model-name comparisons. The keynote tried to step over that trap.

Yes, they highlighted the portfolio: GPT-5 adopted the day it released, plus partner models like Anthropic, plus Microsoft-built models powering specific experiences in Teams and elsewhere. And yes, they’re leaning into a future where you can pick the best model for the job.

But Teper argued that the real differentiation is what they called WorkIQ: the layer that brings your organization’s data into the reasoning process, securely, inside your tenant, without using it for training. This is the “your org’s memory” concept: Copilot becomes more than a chatbot because it can interpret what’s happening across your meetings, mail, files, and collaboration history, while still respecting permissions.

That’s the theory. Then they demoed it.

The Copilot app, Model Choice, and “Agent Mode” in Office

Michaela Barrett’s demo was structured around something everyone recognizes: the panic of pulling together a QBR at the last minute.

She started in the Copilot app, positioning it as an “all-in-one” hub with four modules: Chat, Search, Create, and Notebooks. From there, she asked Copilot for action items, then for the QBR workstream leads, then asked it to check email and confirm whether all the QBR docs had arrived.

The point wasn’t that Copilot can fetch names. The point was that it did it as if it understood the work, not as if it was keyword-searching. That’s where WorkIQ was positioned as the engine: it reasons over your work graph and pulls what matters without requiring you to hunt, attach, or paste content.

Then came Model Choice, which they framed as a direct response to customer feedback: “give me more control.” In the demo, Michaela chose GPT-5 specifically because she wanted fast, accurate synthesis across lots of internal information.

After that, the workflow changed from “answer questions” to “build something,” and that’s where Researcher and Agent Mode entered.

Researcher was described as the tool for deep, multi-step tasks: digest multiple sources, follow guidelines, ask clarifying questions like a colleague would, and produce a leadership-ready report.

Agent Mode was the bigger “aha” moment. It was presented as Copilot becoming a true co-editor inside Office apps like Word, Excel, and PowerPoint. The phrase that stuck was basically: tell the app the outcome, and AI fills in the steps. In Excel, that meant generating tables, formulas, and even visual elements as real, editable spreadsheet objects, not static “here’s what you could do” advice. She called it “vibe working,” which got a laugh because it’s exactly the experience: you describe what you want, and the tool starts constructing it beside you.

This is a subtle but important shift in how Microsoft wants people to think about Copilot. It’s not just a chat window that answers. It’s increasingly a set of capabilities that manipulate the same native objects you do: cells, tables, docs, slides.

Teams as the place where people and agents actually collaborate

After the Office-focused segment, Teper pivoted to the obvious reality: work isn’t solo. If Copilot stays one-person-at-a-time, it will never match how organizations actually operate.

Teams was framed as the hub where multi-user, multi-agent work happens. They summarized Teams priorities as:

• More intelligent (AI in the flow of work)
• Simpler and faster (for novices and power users alike)
• Secure (internal and external collaboration)

They also called out recent UX upgrades, like threaded replies inside channels, customizable chat/channel views, and performance tuning. Not glamorous, but meaningful. Better threads matter even more when agents join the conversation, because otherwise everything becomes an unreadable stream.

Ad-hoc group chat with Copilot

The first Teams demo was about turning a one-on-one Copilot thread into a group chat. You can select which history items to bring over, add participants, and Copilot joins as an equal collaborator.

One detail that matters: the security checkpoint. If Copilot is about to share something one participant has access to but another doesn’t, it asks before revealing it. That’s not just a “nice compliance feature.” It’s the difference between “we can try this” and “legal will shut this down.”

Channel Agents and cross-tool retrieval

The second Teams agent was the Channel Agent, described as something every channel is “born with.” It knows the SharePoint site and group behind the channel, so it can act like a durable teammate for that project space.

Then they pushed it further: cross-service integration using MCP (Managed Connector Protocol). The demo used Jira: ask for the latest “Yes-labeled” issues, get a threaded list of blockers, then ask the agent to schedule a meeting with “Finn and Mario.” The agent inferred the right people, checked availability, and proposed a time, without Alex needing to jump into calendars.

That’s the pattern Microsoft is aiming for: less tab-switching, more “solve it where I’m already working.”

Facilitator: the meeting agent people will actually use

Finally, they highlighted Facilitator, a meeting agent that shows up like a participant in the roster. It extracts agenda items from chat and comments and displays a timed agenda bar at the top of the meeting. You can edit it on the fly (“add 30 minutes for success metrics”), and it recalculates the meeting plan.

Then it does the things people secretly want most: creates a launch plan document based on the discussion and produces live meeting notes that attendees can edit. In other words, it tackles the annoying parts of meetings: structure, capture, and follow-through.

If Channel Agent is about project rhythm, Facilitator is about meeting hygiene.

SharePoint’s new role: grounding, organizing, and “knowledge work”

SharePoint has always been the “S” in ESPC, and Teper made sure it didn’t feel like an afterthought. He emphasized SharePoint as the flexible content platform under everything: sites, libraries, OneDrive integration, Lists, and even Copilot Pages.

He also mentioned a major milestone: SharePoint in the cloud passing one billion active users, boosted by moving consumer OneDrive onto SharePoint infrastructure and expanding SharePoint as an ISV platform through embedding.

Then Zach demoed the SharePoint Knowledge Agent, which may be the most practical “AI in SharePoint” pitch I’ve seen: not Q&A magic, but organization.

The Knowledge Agent sits in the bottom-right corner of SharePoint pages and libraries and changes what it offers based on context: summarize a page, create an FAQ, organize selected files, and so on. It can also help improve a site by identifying broken links, stale pages, and content gaps based on what people are searching for but not finding.

The hands-on example was expense spreadsheets dumped into a library. The Knowledge Agent extracted transaction date, vendor, and total spend, wrote that metadata back to the files, and then let the user create an agent grounded in that library (“Zaba expenses”). Add that agent to a Teams chat, and now the whole team can ask questions like “How many expenses were there in Q3?” and get structured answers based on the metadata, including new files as they arrive.

This is a big deal because it flips the typical dynamic. Instead of humans doing tedious tagging and cleanup so systems become searchable, the system does the cleanup so humans can work.

Governance: Agent 365 and the “billion agents” problem

The keynote ended where most enterprise conversations eventually end: governance.

They said customers have already built over a million agents, and analysts expect more agents than people in the workforce. That’s believable. If agents become cheap to create, they will proliferate the same way Teams channels and SharePoint sites did.

So Microsoft announced Agent 365 as a governance plane: a single place to inspect agents, control what apps/data/systems they can access, and track usage and compliance.

That last part is the part that will decide whether this era goes smoothly. If every team can spin up agents the way they spin up chats, IT needs a way to keep things safe without becoming the “Department of No.”

What I’m taking away

Much like the messaging out of Ignite in November, ESPC25’s Day 1 keynote didn’t feel like a “here’s one new feature” keynote. It felt like Microsoft trying to make a broader argument:

• Copilot is moving from chat to work execution.
• Agents are moving from personal helpers to shared teammates.
• Teams is becoming the collaboration layer for people plus agents.
• SharePoint is becoming the grounding and organization engine for AI.
• Governance is being treated as a first-class requirement, not an afterthought.

And framing it all, back to those uilleann pipes: the tools are changing fast, but mastery still matters. The organizations that win won’t just be the ones who turn features on. They’ll be the ones who learn how to play the new instrument well, together, without letting the sound turn into noise.