I help M365 admins & leaders get real value from Copilot and AI. 16-time Microsoft MVP + RD. Host, The CollabTalk Podcast. Daily practical tips. Sometimes funny
Why you should follow me
Joined May 2008
- Tweets57.3K
- Following10.4K
- Followers22.1K
- Likes22.3K
Microsoft 365 admins: the gap between tenants that wired Copilot into real systems and those still stuck on chat-only pilots is getting wider.
A new Copilot Connectors usage report is rolling out in the Microsoft 365 admin center through November 2026 (Roadmap ID 519571).
It shows connector activity, connector-powered response patterns over time, and user-level adoption.
Three numbers worth pulling on day one:
1. Which connectors actually get used
2. Which ones return answers people keep
3. Which users never hit a connector at all
If Copilot still only sees Microsoft 365 content where you are, this report is worth opening. What's the first connector you'd turn on?
Every agent in Microsoft 365 needs a named human owner. "The AI did it" can't be the answer when an agent sends the wrong email or approves the wrong thing.
Agents now run tasks in the background, work across apps and hand work to other agents without someone approving every step. That's less a tool and more a coworker, and coworkers raise questions tools never did.
The one I'd push hardest: train people to manage agents, not just use them. Delegating clearly, checking the result and knowing when to step back in is closer to supervisor training than software training.
The other five, from owners and identities to deciding which calls stay with people:
linkedin.com/pulse/6-ways-ge…
Microsoft says swapping the model barely moved its agent pass rates. Swapping the harness did: up 10.8 points with an OpenAI model and 22.1 with an Anthropic model, on the same tasks.
That's the part of the Microsoft Copilot harness rename worth paying attention to. After a failed tool call, agents on the new harness kept working the problem in 84 to 91% of cases, versus under 7% on the standard harness.
It matches what I've seen running 30+ bots of my own. Swapping models changed results a little. Fixing retries, memory and triggers changed them a lot.
Read the footnotes, though. These are Microsoft internal numbers, and an agent that keeps trying can still fail the task. Test it against your own work.
My take on the harness rename, the numbers, and what Microsoft 365 admins should test:
buckleyplanet.com/2026/11/th…
Microsoft 365 admins: starting in November preview, Copilot Chat's model selector can include usage-based billing models (Roadmap ID 571400).
General availability targets December 2026. Admins allocate credits first. When a user runs out, they can request more from the tenant.
Two places this shows up in real work:
1. A pilot that quietly burns the credit pool on one team
2. A finance review that asks why Copilot spend jumped with no seat change
Seats were the old scoreboard. Credits are the new one.
Have you set a credit request path yet, or is that still informal?
Hours from prompt to backend is progress. Running that generated app in a real org is another job.
Before launch, someone still has to own it, approve who gets access, and know who takes the support call when it breaks. That part doesn't generate itself.
Describe the app you want. Rayfin and GitHub Copilot generate your full backend in code- schemas, relationships, and access policies. No manual wiring. See how it works. youtu.be/rvP1bpmPE2k
Build production-ready enterprise apps in hours, not months. Describe the app you want using Rayfin's open-source SDK with GitHub Copilot, and generate your full backend in code — schemas, relationships, and access policies included.
Deploy to Microsoft Fabric with a single CLI command and immediately inherit enterprise data security, identity controls, and audit compliance already in place across your data estate. #Rayfin #AppDevelopment #MicrosoftFabric #GithubCopilot #AIDevelopment
Microsoft 365 admins: in December 2026, Work IQ memory can expand to agents through a new MCP path (Roadmap ID 578869).
Authorized agents get a user's memory profile: custom instructions, saved memories, relevant work account data, and memories from chat history.
That personalization is helpful. It also means an agent can inherit more than a prompt.
Three controls worth deciding before it lands:
1. Which agents are allowed to read memory
2. Who approves that access
3. How a user clears or limits what agents can see
Does your agent policy say anything about memory yet?
The Information: Governors including Texas want data-center moratoriums. Nvidia looks at AI servers in existing telecom buildings for fewer permits.
Before betting on a new AI footprint:
1. Name who owns capacity and power
2. Check local permit and moratorium risk
3. Prefer reuse of existing sites
4. Tie expansion to a measured workload
Exclusive: Governors elected Tuesday, including in Texas, have called for moratoriums on new data centers.
Nvidia's plan to put AI servers in existing Brightspeed telecom buildings would likely require fewer permits than building from scratch.
Read more: thein.fo/4AjtXQm
Your speakers and sponsors want the event to succeed, but a lot of them won't write promo copy from scratch. Hand them a share kit: posts, images, the link, and a request for a 30-second phone video. Fellow MVP Daniel Glenn's tip: one post per session naming the speaker. That's twenty promoters.
Full recap, slides, and the 3-page organizer handout from my MGCI session: buckleyplanet.com/2026/10/fi…
Most governance policies were written for static content and predictable behavior. Copilot and AI agents are neither. That gap was the core of my conversation with @m365show earlier this year, and it still holds. Worth a listen if you missed it. Link below.
Here's the episode, Your Governance Policies Were Not Built for AI: m365.fm/show/episodes/your-g…
![AI Governance with Christian Buckley [MVP]](https://pbs.twimg.com/card_img/2118811186707779585/UFxn8Y81?format=png&name=800x419)
TechCrunch: AI usage keeps climbing, but spending has stalled as big buyers cut prices and tighten controls.
Before any team scales another AI seat wave:
1. Name who owns the spend
2. Cap weekly token or seat growth
3. Review idle seats every 30 days
4. Tie renewals to a measured outcome
AI usage keeps climbing, but spending has stalled. Price cuts and smarter cost controls among the biggest spenders could challenge investor assumptions about AI hypergrowth.
spr.ly/6019BHNzY7
