Latest / Microsoft 365 Roadmap Rundown / Human Gates, Identity Recovery, and the Great Agent Administration Era

Transcript

Human Gates, Identity Recovery, and the Great Agent Administration Era · Microsoft 365 Roadmap Rundown

• 0:00
  Speaker 1: Automatically block identified external bots from joining meetings.
• 0:03
  Speaker 2: The bot bouncer. It is a fantastic direct governance action.
• 0:07
  Speaker 1: I see the immediate security value. Absolutely. Yeah. But I also see a massive change management disaster just waiting to happen.
• 0:14
  Speaker 2: Oh yeah. Welcome
• 0:19
  Speaker 3: to the Roadmap Rundown.
• 0:21
  Speaker 1: The podcast that keeps you up to speed on what's new, what's changing, and what's coming next across Microsoft 365.
• 0:28
  Speaker 3: Let's get into it.
• 0:32
  Speaker 1: Imagine having this, highly efficient, abs absolutely brilliant personal assistant, like works at lightning speed.
• 0:39
  Speaker 2: Right.
• 0:40
  Speaker 1: But they have absolutely zero common sense and they will confidently hand your company's most sensitive, completely unverified data to anyone who asks nicely.
• 0:50
  Speaker 2: Just here you go, here's the payroll data.
• 0:52
  Speaker 1: Yeah, exactly. Welcome to the Roadmap Rundown where we take the raw unfiltered source material of enterprise software updates and we translate it into what you actually need to know.
• 1:03
  Speaker 2: Yeah. Are analyzing the source material so you can spend less time, you know, decoding all that dense release language.
• 1:09
  Speaker 1: And more time deciding what actually matters for your environment. Today we're looking at a massive underlying theme across the board, which is control points for AI and automation.
• 1:19
  Speaker 2: Because it's really the era of supervised autonomy now. I mean, we spent the last few years giving these systems incredible privileges, right, to read, summarize, generate content.
• 1:29
  Speaker 1: Yeah.
• 1:30
  Speaker 2: And now looking at the source material, there's just this frantic effort to build the guardrails necessary to keep those autonomous systems from driving the organization right off a cliff.
• 1:39
  Speaker 1: Yeah. But before we get into the, you know, the dense governance controls, I wanna start with a rollout that users are genuinely gonna care about.
• 1:47
  Speaker 2: Always good to start with the fun stuff.
• 1:49
  Speaker 1: Right. So Microsoft three sixty five Copilot is officially integrating with Power BI.
• 1:54
  Speaker 2: Oh wow! Okay.
• 1:55
  Speaker 1: Yeah, this is Roadmap ID 567800And91. It's slated for general availability in August of calendar year twenty twenty six. Copilot will basically be able to reason over Power BI enterprise data and return grounded answers directly from reports and semantic models.
• 2:11
  Speaker 2: Which is a fundamental shift in how employees interact with business intelligence. Mean, think about it. Instead of forcing a user to navigate this, like, labyrinth of workspaces
• 2:21
  Speaker 1: December of the third quarter revenue dashboard is in the finance final folder or
• 2:24
  Speaker 2: Or the marketing sandbox. Right. They can just query the data using natural language. It's huge.
• 2:29
  Speaker 1: Right. It is. But let me pause on a technical detail there for anyone listening who maybe doesn't build dashboards every day. The source material specifically mentions copilot reasoning over semantic models. Yeah.
• 2:43
  What does it actually mean for an AI to interact with a semantic model rather than just, I don't know, reading a raw database?
• 2:50
  Speaker 2: Well, semantic model is basically the translation layer between raw unformatted database tables and human business logic.
• 2:58
  Speaker 1: Okay. So it's interpreting the raw data.
• 3:00
  Speaker 2: Exactly. Because a database might just have a column labeled, cuss status code four.
• 3:05
  Speaker 1: Which means nothing to a normal person.
• 3:07
  Speaker 2: Right. But the semantic model defines that code as an active customer. So when Copilot reasons over the model, it's relying on those predefined relationships to actually understand what the numbers mean before it generates an answer.
• 3:20
  Speaker 1: So it's not just handing the AI a raw spreadsheet, it's handing it the company's official dictionary.
• 3:25
  Speaker 2: Yes, perfect analogy.
• 3:26
  Speaker 1: But okay, here's where my analytical brain sees a massive operational flaw with this.
• 3:30
  Speaker 2: Uh-oh, let's hear it.
• 3:31
  Speaker 1: What if that dictionary has like 17 conflicting definitions of an active customer because different departments built their own models?
• 3:40
  Speaker 2: Then you run face first into the dry unforgiving reality of enterprise data.
• 3:44
  Speaker 1: Right.
• 3:45
  Speaker 2: If the underlying semantic models are a chaotic mess, this highly advanced AI is just gonna confidently hand you the wrong book. I mean, Copilot cannot magically rescue a report that an analyst hastily stitched together for a one off meeting back in 2023.
• 4:01
  Speaker 1: Yeah. And that's the danger, right? Users might look at a beautifully formatted natural language summary from Copilot and assume it's totally authoritative.
• 4:10
  Speaker 2: When it's actually summarizing deeply flawed, outdated metrics.
• 4:13
  Speaker 1: Exactly. So organizations really have to use this rollout as an excuse to lock down their model ownership and metric definitions.
• 4:20
  Speaker 2: Yeah. The Surface presentation is pristine, but the readiness is entirely a data governance problem. You cannot unleash a reasoning engine on a data swamp and expect, you know, a crystal clear well of insights.
• 4:31
  Speaker 1: Man, that terrifying thought leads perfectly into our next major item.
• 4:34
  Speaker 2: Let's do it.
• 4:35
  Speaker 1: So if the AI has access to all this data and is capable of taking autonomous actions based on it, how do we stop it from say, firing off an email or closing a financial ticket based on a faulty semantic model?
• 4:48
  Speaker 2: We install a hard stop.
• 4:50
  Speaker 1: Right.
• 4:50
  Speaker 2: This is Roadmap ID 570434. It introduces Copilot Studio human approval gates for agent tool calls. General availability is set for September 2026 and Microsoft describes this as a deterministic guardrail which is a really vital phrase.
• 5:07
  Speaker 1: So unpack deterministic guardrail for listener. How does this, this per tool, per agent toggle actually function mechanically in the interface?
• 5:15
  Speaker 2: Well, it means the pause is architectural. Right? It's not behavioral.
• 5:18
  Speaker 1: Meaning you aren't just putting a line in the agent's prompt saying, please be very careful and ask permission before spending money.
• 5:23
  Speaker 2: Exactly. And just hoping the AI listens. No. It is a strict system level gate. So if an agent attempts to execute a sensitive action like processing a payment or modifying a live support ticket, the execution physically pauses.
• 5:36
  And the user sees an in line approval request surface natively like right in Teams or M365 Copilot detailing exactly what the agent intends to do.
• 5:46
  Speaker 1: And then the user can approve it, approve it for the session or deny it.
• 5:50
  Speaker 2: Right.
• 5:50
  Speaker 1: But I have to question the operational reality of this. Are we just shifting the liability to a tired employee?
• 5:58
  Speaker 2: What do you mean?
• 5:58
  Speaker 1: Well, if we prompt an employee 50 times a day to approve automated actions, they're gonna develop alert fatigue.
• 6:05
  Speaker 2: Oh, totally.
• 6:06
  Speaker 1: They stop reading the details and just blindly click approve all to clear their notifications and get back to work.
• 6:12
  Speaker 2: Alert fatigue is a massive risk, but, you know, the UI structure attempts to mitigate that by specifically detailing the intended action rather than just asking for a blanket, allow this app to run permission.
• 6:24
  Speaker 1: Okay. That's fair.
• 6:24
  Speaker 2: But your point is exactly why this is a per tool toggle. It forces administrators to map out their agent behaviors meticulously. You have to sit down and decide which of these actions are harmless and reversible and which ones actually deserve a human checkpoint.
• 6:37
  Speaker 1: So if an admin gates an agent drafting a routine meeting summary, they're just causing alert fatigue.
• 6:43
  Speaker 2: Exactly.
• 6:44
  Speaker 1: But if they only implement the gate when the agent attempts to email a finalized contract to an external vendor, the human approval retains its psychological weight.
• 6:52
  Speaker 2: Bingo. The configuration dictates the security. The tool just forces you to be deliberate about where you place the toll booths.
• 7:01
  Speaker 1: Now, if we're asking humans to verify these intended actions, the agent really needs to present his logic clearly.
• 7:07
  Speaker 2: It does.
• 7:07
  Speaker 1: Let's look at how the front end experience is evolving to build manipulate that human trust. Roadmap ID 570435 hitting GA in September 2026 aligns Copilot Studio agents really closely with Copilot chat.
• 7:22
  Speaker 2: Yeah, so agents are gaining similar formatting to the native Copilot experiences. We're talking about smoother streaming, retry behavior, voice dictation, and this thing they call a compressed chain of thought.
• 7:33
  Speaker 1: They're also adding tool execution traces and clickable citations. Yep. But I see a psychological trap here.
• 7:39
  Speaker 2: Always the skeptic. Always. If we give a custom agent a highly polished UI complete with a compressed chain of thought and slick execution traces, doesn't that accidentally trick users into trusting it way too much?
• 7:51
  Speaker 1: Presentation bias.
• 7:53
  Speaker 2: Yes. If it looks incredibly confident and professional, a user is much more likely to click that human approval gate without actually verifying the underlying data.
• 8:02
  Speaker 1: Presentation bias is a formidable challenge, yeah. I mean, if the agent confidently explains its chain of thought, even if that thought is based on some outdated 2021 HR policy, the user might just accept the flick presentation as truth.
• 8:17
  Speaker 2: Exactly.
• 8:18
  Speaker 1: However, from a purely administrative and troubleshooting standpoint, those tool execution traces are an incredible addition.
• 8:25
  Speaker 2: Because it lets the help desk show the agent's work.
• 8:27
  Speaker 1: Precisely. When an agent inevitably returns some bizarre response a technician can look at the tool trace and pinpoint the exact moment the logic failed. It turns a black box into a glass box.
• 8:38
  Speaker 2: I like that, a glass box.
• 8:39
  Speaker 1: You can see the gears turning which is essential for refining these custom agents. Well speaking of refining trust, these front end tools are officially pushing into highly regulated environments.
• 8:49
  Speaker 2: This is big news.
• 8:50
  Speaker 1: Yeah. Roadmap ID 569214 brings teams copilot chat to chats, channels, calls, and meetings across GCC, GCC high, and DOD environments. That's coming in October 2026.
• 9:02
  Speaker 2: For any listeners outside the public sector, GCC is the government community cloud and DoD is the Department of Defense environment.
• 9:10
  Speaker 1: Right.
• 9:10
  Speaker 2: These are tenants with the strictest compliance, data isolation, and residency rules on the planet.
• 9:16
  Speaker 1: So seeing native Copilot chat deploy into the actual flow of calls and channels in these environments, rather than forcing government workers to use some detached, separate web tool, that means the backend data handling has finally achieved an incredibly rigorous certification standard.
• 9:31
  Speaker 2: It really has. Releasing these features into the collaboration flow of a DoD tenant requires proving that the AI can interact with highly classified conversations without leaking context or violating data boundaries.
• 9:43
  Speaker 1: That's a massive maturation of the compliance architecture. Huge. But if that compliance architecture fails or if you know a credential gets compromised and a malicious actor corrupts a critical identity, the blast radius is devastating.
• 9:57
  Speaker 2: Which brings us to the ultimate structural guardrail.
• 10:00
  Speaker 1: Yes. Roadmap ID 56700885. Entra ID Backup and Recovery.
• 10:06
  Speaker 2: Slated for GA in December of calendar year twenty twenty six and it provides a rolling thirty day point in time recovery for untra objects.
• 10:14
  Speaker 1: It's essentially version control for human beings.
• 10:16
  Speaker 2: That's a great way to put it.
• 10:18
  Speaker 1: You can roll identity back to a last known good state via the M365 backup native app or the platform APIs. I do want to clarify from the source material though that the existing seven day recovery, the one that comes with P1 and P2 licenses, is already available and operates completely independently of this new thirty day extended safety net.
• 10:36
  Speaker 2: Yeah and having a thirty day window is a monumental upgrade for disaster recovery. But administrators really have to maintain a healthy skepticism about how restoring an object actually functions mechanically in a chaotic enterprise environment.
• 10:48
  Speaker 1: Oh, for sure. Demo candidates always made it look like a magic wand. You just click a button and the deleted user is instantly restored happily back at work.
• 10:56
  Speaker 2: In reality, an identity is this deeply complex web of relationships. You're dealing with group memberships, conditional access policies, downstream application dependencies.
• 11:06
  Speaker 1: It's messy.
• 11:07
  Speaker 2: Very. Imagine a scenario where an administrator uses this tool to roll a user's object back by fourteen days because of a corruption issue. What happens mechanically to all the legitimate access requests that users submitted and was granted in the last two weeks?
• 11:22
  Speaker 1: Well, they theoretically lose those permissions.
• 11:23
  Speaker 2: Right? And
• 11:25
  Speaker 1: would third party HR or finance systems gracefully accept this sudden rollback in state? Or would they just throw massive synchronization errors because the user profile suddenly looks completely different?
• 11:36
  Speaker 2: That is the challenge of defining a last known good state. A thirty day window is a brilliant capability, no doubt. But IT teams must rigorously test the downstream impact of a rollback in realistic, messy scenarios.
• 11:49
  Speaker 1: While identity is just one vector, managing the actual flow of data across the environment requires much more nuanced tools.
• 11:56
  Speaker 2: Yeah, let's talk about that.
• 11:57
  Speaker 1: So let's dig into how administrators are maintaining control over this shifting landscape. Roadmap ID 569004And77 updates Purview Data Loss Prevention or DLP in September 2026.
• 12:11
  Speaker 2: Purview DLP is adding specific controls for exchange classification failures.
• 12:16
  Speaker 1: Explain the mechanics of a classification failure for us. Why does an admin need specific controls for this rather than just a general failed policy?
• 12:24
  Speaker 2: Because a classification scan can fail for completely different reasons. I mean, a scan might simply time out because the email attachment is enormous and the server is under heavy load.
• 12:34
  Speaker 1: Okay.
• 12:34
  Speaker 2: Better. Or the system might hit a throttling limit. But alternatively, the scan might completely fail to extract any text because the file is corrupted or it's encrypted in a really strange way.
• 12:44
  Speaker 1: So treating a simple timeout the exact same as a totally unreadable file Mhmm. Is basically a sledgehammer approach.
• 12:50
  Speaker 2: Exactly. If a scan times out, an admin might configure the DLP policy to just briefly delay the email and retry. But if the file is completely unreadable, the policy might need to quarantine the message immediately and alert the security team. It just allows for highly nuanced compliance routing instead of disrupting the whole business over a simple server timeout.
• 13:11
  Speaker 1: Makes total sense.
• 13:12
  Speaker 2: Okay.
• 13:12
  Speaker 1: And managing the data from human conversations is also getting an AI overhaul. Roadmap IDs 569000And17 and 568933, both hitting GA in September 2026, bring intelligent AI recaps and a Copilot chat sidecar to Teams queues.
• 13:28
  Speaker 2: This completely changes the mechanics of post call follow-up for customer service teams. How so? Instead of an agent manually scrubbing through a sprawling twenty minute unstructured call transcript to figure out, you know, what discount was promised to a client, they simply query the Copilot sidecar. The sidecar queries the transcript locally and extracts the specific commitment.
• 13:47
  Speaker 1: Well, speaking of bots analyzing meetings, let's look at how we control third party bots trying to join those same calls.
• 13:53
  Speaker 2: Well this is a fun one.
• 13:54
  Speaker 1: Roadmap ID 566201 allows teams to automatically block identified external bots from joining meetings.
• 13:59
  Speaker 2: The bot bouncer. It is a fantastic direct governance action.
• 14:03
  Speaker 1: I see the immediate security value, absolutely. Yeah. But I also see a massive change management disaster just waiting to happen.
• 14:11
  Speaker 2: Oh yeah.
• 14:11
  Speaker 1: An administrator is gonna quietly toggle this policy on and five minutes before a critical client pitch, the CEO's favorite external AI notetaker is gonna get blocked from the meeting.
• 14:22
  Speaker 2: And that administrator will be the one receiving a very loud phone call.
• 14:26
  Speaker 1: Oh, a screaming phone call.
• 14:28
  Speaker 2: Yeah. The technical implementation of a bot bouncer is a single click, but the organizational communication required beforehand takes weeks.
• 14:36
  Speaker 1: Yep. We're also seeing a major fix for the fragmented way these apps are installed. Roadmap IDs 567883 and 569600And8 introduce unified app and agent installation across Teams, Outlook and M365 Copilot rolling out between August and October 2026.
• 14:55
  Speaker 2: This finally closes the classic enterprise side door.
• 14:57
  Speaker 1: What do you mean?
• 14:58
  Speaker 2: Historically, administration was heavily siloed by applications so a security team might rigorously review and block a third party app in Teams.
• 15:05
  Speaker 1: Okay.
• 15:05
  Speaker 2: But because the administration portals were separated a user could simply go and install that exact same app in Outlook.
• 15:10
  Speaker 1: Oh wow.
• 15:11
  Speaker 2: Unified installs mean you configure the policy once and it enforces that boundary across the entire ecosystem.
• 15:18
  Speaker 1: That's great but before administrators build a massive inflexible project plan based on these unified policies and dates we need to look at the volatility of the roadmap itself.
• 15:28
  Speaker 2: Let's play some roadmap roulette.
• 15:29
  Speaker 1: Because release dates are estimates and sometimes features simply vanish into the ether.
• 15:35
  Speaker 2: It is a living document and plans change.
• 15:39
  Speaker 1: Roadmap ID 559600And1 is a perfect example. It was originally slated for GA in June 2026 promising co pilot chat history filtering by experience. It is officially cancelled. Microsoft explicitly apologized for the inconvenience and chose not to move forward.
• 15:55
  Speaker 2: You know beyond marveling at the absurdly corporate name, chat history filtering by experience, the cancellation is actually a useful example of Roadmap Clarity. Yeah. They didn't just quietly delay the feature month after month into infinity, they provided a definitive end to the project. Users were supposed to see their chat history scoped strictly to their specific endpoint, like filtering just for mobile chats with an all chats option.
• 16:17
  Speaker 1: Go, I
• 16:18
  Speaker 2: kill it. The scoping likely created a deeply confusing user experience. Imagine a user remembers having a conversation with Copilot, logs into their laptop, doesn't see it because it was filtered to their mobile experience, and assumes the data is completely lost.
• 16:33
  Speaker 1: Ah, I see.
• 16:34
  Speaker 2: The cancellation saves adoption teams from having to explain a convoluted interface.
• 16:38
  Speaker 1: But a cancellation is clean. A delay can actively sabotage a project plan.
• 16:42
  Speaker 2: Yeah, that's true.
• 16:44
  Speaker 1: Roadmap ID 5610000And26 SharePoint form driven document generation with Power Automate was supposed to hit GA in June. It is delayed by four months to October 2026.
• 16:56
  Speaker 2: This one is going to sting automation teams heavily.
• 16:59
  Speaker 1: Explain how form driven document generation actually works mechanically and why a four month delay hurts so much.
• 17:04
  Speaker 2: It's basically the automated mapping of data inputs from a SharePoint form directly into predefined template fields to generate finalized documents like vendor contracts, customer invoices, offer letters, and storing them natively in SharePoint. Okay. Automation teams rely on this to remove manual data entry errors from highly repetitive legal and financial workflows.
• 17:23
  Speaker 1: So if an IT director planned to sunset their expensive legacy document generation software in August, assuming this native SharePoint feature would replace it, they are now left with a massive gap in their architecture.
• 17:35
  Speaker 2: Which is the ultimate administrative truth. Roadmap dates are planning inputs, not guarantees, not blood oaths.
• 17:42
  Speaker 1: Right.
• 17:42
  Speaker 2: If you are replacing a critical business process like generating vendor contracts, you cannot put your entire operation on pause waiting for a roadmap item to clear preview. You must always maintain a manual fallback process or extend your legacy contracts. The Roadmap is a compass giving you a general direction, not a GPS with a guaranteed arrival time.
• 18:03
  Speaker 1: That's great advice. Let's follow that compass to our final item which might just blow the doors completely off your compliance strategy. Roadmap ID 558933GA in October 2026. Configurable AI archive for Teams meeting knowledge retention. This allows AI generated insights to persist beyond standard transcript retention periods.
• 18:22
  Speaker 2: The compliance implications of this are staggeringly heavy.
• 18:25
  Speaker 1: I have to raise a massive legal red flag here. Let's trace the mechanics of this. Imagine your company policy designed to limit e discovery liability during lawsuits requires the strict deletion of all meeting transcripts after thirty days.
• 18:39
  Speaker 2: Pretty standard.
• 18:40
  Speaker 1: The transcript is destroyed. However, the AI generated insights based on that deleted transcript are configured in this new archive to live on indefinitely. Doesn't this create a shadow compliance nightmare?
• 18:53
  Speaker 2: It is a profound mismatch between legal policy and technical configuration. You are permanently deleting the original source material but you're intentionally keeping the ghost of the data.
• 19:03
  Speaker 1: Right. An AI generated summary of a deleted document is still a discoverable record during an audit.
• 19:08
  Speaker 2: Absolutely. If the legal team mandates a thirty day data life cycle, but an IT administrator configures this AI archive to retain insights for three years simply so Copilot can answer historical questions about past projects, the organization is severely out of compliance. Yeah. Administrators must sit down with their records management teams well before October to define exactly how long the secondary AI archive is legally permitted to exist.
• 19:33
  Speaker 1: The convenience of asking an AI about a project meeting from six months ago does not outweigh the crushing financial penalty of failing a legal audit.
• 19:41
  Speaker 2: Not at all.
• 19:42
  Speaker 1: The governance must strictly dictate the configuration.
• 19:45
  Speaker 2: If you summarize the overarching theme of the source material we analyzed today, it's clear that the control points are finally catching up to the initial hype.
• 19:53
  Speaker 1: We're seeing a distinct shift from look how fast this technology can run to here is how we install the brakes. We have Co Pilot Studio human approval gates stopping rogue tool calls. Yep. We have bot bouncers protecting meetings and unified installations closing policy loopholes across Outlook and Teams.
• 20:12
  Speaker 2: And we have Intra ID Backup and Recovery providing that crucial thirty day extended safety net for the underlying identity architecture.
• 20:20
  Speaker 1: But before we sign off, want to leave you with a final thought to mull over a scenario that ties all these conflicting controls together.
• 20:27
  Speaker 2: The grand contradiction. Let's hear it.
• 20:29
  Speaker 1: Think about the mechanics of what we just discussed. Entra gives us a thirty day window to roll back a compromised identity to a last known good state, effectively wiping its recent permissions and changes from the system.
• 20:42
  Speaker 2: Right.
• 20:42
  Speaker 1: But Teams AI meeting insights can be configured to live on indefinitely well beyond the life of the original transcripts.
• 20:49
  Speaker 2: Oh wow. Okay. I see where this is going.
• 20:52
  Speaker 1: So what happens when an identity is compromised, acts maliciously during a project meeting to introduce false data, and the AI permanently archives those malicious insights?
• 21:02
  Speaker 2: Wow.
• 21:03
  Speaker 1: If you then use Entra to roll back the identity's permissions and wipe its recent actions from the directory, Co Pilot will still leverage those infinitely retained Poison insights generated after the identity state was reverted.
• 21:15
  Speaker 2: Because the AIR type is separate.
• 21:17
  Speaker 1: Exactly. When the source material is gone, the identity is rolled back but the AI's memory persists indefinitely. Where does the truth of your enterprise data actually live?
• 21:27
  Speaker 2: That is the multi million dollar governance question every architecture team is going to have to answer.
• 21:32
  Speaker 1: Something to keep you awake tonight. Thank you for joining us for this Roadmap Rundown. Keep a very close eye on those admin toggles, communicate your policies well before you start bouncing executive bots from meetings, and we will catch you next time as we continue analyzing the source material.
• 21:51
  Speaker 3: Thanks for listening to that Roadmap Rundown. You can find the show notes, roadmap item links, and more at www.roadmaprundown.com. We'll be back next week with another look at what's new, what's changing, and what's coming next across Microsoft three sixty five. Roadmap Rundown uses AI assisted production and synthetic voices based on information from the official Microsoft three sixty five roadmap. See you next time.