Microsoft struggles to address AI notetaker governance nightmare

05/07 18:15 - Microsoft struggles to address AI notetaker governance nightmare
Microsoft this week tried to address the growing challenges surrounding notetaker bots in meetings by giving IT better control over them. Microsoft’s announcement said that users of Microsoft Teams will be able to block non-Microsoft bots “even in meetings where organizers allow participants to bypass the lobby.” When the feature is enabled, Teams automatically detects potential bots, places them in the meeting lobby, clearly identifies them, and prompts organizers to confirm admission, Microsoft said, and even in meetings where organizers allow human participants to bypass the lobby, bots identified through this new policy will continue to require approval before joining. “We’ve strengthened Teams’ ability to distinguish between bots and human participants as they join a meeting,” the company said. “Teams now uses a combination of behavioral and infrastructure signals to identify bots with a higher degree of accuracy. Alongside these improvements, soon we’ll introduce a registration path for independent software vendors (ISVs) that build meeting experiences for Microsoft Teams.” The underlying problem with the strategy is more complicated, however. Although AI bots launched by the meeting owner are typically announced at the beginning of a call, and participants’ bots announce themselves as the attendees log in, alert fatigue is diluting how carefully people watch what they say during those meetings. But the thornier issue is that meeting owners’ approval of their own bot notetakers typically happens right before the start of a call, and the host has no control over whether participants also introduce their own AI notetakers. And even if the intended topic of a call was innocuous, if someone brings up something that needs to be kept secret, such as plans for a hostile takeover or discussion about firing an employee, that is duly recorded by every bot. This expands the threat surface and increases the ways sensitive data could leak. Doesn’t rein in Microsoft bots Analysts and consultants agreed that any effort to restrict notetaking apps is good for enterprise IT, but some questioned whether the Microsoft effort went far enough. “Although this new capability is useful to prevent external bots from attending recurring meetings even if they were needed for just one instance, it doesn’t seem to me that it does anything to prevent Microsoft’s own bots from doing so,” said Flavio Villanustre, CISO for the LexisNexis Risk Solutions Group.  Indeed, the Microsoft statement solely talks about managing “external bots and their access to meetings.” In fact, Gartner VP analyst Nader Henein said the limited controls that Microsoft is offering may actually dilute IT’s ability to control access to sensitive information. Allowing any additional AI notetaking “takes the option to restrict/redact off the table,” and that control is what he thinks IT leaders should demand. The only practical way to do that is to allow only one notetaking app for any meeting and it needs to be controlled by the meeting owner. “Allowing attendees to ask for an AI summary from the meeting owner and giving the owner the capacity to provide different versions that potentially shield sensitive data is a better choice for organizations looking to support better meeting follow ups without adding more work on the meeting owner,” Henein said. “It could even be set up in advance so a ‘sanitized’ summary is available for download.” Tom Findling, CEO of Conifers.ai, agreed with Henein and argued that these notetaking app controls have to be centralized with IT. “Microsoft basically built a bouncer for meeting bots and that is a good thing. But the real risk shows up later, when a normal meeting turns into M&A, legal, HR, or board-level discussion while an AI notetaker is still running,” Findling said. “Now that transcript may be sitting in a cloud nobody approved. You do not fix that live. You fix it upfront. For legal, finance, HR, and exec meetings, external AI notetakers should be blocked by default unless explicitly approved.” Existing governance not enough Sanchit Vir Gogia, chief analyst at Greyhound Research, said the slowly changing nature of AI notetakers has allowed them to slip by IT governance rules. “A meeting note was once a harmless aid. It is now a searchable corporate record that can hold intent, allegations and material non-public information. Once a conversation is transcribed and saved, it has left the room, and it begins to travel through mail, search, and discovery with a life of its own,” Gogia said. “Microsoft’s control is useful, but should not be oversold. It detects external bots and puts them before an organizer for approval. It does not yet block them, and approval at the lobby is not a governance model. Capture also arrives by routes the lobby never sees, through browser extensions and personal devices.” Gogia also argued that the inevitable errors in these bot-generated transcripts or summaries, whether caused by hallucination or simply incorrect interpretation what was actually said, is also a massive risk. “AI summary does not merely create a record. It creates an authoritative-looking one that is often wrong and, in doing so, it inverts the burden of proof. Once a summary exists, the question shifts from proving what was said to disproving what the machine wrote,” Gogia said. “A tentative ‘we should look at acquiring them’ can harden into ‘we agreed to acquire them’ and that version becomes the default until someone corrects it.” And, noted Justin Greis, CEO of consulting firm Acceligence, the problem will only get worse as AI summary generators morph into agentic systems, with action-taking autonomous agents. “Over the next few years, we’ll see AI agents that summarize, extract decisions, assign work, update business systems, prepare follow-up documents, and collaborate with other AI systems after the meeting ends,” he said. “In fact, we are already seeing that integration happen, and it is simultaneously incredibly valuable and outrageously risky. The real question isn’t whether to allow an AI notetaker. It’s how organizations will govern an increasingly machine-readable workplace.” Greis said that he sees the Microsoft approach as a good start, “because they’re treating AI participants more like digital identities than software features.” He pointed out, “detection, verification, explicit admission, auditability, and policy-based control are exactly the kinds of enterprise controls we’ll need as AI agents become commonplace. This feels very similar to identity and access management twenty years ago. We eventually realized we weren’t managing employees, we were managing identities. AI agents deserve the same treatment.” ...