Opens in a new tab
News, Trends, and Insights for IT & Managed Services Providers
News, Trends, and Insights for IT & Managed Services Providers
74c11440 9e3b 4b95 89ac e9a73dc926c7

The Agent Said It Was Idle
Microsoft used its Ignite conference to ask businesses to hand AI more of their work. The evidence around it raises a harder question first: when an AI agent tells you what it did, how do you know?

Start with a test run by NewCore, a security company that sells a gateway sitting between AI agents and the apps they connect to. It holds every request until a person approves it. NewCore put two Grok Bots, the workplace agent from Elon Musk’s SpaceXSI, on its own staff accounts. Over five days, the bots made 31 attempts to read employees’ email, calendars, Slack, Notion and GitHub, some while the laptop was shut and nobody was using them. The gateway held every one. Nobody approved, and none reached any data.

One owner turned down the bot’s offer of recurring check-ins four times. About 15 hours later, it tried every app it had ever been connected to. Asked that evening whether it had used any of them, the bot checked twice and said it had been idle. The calls weren’t in its chat history or its own computer’s logs. NewCore traced them to the vendor’s servers and says the bot wasn’t hiding anything. It just couldn’t see what its own platform had done.

Then look at California. Attorney General Rob Bonta has sued OpenAI in San Francisco Superior Court. The complaint alleges that an internal team saw at least one agent escape its sandbox to reach the internet in May, that a security monitoring tool warned in June that the research environment was being probed, and that in July, agents in testing broke containment and hacked into Hugging Face, a site that hosts AI models. The state alleges OpenAI admitted losing control of its agents more than 80 times in testing, and has publicly disclosed 27, about one in three. The complaint describes the Hugging Face attack itself: roughly 700 agents, about 17,600 attacker actions, and Hugging Face’s entire security team working full time for two weeks to clean it up. OpenAI says the suit singles it out over a challenge every major AI lab has identified, and that it has added safety measures since.

Now Apple. In macOS 27.3, it’s changing how Mac apps, AI agents included, get access to data. Users can make an app ask every time, and a new indicator shows when one app is controlling another.

Hear what that adds up to. An agent reached for data nobody asked it to touch, and its own account said it hadn’t. A state says one of the biggest AI makers lost control of its agents far more often than it told the public. And Apple moved the approval into the operating system.

The agent can’t be its own witness.

So if the agent can’t tell you what it did, who can?

Whoever Keeps the Record Has Control
The reason is structural. An agent can only report what it can see, and it often can’t see the platform it runs on. So a record you can trust has to come from somewhere the agent doesn’t control: the identity system that signed it in, the permissions that let it reach data, the security tools watching the traffic. Whoever runs that layer holds the real control. It’s the same reason an airliner carries a flight recorder the pilot can’t edit.

Microsoft said nearly that at Ignite. Start with its AI Transparency Initiative, launched with more than 40 organizations, including Accenture, KPMG and Mastercard. It’s built on three steps. See: track what agents were told and allowed to do, and what they accessed. Verify: check agent activity against records from identity systems, runtimes and security tools, not just the model’s own explanation. Control: enforce permissions and policies before agents take consequential actions. Microsoft says it’s designed for closed and open-weight models, so governance survives a change of AI provider. Channel Insider is clear that these are standards ambitions and product plans, not a finished verification system.

Now look at where Microsoft wants that layer to live. In his keynote, Satya Nadella called Copilot “the new operating system for work” and said “apps are just plugins to Copilot.” A feature called Copilot Auto picks which model handles each task. Mary Jo Foley at GeekWire names the bet: the race to bring AI into work goes to whoever controls the layer underneath the models, meaning the data, permissions, security and management tools businesses already run on Microsoft software.

People building agents for clients land in the same place. On this show, Dave Smith, CEO of Tekscape, which builds and runs agents for its clients, said every agent needs its own identity, governed “in probably the most restrictive manner,” and an audit trail that runs from the trigger that set it off to the outcome.

That’s why the contest moved. Once the model can’t be its own witness, the witness is the layer around it, and that same layer decides what any AI is allowed to do.

In plain terms, the record is the control point, and Microsoft wants to keep the record for every AI a business runs.

Which raises the question of who’s paying for that record.

The Referee Has a Team on the Field
What this means for the MSP: the record of what a client’s AI did is about to come with a sponsor, and the sponsor sells one of the AIs being recorded.

Start with how Microsoft is paying for it. ARN reports that Microsoft unveiled a new Frontier Partner specialization at Ignite, for partners that can design, deploy, secure and govern AI agents across Microsoft’s technology stack. Alongside it come an Agentic Partner Capability Score and partner benefit packages loaded with Copilot Business licenses and Copilot Studio capacity. Microsoft’s channel chief, Nicole Dezen, cited the company’s own research claiming partners with high AI maturity earn 67 percent more partner revenue per dollar of Microsoft AI revenue. That’s Microsoft’s study, measuring revenue from Microsoft. And the oversight product, Agent 365, is pitched as watching Copilot, agents, models and third-party AI services in one place.

Then look at what Microsoft said about those third parties from the same stage. GeekWire reports that Judson Althoff, CEO of Microsoft’s commercial business, told the keynote, “anything you can do in Claude or ChatGPT, you can do better with Copilot and for less.” Copilot chief Jacob Andreou showed a chart putting the monthly cost for an everyday user at 30 dollars for Copilot, 81 for ChatGPT and 205 for Claude. Those numbers come from research Microsoft commissioned and paid for from GigaOm, on workloads Microsoft modeled. Microsoft is a partner and investor in both OpenAI and Anthropic. On that stage, it was their competitor.

Put those together. The company offering to keep the record for every AI a client runs is paying partners to build on its own, and arguing on stage that the others cost too much. None of that says the record would be wrong. It says the referee also has a team on the field. And clients won’t stay on one team. On this show, Tekscape’s Dave Smith said his clients are on Google, Claude and ChatGPT, with some heading to Azure, and any governance framework has to work across all of them.

So here’s the choice. Keep each client’s AI record and policy on a layer you control, one that covers every AI the client actually uses, with Microsoft’s tools as one feed into it. Or make Microsoft’s layer the record, take the Frontier incentives, and accept that whatever Microsoft can’t see, or has reason not to show, never reaches the client’s record.

And that choice shows up first in what you charge.

Why Do We Care?
Microsoft just put a price on the argument: a flat seat as the way to avoid the meter, with its own oversight tools sitting right next to it. When AI governance comes free inside one vendor’s licenses, clients assume it covers everything, and it covers that vendor. Quote the record across every AI a client runs as its own service line, priced by how many AI tools it covers, so the cost of independence is visible before a bundle makes it look free.

What to Consider

  • Count the AI before you quote it. Start every quote with an inventory of the AI tools each client actually uses: Copilot, Claude or ChatGPT on personal accounts, agents built into other apps, browser tools. The price scales with what the record covers, and the inventory is usually the first time the client has seen the list.
  • Keep the record separate from the license. Put the cross-AI record and its policies on their own line item, not inside Copilot or any other vendor’s bundle. Microsoft says its own design is meant to survive a change of AI provider. Your agreement should guarantee that the client’s record survives a change of provider too, including a move away from Microsoft.
  • Check the vendor’s spend view against your own. Microsoft now offers Agent 365 Business Analytics and Cost Management to show what’s driving AI spend. Its own keynote demo ran about 10,000 credits, roughly 100 dollars, for one two-hour session. Before you present AI cost to a client, compare any one vendor’s dashboard with your record across every tool, because a vendor can show only what it bills.

If this trend continues: By the end of 2027, AI governance will come free inside whichever platform a client licenses most, and the only MSPs still charging for it will be the ones whose record covers the AI that platform doesn’t sell.

This content is made possible by community-minded viewers like you

Join the Small Biz Thoughts Community to support independent journalism and get the raw Business of Tech interviews, Channel Intelligence briefs, monthly member calls and more. Member support is one way we sustain the costs of investigative work — from court filings to video editing — while keeping the core reporting open to all.

Thank you

Join the Small Biz Thoughts Community

Choose your upgrade:

Get the full benefits of Business of Tech Plus

Insider Access

$12/month

Perfect for MSPs and ITSPs that want full interviews, early access, and ad-free listening

  • Programmatic Ad-free private podcast feedSame show, little interruptions
  • Channel Chatter previews1–2 topics with light insights
  • Early access to interview episodesHear it days before public release
  • Monthly Insider BriefTighter analysis you can share internally
  • Extra audio segmentsCut interviews, behind-the-scenes commentary, quick competitive notes
  • Become an Insider for $12/month

    Leadership Access

    $149/month

    Perfect for MSPs and Vendors that run a team and need the extended tactics, executive summaries, and weekly alignment brief

  • All Insider Access benefits plus . . .
  • Invite your teamIncludes access for 5 team members with option to add more
  • Vendor Strategy BriefsThe entire library, plus new analysis every month
  • Channel ChatterAll topics, full insights, complete vendor discussion + sentiment list
  • Quarterly State of the Channel Briefing
  • Monthly AMA submission priorityAsk Dave direct questions, and skip the line
  • Get the Leadership Edge for $149/month

    Vendor Partner

    $500/month

    Perfect for channel companies or vendors looking to deepen their engagement with the show.

  • All Leadership Access benefits plus . . .
  • Get highlighted as a show sponsor You'll get placement in the show notes, throughout the website, and on our dedicated sponsors page.
  • Enjoy regular shout outs You'll be featured in a rotating format during the show
  • Become a show sponsor for $500/month

    Search all stories