SHEYU.AI舍予基业
AI Daily Brief

Microsoft and 40+ partners set AI transparency rules: who answers when agents act?

Microsoft's AI transparency initiative with 40+ organisations targets evidence and accountability for autonomous agents. For SMEs, the practical step is deciding what must stay human-signed.

·3 min read
Microsoft and 40+ partners set AI transparency rules: who answers when agents act?
AI-generated illustration, not a news photograph

Microsoft used Ignite to launch an "AI transparency initiative", bringing together more than 40 organisations — Accenture, Adobe, KPMG, Mastercard and Volvo Cars among them — to develop cross-model, cross-vendor standards for AI transparency and governance. According to Gelonghui and Investing.com, the announcement came on 17 November local time.

That may sound remote from small and mid-sized businesses. But if you have ever considered letting AI watch information, fill in forms or run processes on your behalf, it concerns you.

What Microsoft is actually trying to fix

Once companies connect AI to data, tools and credentials, agents can carry out operations without human approval at every step. The problem: existing security mechanisms can see what happens around the model, but not how the model reasons internally, and they cannot independently verify what the model says about its own behaviour. The gist of comments from Hayete Gallot, corporate vice-president of Microsoft's security business, is that vendors' own accounts are not enough.

Put plainly: when AI does work for you, you need to be able to produce evidence of what it did.

If AI does the work, who carries the liability?

You do.

Microsoft is direct on this point: an agent is likened to a "privileged insider" operating at machine speed and machine scale, and the security, compliance, audit and accountability risks still fall on the organisation that deploys it. When you have to explain matters to customers, employees, boards, auditors and regulators, you cannot pass the blame to the model vendor.

For teams going global, the constraint is tighter. You face not only domestic customers but also compliance enquiries from overseas platforms, payers and partners. Where AI participated in a judgment, you need to be able to reconstruct the basis for it.

Do smaller firms need this "evidence chain" too?

It depends on what AI does in your business.

If AI only polishes copy, the risk is limited. But if AI is involved in pricing judgments, supplier screening or bidding decisions — matters where an error costs money — an evidence chain is no longer a requirement only for large companies.

Microsoft's framework breaks the requirements into three parts. The most useful way for an operator to read them:

  • Observe: requests and authorisations, the model and runtime environment, which data and tools were accessed, and what was ultimately done — all must leave a record;
  • Verify: records must make clear whether information came from the model, the runtime environment, or an independent identity and policy system — not rely solely on the vendor's own account;
  • Control: major actions are gated by policy beforehand, and afterwards can be reconstructed and audited; those records must remain usable when you change models or vendors.

Whether the framework can be implemented depends on whether participants can turn it into common standards and portable evidence. That is not yet settled; the standards are still being drafted.

What can be done now?

Start with the task you most depend on AI for, and change it so that a human makes the final judgment while AI only assembles the information.

This is not defeatism. The heavier the decision and the responsibility, the less AI should be left to decide on its own. Shu Yi, executive director of Sheyu, puts it in a way we find practical: AI will not make decisions for the boss; it puts the numbers the boss needs to see in front of him first.

Bidding is a typical case. When a notice is issued, where it appears, how many days remain until the deadline — miss that information and the best-written proposal is useless. What SHEYU AIBID (舍予AI投标官) does is aggregate national government procurement and tendering information, with subscription pushes by keyword, region and industry; the original notices can be browsed, and AI can then be used to draft the proposal. What it cannot do is read the workings behind a project for you, or decide for you whether to bid. That judgment stays with people.

Shu Yi has another line: the outcome of a bid is often half-decided the moment the notice is seen. In today's context, we read it this way — seeing is only the first step; what evidence you use to decide after seeing is where the real gap opens.

The rules being drawn up by Microsoft and its 40-plus partners will not land in your contracts any time soon. But the direction is clear: the more autonomous the AI, the more evidence you must keep. Smaller firms do not need to build a governance system now, but they can draw the line on which matters require a human signature. The earlier that line is drawn, the less trouble later.

AI can make you faster. It cannot carry the responsibility for you.

閱讀繁體中文版 →