On 22 September 2026, Reuters reported that Meta had tested human contractors handling some calls made through its Muse AI assistant. The report concerned an internal employee test and raised questions about disclosure and access to sensitive information.

Human assistance can be a sensible part of an AI service. The business needs to explain that arrangement accurately: who might receive the task, what information they will see and what the customer is being asked to authorise.

What happened

Reuters’ investigation, based on internal posts, described privacy concerns from employees. It reported that Meta rolled back the contractor-calling feature. A company spokesperson said any public release would require appropriate disclosures.

The findings should not be stretched into a claim that every Muse call was handled by a person or that the reported privacy concerns establish a confirmed data breach. The important documented issue is the design and disclosure of a human handoff during testing.

Why it matters

A customer may be comfortable asking software to organise a task while having different expectations about an external person receiving the same material. A workflow should make that distinction understandable at the point where it affects the customer’s choice.

Imagine a hypothetical appointment assistant that cannot resolve a scheduling conflict. It transfers the request to a service colleague. The useful handoff includes the necessary details and the unresolved question. It does not require unrestricted access to the customer’s entire message history.

The organisation should specify what the colleague can do next. Can they only propose a time, or can they make a commitment? Can they contact another party? Where can the customer see what happened? A transfer between software and a person should not quietly expand the original authority.

Internal teams need the same clarity. If human support is essential to reliable delivery, include its capacity, working hours, quality checks and cost in the service design. A performance result that depends on people should be evaluated as the result of that combined process.

The bigger shift

The boundary between an automated service and a human-operated service can be hard for a customer to see. That makes accurate descriptions of the operating model more valuable as systems take on longer tasks and interact with other organisations.

Disclosure alone will not fix an unsuitable design. The handoff must also be proportionate, controlled and workable. A lengthy notice cannot substitute for limiting the information shared or ensuring that the person receiving it is properly supported.

For leaders, the question is how the full service works, including its fallback. The CEO’s Guide to AI Governance provides a broader starting point for assigning responsibility. The supplier’s presentation should be tested against the people, systems and permissions actually involved in delivery.

My take

I would ask any provider of an action-taking assistant to demonstrate a difficult handoff. Show what triggers human involvement, what the user is told, what information moves and how the recipient’s authority is limited.

Then examine the experience from the user’s side. Could a reasonable person understand who might handle the task? Can they decline an optional transfer and still reach an appropriate alternative? Are the explanations accessible before the decision matters, rather than buried after the action?

Keep the human contribution visible in the business case too. Skilled support can be valuable, but it needs realistic staffing and honest measurement.

The ambition should be a dependable service whose operation the customer can understand. Needing a person is not evidence that the entire proposition has failed. Concealing or poorly explaining that person’s role creates a separate trust problem that leaders should address before expanding the service.