Human Work Behind AI Agents Must Be Disclosed and Governed

A human fallback changed the nature of the service

Reuters in September 2026 reported that Meta tested human contractors to handle some telephone calls initiated through Muse, its personal AI agent. Employees raised privacy concerns, and Meta rolled the feature back after acknowledging that the test had started without adequate disclosure. Internal tests cited by Reuters indicated that human handling could raise call success to between 95% and 98%, compared with a lower rate for fully automated calls.

The episode is not simply a product-design detail. When a person receives the task, the data path, confidentiality position, quality controls and accountability all change. A user who believes an automated system is acting may make different disclosure choices from a user who knows a contractor will read the request or conduct the conversation.

Human in the loop is not one control

Organisations often describe human review as a safeguard, but the term can cover very different arrangements. A trained employee may approve a recommendation. A service centre may take over an interaction. A contractor may label data, correct an output or complete the task. Each model creates different access rights, competence requirements and employment or outsourcing risks.

The operating design should specify when the handoff occurs, what information the person can see, where the work is performed, whether subcontractors are used and how the intervention is recorded. Users should know when a human has entered the process, especially where health, benefits, financial or employment information may be involved.

Performance metrics need to show the handoff

A headline success rate can be misleading if it combines automated completion with human rescue. Management should report at least three measures: tasks completed without intervention, tasks completed after human assistance and tasks that failed or were abandoned. Cost, turnaround time and error rates should also be separated by path.

This distinction matters for business cases. A service may still be valuable with substantial human support, but its economics and scalability will differ from those of an autonomous agent. The organisation should decide whether human fallback is a temporary training mechanism, a permanent premium service or an exception route for defined cases.

Benefits assistants should make human review visible by design

A specialist benefits assistant will sometimes need professional review because sources conflict, local rules change or a recommendation has material consequences. That review should be an explicit feature rather than a hidden correction layer. The interface can show that a case has been escalated, identify the reviewer role and preserve the sources and changes.

Contracts should cover confidentiality, location, subcontracting, training, audit rights and incident reporting for every human support provider. The product description and metrics should state the true division of work. Trust depends less on claiming full autonomy than on explaining reliably who or what performed the task.

Contract for the real operating model

Procurement and legal teams should verify the service behind the interface. Contracts need to identify the provider responsible for the model, the operator of any human support, approved locations, access controls and retention periods. They should state whether customer content may be used for training and whether a contractor can make external calls, send messages or complete transactions.

Audit rights and incident duties should follow the full chain. A provider should be able to reconstruct the task, the model output, the reason for escalation, the information shown to the human and the action taken. Where that evidence cannot be produced, the client may be unable to investigate a complaint or demonstrate compliance. Service design, commercial claims and contractual terms therefore need to describe the same operating reality.

Clients should also decide whether users may opt out of human assistance. In some services, a visible choice between automated processing, professional review and no continuation may be appropriate. Where regulation or safety requires review, the condition should be stated before the user submits sensitive information. Consent is meaningful only when the alternatives and consequences are understandable.

Source: 2026-09-24_GBV_Human_Fallback_AI_Agents.docx