95% 16 to 98% success. That was the range some Meta tests suggested human callers could reach, according to Reuters on September 22, 2026 10. AI-only calls scored lower. The report did not give their exact rate.
Those people were handling requests placed through Muse, Meta’s personal AI assistant. According to Reuters, Meta was testing the use of trained contractors to handle some phone calls placed via Muse. Employees raised privacy concerns. Meta acknowledged inadequate disclosures in Muse’s human concierge test and rolled it back.
The human help itself deserves a fair hearing. A user who wants a haircut booked may welcome a person who can finish the job. But permission to use an assistant should not quietly become permission to share personal details with a contractor.
I think Muse exposes the next hard problem in agent engineering: transferring responsibility without losing the user’s intent. A better model may reduce the need for help. Builders still need a clear contract for when help arrives, what it can do, and how completion gets checked.
The Handoff Contract
Call that contract the Handoff Contract. It defines how a task changes hands while keeping the user’s limits intact.
What the Muse concierge test actually put on the table.
For a Muse-style calling service, I would make four fields mandatory:
- Trigger: The specific condition that stops the automated path. A request for identity verification could require the user to step in.
- Brief: The task objective and the minimum context needed to act. A haircut booking should not expose unrelated conversation history.
- Permission: The actions the next operator may take. Asking about availability and accepting a cancellation fee require different authority.
- Receipt: The evidence returned after execution. A booking record should include the confirmed time and any unresolved conditions.
Define these fields before choosing an operator. Both a contractor and a user taking back control need them.
Reuters reported that Meta was testing human contractors to handle some Muse phone-call requests. That does not establish that Meta transferred live calls midway through conversations. Builders should keep that distinction clear because each workflow needs different coordination.
A request transfer needs a complete brief. A live transfer also needs a reliable account of what has already happened. Otherwise, the next caller risks repeating an action or contradicting a commitment.
The Handoff Contract also allows a clean stop. If nobody has authority to proceed, the system should return the decision to the user.
The Call Needs An Owner
A salon booking makes the engineering problem concrete. Imagine a user requests an appointment within a stated window and refuses any cancellation charge. This is a proposed test case, not a reported Muse interaction.
The assistant reaches a point where it cannot proceed. Perhaps the business offers a different time or asks for details the user never approved sharing.
The tempting implementation is simple: send the conversation to a worker and ask them to finish. That leaves the hardest questions unanswered.
Who owns the task now? Can the automated caller keep retrying? What happens if the user cancels while a contractor is waiting on hold?
I would use explicit task states: recorded labels that show where a request stands.
A request could move from automated work to awaiting assistance. Once a worker accepts it, the system should block another operator from acting on the same request. Cancellation should invalidate outstanding permission to make a booking.
That is the boring plumbing worth building first. A slick voice cannot repair two callers making incompatible commitments.
Keep the worker’s brief compact and structured, with named fields rather than a loose paragraph.
For the salon example, include the approved appointment window and the rule against cancellation charges. Show only the contact details needed for the booking. Keep unrelated chats out of the worker’s view.
Disclosure belongs before sensitive context changes hands. Ask the user whether human help is acceptable. Tell the business who is calling and whose instructions govern the request.
The reported failure explains why this matters. Reuters said a Meta executive called testing contractor-placed calls without proper disclosures “a miss.” The operating design needed better disclosure.
Human help also needs a conduct policy. Reuters reported that an employee said a transcript showed a contractor making a racist reference. Having a person involved is not a complete safety argument.
Give the return path equal attention. Require the worker to record the outcome against the original constraints. Let the assistant produce its user-facing summary from that record.
If the appointment falls outside the approved window, the task remains unresolved. A pleasant conversation should never erase an unmet requirement.
The reported success range needs this same discipline. It is unclear whether “success” meant completing the task correctly while honoring every user constraint.
My proposed scorecard would separate autonomous completion from human-assisted completion. Failed transfers would get their own count. Unauthorized actions would remain failures even when a business confirmed the appointment.
Then put a dollar figure on the workflow:
Cost per verified completion = (model costs + phone costs + human labor + oversight) / verified completed tasks.
This is a proposed accounting rule, not a Muse cost estimate.
A contractor queue belongs in that calculation too. Set an operating budget and an expiry rule for waiting requests. If nobody accepts a task, tell the user it remains unfinished.
The useful question for a builder is what a correctly completed booking costs. Cheap model output alone cannot answer it.
Where a booking changes hands and breaks
Permission to use an assistant is not permission to brief a contractor.
Reuters reported that a Meta executive called testing contractor-placed calls without proper disclosures a miss, and the internal human concierge test was pulled back. Employees had already raised privacy concerns. Disclosure has to land before sensitive context moves.
Nobody has said what success counted.
Some Meta tests suggested human callers reached 95% to 98%, with the AI-only rate lower and unstated. It is unclear whether success meant completing the task while honoring every user constraint. A scorecard should split autonomous completion, assisted completion, failed transfers and unauthorized actions.
Human escalation creates a second budget line.
The reported GPT-6 prompt-caching update for Sol and Luna would cut repeated-prompt cost, but labor and oversight sit outside that saving. Cost per verified completion divides model, phone, labor and oversight spend by verified completed tasks. The four Handoff Contract fields are what make a completion verifiable at all.
2031: The Service Behind Glasses
Over the next five years, I expect dependable delegation to give some companies an advantage over others. Companies that preserve permission through every transfer could earn the right to handle more consequential tasks.
Meta’s reported rollback offers a case study. Reuters reported that Meta temporarily pulled back Muse's internal human-concierge calling test after disclosure concerns. Capability gains did not settle the launch decision.
If Muse reaches that form factor, I expect unresolved tasks to become harder to overlook. A spoken assurance follows the user into the next activity. A missing appointment can surface long after the conversation ends.
That makes clear status updates more valuable. The brief also describes an OpenAI prompt-caching update for the GPT-6 family, specifically GPT-6 Sol and Luna, aimed at reducing latency and cost for repeated prompts. Prompt caching generally means reusing prior computation for repeated prompt prefixes.
Treat it as a reported infrastructure direction rather than measured evidence about Muse.
The economic connection is still useful. Reusing computation would address one part of a task’s cost. Human escalation would introduce a separate labor budget that needs its own controls.
The strategic goal should be to learn from that budget. Record why help was required, then test whether a narrower workflow can safely remove the dependency.
The gains could compound. Each reviewed failure can improve the next version of the service.
None of this proves every agent needs a contractor pool. Some products should offer fewer actions or return hard decisions to the user. Scope is an engineering choice too.
Prototype A Booking Handoff Desk
This week, build a small handoff desk for appointment requests. Use fictional details and willing testers before contacting real businesses. You can prototype the control flow with a form and a spreadsheet without a computer science degree.
First, create the Handoff Contract as a task form. Include the requested service and acceptable appointment window. Add separate fields for forbidden commitments and permission to involve a person.
The form should reject requests missing an authority boundary. “Handle it for me” leaves too much open.
Next, make a queue that one tester can operate. Give each request a unique identifier and an owner field. Only the owner should be allowed to record a booking.
Keep the assistant’s role small during this test. It can draft the brief, but a tester should verify that the draft preserves every user constraint.
Then write the disclosure messages before connecting any real calling service. The user message should explain what information a person would receive. The business message should identify the caller’s role.
A short message can be clear without pretending to settle every legal requirement. Get appropriate review before using the workflow with real customer information.
Finally, build the receipt form and test failure paths. Offer an appointment outside the permitted window. Cancel a request after a worker accepts it.
Also test duplicate acceptance and a queue with nobody available. The prototype should preserve an honest status in every case.
Track the time spent repairing each failed transfer. Record why a tester had to reconstruct the brief. Those observations will tell you where the next engineering hour belongs.
For tools, inspect the adam.new demo before committing. Today’s digest does not establish its capabilities or pricing. Keep the prototype tool-independent until you know whether it fits.
The digest also lists AI for News v1 without a substantive feature description or verified product link. Check the actual product before assigning it research work. Neither listing establishes suitability for sensitive customer data.
Expect the first version to break. Fix the transfer that lost a constraint before polishing the assistant’s voice.
Muse’s reported test offers a useful starting point: human help can be a legitimate product feature. Build it so everyone knows when it happens. The user should never have to guess who acted or what they agreed to.
Prototype a booking handoff desk with a form and a spreadsheet.
- Write the Handoff Contract as a task form. Give it four mandatory fields: trigger, brief, permission, receipt. Include the requested service, the acceptable appointment window, forbidden commitments such as cancellation charges, and explicit permission to involve a person. Reject any request that leaves the authority boundary blank.
- Run a single-owner queue. Give every request a unique identifier and an owner field, and let only the owner record a booking. Then test the ugly paths: duplicate acceptance, a cancellation after a worker accepts, and a request nobody picks up before its expiry.
- Draft the two disclosure messages first. Tell the user exactly what information a person would receive, and tell the business who is calling and whose instructions govern the request. Use fictional details and willing testers, and get appropriate review before any real customer data enters the workflow.
Ship the plumbing before the voice.
Meta's reported Muse test showed that a human fallback can raise completion rates and still fail the user, because the failure was disclosure and coordination rather than capability. A task that changes hands needs a trigger, a compact brief, a stated permission and a receipt checked against the original constraints. Explicit task states stop two callers from making incompatible commitments, and a cancellation should void outstanding authority to book. Measure cost per verified completion so that model savings, including reported prompt-caching gains for the GPT-6 family, do not hide a growing labor budget. Better models will shrink how often help is needed, but they will never decide who owns the call.