Meta's chief AI officer Alexandr Wang says the company's new Muse agent can find $1,000 in savings through the, yet a TechRadar test recovered only $35 a month from forgotten subscriptions and a cheaper phone plan. The gap between the promised amount and the verified result shows why action-taking agents need separate evaluation from conventional chatbots. For companies, the case is a practical check of what autonomy, access and supervision actually deliver.
How the $35 test result was reached
Wang promoted the challenge on September 19, 2026, inviting users to put Muse to work on personal budgets. Shared results so far include cheaper insurance policies, cancelled forgotten subscriptions and rediscovered unused gift cards, although many early publicised cases come from Meta employees. The TechRadar journalist started with a broad request to find ways to save money and granted access to relevant services. Muse then flagged Patreon and Substack subscriptions as candidates for cancellation and pointed to newly released data plans that could lower a phone bill. After manual review, the usable saving totalled about $35, or 3.5% of the $1,000 target.
Muse is positioned not as an adviser but as software that completes tasks. According to Meta, it runs on its own secure virtual computer with a browser, connects to services such as email and calendar, fills out forms and interacts with customer service systems. It keeps working after the user closes the app, while purchases and outgoing messages can require explicit approval. Access is optional and controlled through permissions for other apps and services, with login credentials stored separately so the agent cannot read them directly. That architecture explains the difference from asking ChatGPT for saving tips about meal planning and streaming cancellations: Muse is expected to locate the charge itself and prepare the cancellation or switch.
The test underlines how much results depend on accumulated financial clutter. Recovering a year of payments is possible only when a subscription has actually run unnoticed for a year, which explains why outcomes vary from large one-off recoveries to small monthly adjustments. The author noted that $35 still represented money that would otherwise have continued to leave the account each month. The more durable use case may be administrative work that people postpone because a 40-minute exchange with customer service outweighs the expected refund or credit. In that framing, the value lies less in hitting $1,000 than in delegating tedious checks for refunds, credits and recurring charges.
What this means for business use of agents
For businesses, the pattern translates directly to recurring-cost audits. Small teams often carry overlapping subscriptions, newsletter tools and mobile plans selected long ago, and an agent with read access to email invoices can surface them faster than a manual review. Larger organisations face the same issue at scale across departments, where forgotten seats, duplicate tools and outdated telecom tariffs accumulate quietly. In both cases the noticeable effect is not a strategy report but prepared actions: identified charges, alternative plans and draft cancellations. The monthly nature of the $35 finding matters here, since small recurring amounts compound across teams and contract cycles.
Limits centre on access, verification and expectations. Muse becomes useful only after it receives entry to inboxes, calendars and accounts where billing evidence lives, which raises permission reviews and internal rules about what an external agent may open. Credential separation and approval steps for purchases and messages reduce risk but do not remove the need for human review, as the tester checked each suggestion before acting. The news by itself does not prove that $1,000 is available in every account, especially since prominent examples come from inside Meta. Questions to ask a vendor therefore include what data sources are required, which actions need approval, how credentials are isolated, and how proposed savings are documented before execution.
The marker to watch is whether Meta or independent users publish verified savings beyond employee accounts, with a breakdown between one-off refunds and recurring cancellations. If outside tests consistently show documented reductions in insurance, telecom and subscription spend with approval logs intact, agentic savings can be treated as an operational tool. If results stay anecdotal, the $1,000 figure remains a marketing benchmark rather than a planning input.
