Musings about Muse, Meta's red-hot new AI agent

Meta's new AI agent, Muse, took just 10 days to reach No. 1 on the U.S. free iPhone app chart. Its promise is easy to appreciate: An assistant that can shop and make phone calls on our behalf could relieve us of some everyday chores. 

For those of us following AI adoption in financial services, however, developments surrounding Muse's launch are more instructive than the consumer hype might suggest. For banks, consumer-facing apps like Muse serve as a crucial test for the operational and security norms they will inevitably have to navigate.

AI + HI still applies

Meta was testing a "human concierge" feature internally, with contractors handling some phone calls placed through Muse, according to Reuters, which reported that internal posts showed pure AI calling struggled with low completion rates because merchants rejected the synthetic voice and automated phone trees. Routing difficult calls to trained human contractors pushed success metrics as high as 95% to 98%.  

Meta ultimately shut down its previous version of virtual assistant, M, in 2018 as it often relied on human help, sometimes for as much as 70% of the work. Meta's renewed attempt on the human concierge approach demonstrates that this fundamental dependency has evolved rather than disappeared. 

I first proposed the AI + HI (human intelligence) approach as best practice for AI adoption in finance also around the same time. Back then I suggested that AI in finance will continue to be "assisted driving" rather than "self-driving" for many years to come. Looks like despite all the technical progress since then, this largely remains true. The key is understanding where the frontier between human- and AI-led tasks is shifting and what underlying drivers are moving it.

Is full agent authority worth the risk?

Amazon's response to Muse brings another issue into focus. The retailer has blocked the agent, citing unauthorized access and concerns about identification and customer credentials. There is obviously a war over customer loyalty going on here between the two tech giants but these are also legitimate concerns. They are questions Amazon should be asking.

Platform-level blocks are precisely the type of operational friction and third-party risk that regulators are warning about. In its August letter to the G20, the Financial Stability Board warned that frontier AI could materially change the speed, scale and economics of cyber risk. In September, Fernando Restoy, chair of BIS' Financial Stability Institute, highlighted operational risks from AI-enabled attacks and disruptions at AI providers. 

These warnings extend beyond any individual product. For banks adopting agentic AI, they reinforce the need to assess how the entire workflow will function under stress, including when an external service fails or an agent encounters circumstances its designers did not anticipate. 

Another subscription to save on subscriptions?

One of Meta's pitches for Muse is that it can help consumers save money. Alexandr Wang, Meta's chief AI officer, has promoted a "Muse Money Challenge." Examples include canceling unused subscriptions and negotiating lower bills.

There is, of course, some irony in paying for another subscription to reduce subscription expenses. Yes, Muse does have a free basic version but heavier users can pay $20 or $100 a month. As I argued in my recent writing on AI ROI, the starting point is the baseline before AI is introduced into the equation: Do the reduced expenses cover the added cost? Users' calculation here would be the same.

I am not sold on Muse, but that could just be me: I was and still am not sold on grocery shopping services like WebVan. Personal skepticism aside, Muse provides a timely test for AI adoption. Balancing operational risks, human oversight and ROI consideration will determine how much financial services can realize agents' full promise.