Outcomes Over Agents Is the Whole Game
Meta's Muse plugs into your email, calendar, payments, and health data. It just does stuff. Since launching on September 8, Muse has connected to Gmail, Google Calendar, Outlook, Peloton, Plaid, Spotify, and payment rails via Stripe's Link, and it briefly overtook ChatGPT as the top download on Apple's App Store]. Nobody using it is asking which model sits underneath. That's not an accident. It's the whole strategy.
The Model Isn't the Point
Meta CTO Andrew Bosworth said this explicitly on the Big Technology podcast in July: "The model itself isn't the value... we're going to get to a world very soon where consumers don't care which model they're using - they don't care if it's 4.7 or 4.8, the same way you don't care whether you're using Oracle or a SQL database. You just want the functionality to work well".
His argument is that a state-of-the-art model is table stakes you can rent from Anthropic, OpenAI, or Google. The durable value sits in the layer around it - the product, the distribution, and the consumer relationship. Most competitors have only one of those pieces. Meta is betting it can own all four .
That's exactly what we've been building toward in public services. The intelligence has to disappear. If a caseworker or a resident notices the "AI bit," we've already lost. They don't want a chatbot. They want the case resolved, the form to stop bouncing back, the payment to land on time.
The Line Everyone Skips Past
None of that works without trust. Meta can only get away with an agent touching your payments and health data because enough people trust the outcome more than they fear the access - and that trust is already being tested. Within two weeks of launch, Amazon blocked Muse from shopping on its site, accusing Meta of letting the agent move through customer accounts without disclosing itself or Amazon's consent. TechCrunch framed the whole launch as "a major test of whether people still trust Meta with their data".
Meta's answer has been to build in visible guardrails: a separate "Sentinel" monitoring agent has to approve anything Muse sends to the internet, the agent checks with users before sensitive actions like sending an email or making a purchase, and Stripe's Link issues one-time-use card numbers so real payment details stay hidden.
In Public Service It is not cost. Not tech. It is Trust.
Public sector is no different, and the evidence is sharper here than almost anywhere else. Capgemini's research into government AI adoption found that 74% of public sector leaders cite limited trust in AI-generated outputs as a primary barrier to scaling AI, second only to data security concerns at 79%. Not cost. Not tech. Trust.
Three Non-Negotiables
So for us, three things don't get compromised on:
Codesign. Build with frontline staff and users, not for them. That's where legitimacy comes from, not the model card.
Trust. Earned through visible safeguards, human override, appeals, and transparency. Not a compliance tick.
Impact. Measured in outcomes, not activity. Not forms processed. Did the person's actual problem go away? Are they satisfied with the service?
Higher Stakes, Same Bar
Meta is proving this thesis at consumer scale with real money behind it, Muse now sits across email, calendar, health, shopping, and smart-home data for millions of users, with Meta layering on a keychain hardware device, a dedicated agent email address, and Mac computer-use within weeks of launch.
Our job is proving the same thesis in public services with OSCAR AI, where the stakes and the trust bar are even higher. A resident whose benefit payment doesn't land, or whose case gets stuck in an AI-mediated loop, doesn't get to just uninstall the app.
The agent isn't the product. The outcome is.
