September 21, 2026
TechCrunch, citing Appfigures estimates, reported that Meta's new AI agent Muse has out-downloaded and out-engaged ChatGPT's early mobile numbers in the U.S. and Canada. That's a genuinely impressive distribution win -- Meta has billions of existing app users to funnel into a new product, something OpenAI didn't have when ChatGPT first launched on mobile. But raw install numbers are a vanity metric until you ask what the agent can actually do once it's on your phone. And on that question, the same day brought a much less flattering data point: Amazon has blocked Muse from operating on Amazon.com, according to TechCrunch. Amazon runs its own foundation models and one of the largest inference businesses on the internet -- there's no legal requirement forcing it to let a rival's agent shop, compare prices, or check order status on its turf, so it simply didn't.
This is the part business buyers should sit with. An AI agent's usefulness isn't just a function of its model quality -- it's a function of what it's allowed to touch. If the biggest platforms start selectively locking out competing agents, the promise of a single assistant that handles your shopping, your scheduling, and your errands quietly evaporates into a patchwork of walled gardens. For companies evaluating agentic tools for internal use -- not just consumer novelty -- that's a real warning sign: an agent's value depends heavily on integration access you don't control, which is exactly why we've argued that owning your own workflow automation and integrations layer matters more than betting on someone else's agent to play nice with every vendor you rely on.
OpenAI has formed a math advisory group after its models reportedly helped resolve more than 100 open problems, TechCrunch reported -- but notably, the group has no mandate to slow down or redirect that research. Read that combination carefully: OpenAI wants credentialed mathematicians attached to the story of its models cracking real open problems, but it isn't handing them any authority to say 'wait, we should check this more carefully before publishing' or 'this claimed proof needs independent verification first.' That's advisory theater dressed up as scientific rigor.
For a business audience, the lesson isn't about number theory -- it's about how AI vendors talk about their own capability claims generally. A model 'resolving' an open problem is a headline; whether that resolution holds up to peer review is a separate, slower, less exciting process, and vendors have every incentive to let the headline travel faster than the correction. If you're evaluating any AI tool's claims about what it can verify or guarantee -- financial reconciliation, compliance checks, code correctness -- the same skepticism applies: ask who actually has the authority to say no, not just who's been given a nameplate.
TechCrunch profiled Tabby, a startup founded by a former accountant, built as a real-time bookkeeping interface that gives business owners live profit-and-loss visibility instead of waiting on a monthly close. This is a far more concrete and near-term story than Muse's download numbers or OpenAI's math claims, because bookkeeping is exactly the kind of structured, rules-based, document-heavy work that current AI genuinely handles well. It's less flashy than an agent that shops for you, but it's the category where displacement is already happening, not theoretical.
The honest counterpoint is that 'obsolete' is doing a lot of work in that headline -- most small businesses still want a human who can explain a weird quarter to a lender or catch something the software missed, and that judgment layer isn't going away soon. But the real-time data layer underneath it is exactly the kind of thing worth automating rather than paying someone to reconcile by hand, which is the same logic behind pairing live finance & billing tooling with broader analytics & reporting instead of waiting for a monthly spreadsheet. Tools like Tabby are a preview of a broader pattern: AI doesn't need to be flashy to be disruptive, it just needs to remove a slow, expensive middle step.
So which worries you more as a business owner: an AI agent that can't shop where you shop, a research claim you can't independently verify, or a tool quietly aiming at your bookkeeper's job?
Sources