The Market Is Overestimating Meta’s Muse. AI Isn’t Ready to Take Over Consumer Tasks

There’s been a lot of discussion about MUSE lately. My view is simple: **the market is overestimating it. MUSE won’t fundamentally change the AI ecosystem, because AI agents still aren’t ready for mainstream consumer tasks.**

Today is September 24, 2026. Come back in a few months and see if I’m wrong.

The post is quite long, here's key takeway to save ur time

• **MUSE lowers the barrier to consumer AI agents.** It makes existing PC-based agents easier to use, but doesn’t introduce a fundamentally new capability.

• **MUSE can only amplify existing demand.** If PC users haven’t already embraced agents for consumer tasks, lowering the barrier won’t create that demand.

• **AI is still mainly an info advisor, not an executor.** Even the most AI-enthusiastic users often research with AI but still handle the actual transaction themselves, as seen in the shopping-agent experiments by OpenAI and ByteDance.

• **Consumer agents still struggle with long-task execution.** They need to complete complex tasks with minimal human intervention, but four structural problems remain:

1. **Web infrastructure:** Websites are built for humans, not machines.
2. **Reliability:** AI still struggles with memory and hallucinations over long tasks.
3. **Cost:** Long-running agents consume too many tokens which priced too high to cover
4. **Task complexity:** Consumer tasks are highly non-standardized, making correctness difficult to evaluate and training harder than in B2B applications.

Overall, consumer agents still require major improvements in both technology and infrastructure. **It’s more likely to follow the path Amazon took in e-commerce: start with a few verticals, solve the problems there, and gradually expand into everything else.**

That’s why I think **Zuckerberg’s strategy of building a general-purpose consumer agent from the start is misguided.**

**--------------------------------------here's the full post-----------------------------------------------**

A simple question to testfiy my view to ask **whether PC tech geeks are actually using Codex for everyday consumer tasks.** They’re the most AI-friendly users and among the most willing to pay. If even they aren’t adopting agents, I don’t see why mainstream consumers would.

I’m not saying geeks haven’t embraced AI in their daily lives. **They’re the most enthusiastic users.** **The problem is that they still mainly use AI as an information advisor.** When it comes to actually making purchases, filling out forms, or completing transactions, they still do the execution themselves.

I’m fairly confident about this for two reasons. First, Sam Altman admitted in an August interview that even with a tool as powerful as Codex, he still can’t resist traditional computer workflows.

Second, OpenAI and ByteDance are arguably the strongest consumer AI players in their respective markets, and both are aggressively pushing shopping agents. Yet early results suggest that after using AI to research products, users often don’t complete the purchase through the agent. They go to consumer communities for more information or simply buy the product themselves.

**That tells me even AI’s most enthusiastic users still aren’t ready to hand execution over to an agent.**

**MUSE mainly makes AI easier to use. But that’s just a demand multiplier. It doesn’t matter if the underlying use case isn’t compelling.**

The real value of an agent is saving time on **long, complicated tasks**—planning a week-long trip, filing taxes, buying insurance, or doing the weekly grocery shopping. Cutting a 3-minute task to 30 seconds isn’t transformative; cutting a 3-hour task to 5 minutes is.

MUSE still struggles with the latter because of three problems:

**1. The web wasn’t built for AI agents.**

Websites are designed for humans, not machines. CAPTCHAs, bot detection and anti-automation systems actively get in the way.

More importantly, there’s no standardized way for websites to handle exceptions. If a hotel is sold out, one site may recommend another hotel, another may offer a cheaper room, and another may simply stop the process. An agent has no universal protocol for deciding what to do next, so it often has to ask a human.

For agents to really scale, merchants will need **AI-specific APIs and standardized exception handling**. That’s a massive infrastructure project.

Even in travel, Meta has connected major US airlines, but hotels, restaurants and ground transportation are still far from fully integrated.

**2. AI is still unreliable on long tasks.** Agents forget things, hallucinate, and lose track of steps. If users still have to check everything, much of the time-saving benefit disappears.

**3. Token costs may kill the economics.** Long-running agents consume far more tokens than simple queries, while consumers are less willing to pay than businesses.

**4.B2B tasks are generally more standardized, while consumer needs are far more subjective.** Coding is a good example: if the code runs, the task is usually done. Many office workflows can also be tested in a controlled environment with internal employees helping evaluate the results.

Consumer tasks are different. Whether a suit actually fits or looks good is something only the customer can judge. That makes consumer-agent training much slower. **Early breakthroughs will have to come from highly standardized verticals.**

**Overall, consumer agents are more like early e-commerce: they require years of technical breakthroughs and infrastructure buildout, rather than an overnight, all-category breakthrough.**

Amazon started with books before expanding into everything else. **Consumer agents will likely need the same focused strategy. That’s why I think Zuckerberg’s attempt to sell a “do everything” consumer-agent story to the market is the wrong approach.**

**For now, I still think the biggest opportunity for consumer AI is advertising.**

Come back in a few months and let’s see if I’m wrong.