Three questions you should ask about Muse.

The questions set you apart from everyone else.

Meta’s AI assistant Muse has taken off, and I think it raised three important questions.

The first one that caught my attention was how two companies reacted completely differently to the same agent. Amazon blocked Muse from shopping on its site. Shopify, meanwhile, rushed to make sure the merchant's product is recommended by Muse, with direct checkout in certain cases.

To me, this split in strategy points to the first question: Who Owns the Customer?

---

Question 1: Business to Agent: Who owns the customers?

You likely know this: B2B, that’s businesses selling to businesses. And B2C is businesses selling to consumers.

But what would it be if your agent is the one finding products, comparing prices, and narrowing down your options, so a business will have to win over that agent first, instead of you?

You’re still the one making the final decision and paying for the product. But you’re no longer opening 10 tabs and reading 100 reviews. Your agent gets the first look at what you want and decides which products you see.

That puts it between the shopper and the store.

Now look at how differently two companies responded to Muse. Amazon blocked it from shopping on its site. Shopify, meanwhile, lets merchants’ products appear in Muse, with direct checkout available in some cases.

And you can see the tension if you read between the lines.

Amazon’s stated position is that an outside app shopping on a customer’s behalf should be open about what it’s doing and respect a platform’s choice to participate. That matters to the platform owner, like Amazon.

But there’s also an obvious business question here: if people start by telling Muse, “I want to get this,” Who learns what the customer wanted in the first place, and who gets to recommend the product?

Currently, Amazon controls this. For example, Amazon always sees which third-party products are viral and profitable. After Peak Design’s Everyday Sling became a popular camera bag on Amazon, Amazon launched a cheaper Amazon Basics sling.

However, in this business-to-agent model, Amazon is no longer in control of this momentum. Personal agents won’t just stop Amazon from getting more revenue; it would even slowly dry out the current revenue flow.

Shopify, on the other hand, is in a more favorable position when working with personal agents.

Essentially, Shopify is never a destination you go shopping at, but rather the platform merchants use to build their own shops. And because of this, Shopify’s job is to help its merchants max out their discoverability. So, as an early personal agent adopter, Shopify gives its merchants the best chance to be in front of these agents.

But the question is bigger than just these two marketplaces.

For this, we need to go beyond the concept of shopping on one website.

Meta already has a place in the moments when we discover things. Say I save a cooking video on Instagram. The recommendation has done its job: it caught my attention. But if I decide to make that dinner, I might leave Instagram, search for the ingredients, compare prices, and buy them somewhere else.

Muse is designed to take over a big part of this job. Just imagine I tell it, “I want to make this tomorrow. What do I need, and where should I buy it?” Muse hears that request before any other store does. It can decide which products and sellers make it onto my shortlist.

That’s what I mean by the intent layer.

It is still me who decides whether to buy. But the first conversation about what I want has happened with my agent.

So when I ask who owns the customer, I really mean: who gets that first request, and who helps me decide where to go next? That, to me, is the Business to Agent shift.

And if Muse becomes that starting point for enough people, you can see why the next question is whether Muse would earn back Meta’s enormous AI spending.

---

Question 2: Can Muse make the AI math work?

The second signal I see is that Muse gives Meta’s enormous AI spending a more visible path to a return.

As I mentioned last week, in its first 12 days, Muse reached about 2.8 million downloads worldwide and roughly 640,000 daily mobile users in the US. Its US daily user count is already ahead of ChatGPT’s at the same point after its mobile launch.

Muse added roughly $200 billion to Meta’s market value. This market reaction makes complete sense when you look at the bill investors have been waiting for Meta to justify.

In its July report, Meta expects to spend between $130 billion and $145 billion on capital expenditures in 2026.

So Meta faces the same question as other tech giants: how does all this spending pay off?

And Muse saved the show (at least a little) by giving people something concrete to picture.

There’s a free tier, plus paid plans at $20 and $100 a month. Not to mention, when enough people use it to discover products and complete purchases, then Meta will naturally expand its ad business around agent-enabled discovery and transactions.

Yet, one popular launch doesn’t mean Meta has earned back what it spent.

Muse has now reached 6.6 million downloads, but downloads don’t equal money.

Let’s do a back-of-the-envelope:

Citi thinks Muse could bring in about $27.5 billion a year by 2030, while Meta expects to spend nearly $145 billion on capital expenditures in 2026. Even if Citi is right, Muse’s projected annual sales in 2030 would amount to only about one-fifth of Meta’s projected capital spending in 2026.

And sales aren’t profit. Even if I were generous here, giving Muse a hypothetical 50% profit margin, and you’re down to roughly $14 billion a year. Not to mention the immediate competition from the likes of Dots by OpenAI and  Grok Bot.

Now, I’m not saying Muse should pay for every data center Meta builds.

Those serve a lot more than one app, of course; it’d also pay for the ad and AI features on other Meta platforms.

But that’s exactly why I’m in two minds. Even though Muse has made it easier to imagine how some of the investment might pay off. There is still a huge gap to fill, not to mention too many unknowns, including user stickiness and the costs of running training and inference.

Anyhow, Muse, as a convincing story for investors, is a good start.

Let’s bring the focus back to you, the user.

---

Question 3: Would you give it a second chance?

Why are so many agents showing up now? My first thought was, didn’t OpenClaw already have memory and can control your computer? What exactly is new?

To me, the difference is that models can now get through more of a task without a complex setup and don’t require babysitting every step.

Fine. The hype will push me to try it once. But it won’t necessarily make me keep using it just because it’s popular. If I have to repeat all my context, chase down what it’s doing, or second-guess every permission, I’m probably doing the next job myself.

That’s my third question: which agent earns the second task? I don’t think the smartest model automatically wins that one.

Put Muse next to Dots, Grok Bot, and Manus Cue, and you can see how differently these companies are trying to answer that question. Each wants to become useful at a moment you’ll encounter again and again.

For example, I’ve been using Dots, and the thing that won me over was pretty mundane.

I keep different parts of my work in different projects—recordings in one, research in another. Half the time, I can’t remember where I put the context I need. With Dots, I can just ask it to find that context and carry on with the task. It’s not perfect, but that saves me a small, recurring headache. And because it worked today, I’m more likely to hand it something else tomorrow.

The other products are making different bets on what that recurring job might be.

Grok Bot gives you a team of agents you can delegate work to. Muse wants to handle shopping, travel, and the errands that pile up outside work. Manus Cue gives an agent its own email, phone number, wallet, and computer, so it can do things online without always borrowing your identity.

They’re all personal agents, but they’re trying to become useful at different moments in your life. These small product differences matter a lot more to me than their position on a model leaderboard.

Because, honestly, most of us don’t live the life you see in personal-agent demos.

We’re not constantly booking flights and rearranging dinner reservations. We’ve had a long day at work, we need to buy something for the house, deal with a return, and maybe sort something out for the kids. Then we watch YouTube or scroll for a while and go to bed.

So, where does an agent fit into that day? If I’m going to give it a second task, it probably has to help with one of those small things I run into again and again.

In a way, AI is already there. Recommendation algorithms decide a lot of what shows up in our feeds. They’re good at guessing what might catch our attention.

Continue the same example, a dinner idea from a TV show. That’s where the recommendation stops, and the work comes back to me: find the recipe, check what’s in the fridge, make a shopping list.

With agents, a sentence like “ask me before it orders” is doing a lot of work. To know what’s in my fridge, maybe I can send it a photo or connect my grocery list. To compare prices, it needs to sign in to a store. To buy anything, it needs my approval.

None of that is as neat as the demo makes it look.

If it stops asking me about every little step, I might as well do the shopping myself.

But if it spends my money without asking, I’m not giving it another job. It has to know when to keep going and when to hand the decision back to me.

The best model might know exactly what the recipe needs; however, a valuable product must make the entire handoff feel easy and safe.

If it gets that right once, the next time I see a dinner idea, I’m more likely to send it straight to the agent. That’s how it earns the second task.

For now, every time I catch myself thinking, “Forget it, I’ll just do it myself,” that’s a clear signal that the agent hasn’t quite earned it yet. But also an opportunity for the next generation of AI products to seize.