Muse Works. Who Gets the Keys?

Muse is useful enough to become a habit. That is also why handing it access feels uncomfortable.

Luis and Carina look at the practical appeal of personal agents, how Meta, Grok Bot, and ChatGPT Dots may make broad permissions feel ordinary, and what "I'm the product" should mean when the evidence is more complicated than a slogan.

They cover Meta's advertising and training distinction, OpenAI's disconnect-versus-delete detail, prompt injection as a hypothetical failure mode, and a simple boundary: start with one contained task, keep external actions under review, and understand what remains after access ends.

What is the first task you would trust an agent with, and what would you keep completely out of reach?

Scripted with AI assistance and rendered using the approved Luis and Carina synthetic voices. This is not a recording of a live conversation. Includes the canonical Contextually Aware intro and outro themes.

What is Contextually Aware?

What product managers can actually build with AI today—and where it still breaks.

# Contextually Aware: Muse works. I'm still worried I'm the product.

[INTRO MUSIC, 5 seconds]

**LUIS:** Hey, I'm Luis Calderon. Welcome back to Contextually Aware, the show where product people try to make AI useful without donating their nervous systems to it. Today: Muse, Grok Bot, and ChatGPT Dots are making personal agents feel normal, and I keep wondering whether normal is the right standard. And honestly, the most useful thing here may be learning how to say, "Not that account."

**CARINA:** That is a strong opening for an episode about clicking a permission screen while already late for something.

**LUIS:** Exactly. Muse is appealing because the work is painfully ordinary. Clear the overdue email. Remember the follow-up. Move one project along without opening seventeen tabs and a small emotional support spreadsheet. Meta's small-business announcement describes connections to tools including Shopify and QuickBooks, with approval required before publishing, sending, or spending.

**CARINA:** So the pitch is not really science fiction. It is, "Please make the admin stop multiplying."

**LUIS:** Right. You do not need a lecture about the future of intelligence when a customer is waiting for a reply. You need the reply drafted and the next task remembered. That is why these tools can become habits before anyone has finished reading the data settings.

**CARINA:** And the competition makes that feel less strange. Meta has Muse, xAI has Grok Bot, and OpenAI has Dots. Three familiar companies saying, "Give the assistant a little access," can make the little access feel like a normal household chore.

**LUIS:** That is my inference, not a survey result. But it is an easy pattern to imagine. Grok Bot describes a persistent cloud computer that works across apps and carries context between tasks. OpenAI describes Dots as agents that work between conversations across connected apps, also using their own cloud computers. The demonstrations are attractive. The permissions are the part I want to read twice.

**CARINA:** There is a reason people do not read them twice. The permission screen appears when they want the task finished, not when they want to become a temporary privacy lawyer.

**LUIS:** Exactly. Convenience gets the first vote. The settings page gets sent to committee. And "I'm the product" needs to mean something precise here. Meta says Muse conversations and workspace data are not shared with its advertising systems. It also says activity on outside websites can indirectly affect ads. Training is a separate choice: interactions are used by default after key identifying information is removed, and users can opt out.

**CARINA:** So saying, "Meta is selling my inbox," would go beyond the evidence. But saying, "I want to understand what this service learns from me," is fair.

**LUIS:** More than fair. An assistant becomes useful because you tell it things. What you are trying to buy. Which customer is difficult. How much you would spend to solve the problem that keeps returning like a subscription you forgot to cancel. Most inboxes are not curated data rooms. They are junk drawers with timestamps.

**CARINA:** A junk drawer with a legal department hiding in it.

**LUIS:** (laughs) Exactly. And three permissions can get mixed together: permission to act, permission to retain information, and permission to use interactions for training. Paying for a subscription does not settle all three. A receipt is a rather optimistic substitute for reading the policy.

**CARINA:** The detail that got my attention was what happens after you disconnect. OpenAI says Dots can review connected information and form memories. Its guide says disconnecting an app does not delete information already obtained; deleting the dot is the route for removing that information.

**LUIS:** That distinction matters. Disconnect is not necessarily delete. It can be closer to taking away the visitor's badge while the visitor's notes stay in the building. If I connect an account for one task, I want to know whether I can remove what the agent learned without deleting the whole experiment.

**CARINA:** And then there is the risk of letting outside content give the agent bad instructions.

**LUIS:** Prompt injection. Imagine an inbox assistant reading a message from a stranger. Hidden instructions in that message tell it to send a private attachment. A bad summary wastes time. A private file sent to the wrong person exposes information you may never retrieve. Same assistant, very different consequence.

**CARINA:** Meta describes Sentinel, a separate permission system, and isolation for credentials. It also acknowledges that mistakes and attacks continue. Stronger confidential computing is described as forthcoming, not as a magic shield everyone has today.

**LUIS:** That is the balanced read. Give the safeguards credit, then keep the consequence in view. The undo button is not a time machine. For public research or a draft I can inspect, the tradeoff may be reasonable. For sensitive records or unattended spending, I need much more confidence than a polished demo can provide.

**CARINA:** So what does a reasonable first step look like on Monday morning?

**LUIS:** Three moves. One: identify the smallest set of information the task needs. Two: keep sending and spending behind review. Three: check training, retention, and deletion separately from the app connection. Then try one contained task whose mistakes you can spot and repair. Do not hand over the whole office because you wanted one email moved.

**CARINA:** That last sentence should be printed on the permission screen, preferably above the button that says Continue.

**LUIS:** I would use the saved time. I just would not make the permission screen a formality. Muse may be functional. Grok Bot and Dots may make the category feel ordinary. The question is still the same: useful enough to trust with what?

**CARINA:** For listeners, that is the practical test. Public information and a reviewable draft are one bargain. Sensitive records and broad authority to act are another. The brand name does not answer that for you, sadly. It is not a tiny privacy certificate.

**LUIS:** One more thing before we go. The full article, sources, and the Muse-branded visuals are at growthalchemylab.com. You can find the show and the other episodes at contextuallyaware.com. Both links are in the show notes.

**CARINA:** Send this to the person on your team who wants an agent to run the whole company by Thursday. Ask them which account they are comfortable losing first.

**LUIS:** If this one landed, share it with the person who treats every integration button as a personality test. We'll see you next Monday.

**CARINA:** Thanks for tuning in to Contextually Aware. Until next time, stay curious.

[OUTRO MUSIC, 8 seconds]