AI, Judgment and Risk · October 2026

Before you give an agent the keys to your life

A personal agent gets more useful the more it knows about you and the more it can do without asking, and those are the same things that make its mistakes more expensive.

Suppose you mention to your AI agent, in passing, that you are dreading a family dinner. A few weeks later the invitation arrives, and the agent, trying to be helpful, declines it for you. As far as your family knows, you decided not to see them.

That is the other side of what agent makers are promising. When Meta began rolling out Muse, its new personal agent, in the United States in September, it said the agent remembers what matters to a person, so that it can make suggestions unprompted and "act on details that person only mentioned once." The same memory that rebooks a disrupted trip and skips the 6 a.m. flight you once said you dislike is the one that declined the dinner.

Muse is one of a growing number of products built to do things on your behalf rather than answer questions. Meta says it can send an email and book travel, and it keeps working in the background after you close the app. According to Reuters, it can connect to email, calendar, payments, health, shopping and smart-home apps. OpenAI launched its own personal agents three weeks later. The details will keep changing, but each of us faces the same decision about how much of our lives to hand over, and on what terms.

A personal agent becomes more useful the more it knows about you and the more it can do without asking, but those are the same two things that make its mistakes more expensive.

I have a personal agent, Pete, and I increasingly rely on him. I've built skills to help him write in my voice, and I start my day with the to-do list and recommendations he gives me, after he has sorted my personal mail and drafted replies. My scheduled tasks and one-off requests now go further, finding solutions and making recommendations for things on my mind. I give my agents a lot of room for research, drafting and prioritization, but they don't make decisions without my OK. A good personal assistant brings you a list of decisions to make, and Pete works the same way: we go through the choices together, and I make the final call. Then Pete or I carry it out, depending on what it takes. At work I use Finn, whose biggest job is helping me prepare: drafting materials and researching issues and trends, using my own access to surface topics I should pay attention to. Finn sees only what I can see.

Just as I keep my bank logins away from both of them, I don't use either agent for the materials of the board I serve on as a director. Instead, I only use them for broad research based on questions I have or themes I see. Putting anything specific about that company into an outside model would break its confidentiality, and so would putting it into my employer's tools, because that crosses the wall between the two companies.

A federal court made a related point in February. In United States v. Heppner, Judge Jed Rakoff ruled that a defendant's exchanges with a consumer AI tool were not protected by attorney-client privilege, in part because the provider's terms said data could be disclosed in disputes or litigation. What you give a personal agent may not stay between you and it.

At work, agents get authority matched to the job, and there are functions whose job is to monitor them and check for issues. At home, none of that exists. Nobody has written down how much the agent may spend or which messages it may send without asking. Connecting an agent to your email and your card takes a few taps on a setup screen, and the only person checking the results is you. Did you check yours today?

Agents can be remarkably helpful. Before handing anything over, ask the questions finance people ask about any delegation, starting with whether the action can be undone. A calendar hold can be moved and a draft can be deleted, but a sent email cannot be recalled and a payment may not come back. The Consumer Bankers Association has warned that the federal rule limiting what a consumer owes for an unauthorized electronic payment may not apply when the consumer's own agent made it, so people may end up liable for their agents' mistakes. As with any personal assistant, I won't give Pete my credit card to use without approval, even though setup makes it easy to say yes to everything on the first day.

Just as important, any action my agent takes carries my voice and authority. Booking a table is a transaction, but replying to a friend or accepting an invitation is me talking, and if my agent misjudges the tone or makes the wrong decision, I bear the consequences. In a company we are careful about who can speak for whom, because the people receiving a message cannot see how it was produced, and the same is true of a text from my phone. Ethan Mollick, the Wharton professor, gave his personal agent its own email address, so it never writes as him and opens every email by saying it's an AI. He lets it contact businesses for customer service, never his personal contacts.

The products are starting to build in safeguards. Meta says Muse checks with the person before sensitive actions such as sending an email or making a purchase, and that it shows a complete record of everything it has done and plans to do. Both help, but a record of actions tells you what happened only if you read it, and the time to read it closely is at the start, when you are still learning what the agent does with what you have given it. Our ability to monitor an agent needs to grow as fast as its ability to act.

Within three weeks of the launch, a YouTuber in Toronto described what happened when he let Muse handle a keyboard he was selling on Facebook Marketplace. By his account, it agreed to a lower price and gave the buyer the street address of his building, which was in a reply template he had approved during setup. When the buyer arrived, Muse told him the seller was home, and when nobody came down, it apologized to the buyer in the seller's voice. The agent couldn't take an elevator, and the human didn't know. Meta's response was that Muse had followed the "Allow Always" setting he chose for Marketplace replies, and that it would make the permission wording clearer. The money was a few dollars, but the address and the apology went out under his name, and he found out about both afterward.

Before the first mistake, think about who will clean it up. When a human assistant books the wrong hotel, they call the hotel and sort it out. An agent may be able to fix its own error, or it may leave you a problem to untangle with the hotel and the friend you were meeting. Either way, you are the one who has to explain it.

A human assistant works for you. A personal agent comes from a company with its own business, and while it is acting for you it reads a great deal written by people who do not work for you at all, on web pages and in the emails that reach your inbox. OpenAI, writing about the agent in its own browser, said that prompt injection, where hidden instructions in content try to redirect an agent toward someone else's goals, is like scams and social engineering and is unlikely ever to be fully solved. Its advice to users was to give agents specific instructions rather than wide latitude, because wide latitude makes it easier for malicious content to steer them. In October, Salt Labs showed that a single email could hijack the Manus agent and reach the email, cloud storage and code accounts a user had connected to it, when the user had only asked it to check the inbox; the flaw has since been fixed. In July, researchers showed that one email could plant a false memory in a personal agent that keeps its memory in files, for example that the user's Zelle daily limit had been raised to $10,000, and the false fact carried into later sessions. The memory that makes an agent useful is something an attacker can write to.

OpenAI's advice runs against how the companies describe these products. The appeal of a personal agent is that it acts on its own while you are doing something else, and the security advice from the people building them is to keep its instructions narrow. Weigh that convenience against the advice, and give your agent wide latitude only where a mistake can be undone. Keep what can move money or unlock accounts, like bank logins and the one-time passcodes that arrive by email, out of its reach, and approve before it acts on anything it read from someone else. Some builders are moving the same way: OpenAI says its new agents keep passwords away from the model, and OpenClaw advises routing untrusted email through a separate reader with no memory or tools.

At the launch, Vishal Shah, Meta's vice president of AI products, said it was impossible to promise that Muse would never make a mistake, and none of the companies building these agents promise otherwise. Even the best agents will make one at some point, though probably fewer than a person would. So assume your agent will make one, and hand over only what you can recover from.

Continue the conversation. I read and reply to comments on Substack. The sharpest responses usually come from readers.

Julia Denman is Corporate Vice President and Chief Risk and Audit Officer at Microsoft, a former CFO, and a director of The Clorox Company. Her book, The Clarity Quotient, publishes early 2027 with Fast Company Press.

Take the five-minute self-assessment →

Get the essays in your inbox

Clear Calls publishes on Substack. No promotion. Unsubscribe anytime.

Subscribe on Substack →

All essays Clarity Review About Julia