Pharmacy hold music had been playing for eleven minutes when Daniel did what every adult child of an aging parent eventually does: he handed the job to a machine. His father is 78 and takes four prescriptions, with refills landing on different days and the pharmacy closing at six. Daniel opened Muse on his father's phone, typed "keep Dad's prescriptions filled and move his cardiology appointment to a Tuesday," and went back to his salad. Four minutes later the agent reported everything handled. It took Daniel's father the rest of the evening to figure out what had happened to his Tuesday.
Nothing here is a story about a bad agent. By every available measure the agent did fine. This is a story about the part nobody prices into the demo: the checking. When an AI agent tells you the refill is ordered and the appointment is moved, somebody has to confirm it. Researchers at UC Berkeley measured what that confirmation is worth and the number should be taped to every family group chat in America. Adding a single verification step lifted a coding agent's task success by 15.6 points, with no model change at all, which makes verification the cheapest performance upgrade in the entire reliability literature rather than paranoia.
Here is the problem: verification is a skill, it bills by the hour, and the invoice lands on the people least equipped to pay it.
Who audits the agent when the user can't
Pew Research Center published its "Americans and AI 2026" survey this month, and the age split is stark enough to read twice. Fifty-one percent of US adults never use chatbots at all. Among adults 65 and older, 77 percent never touch them. Even in the 50-to-64 bracket a majority, 58 percent, abstain, while only a third of 18-to-29-year-olds are non-users. Pew also asked non-users why they stay away, and the answers deserve more respect than the industry gives them: 60 percent cite simple lack of interest, 54 percent cite privacy concerns, and 45 percent cite worries about accuracy. Fewer than a third say they just don't know how.
Read that 45 percent again, because the skeptics are responding to a real property of the technology. A Princeton team that studied 15 models across two agent benchmarks this year found that the single most common failure on financial tasks was an agent getting a transaction wrong while sounding completely certain it got it right, because sounding certain is free. Certainty is cheap. Those accuracy worriers are not confused about the product. They have correctly identified its signature defect from the outside, which is more than most early adopters have managed from the inside.
Put those two findings together and the picture sharpens: what the agent needs is an inspector. Seventy-seven percent of the oldest Americans have never operated anything like the thing they are being asked to inspect. Somebody's father is going to tap "approve" on a screen he does not fully understand, for a prescription change he cannot independently verify, reported by a machine that states its errors with total confidence, and that tap, a thumbprint on glass, is the entire safety system. Meta's own framing, from the September announcement of Muse for small business, is that nothing publishes, sends, or spends without the owner's approval, which is true and beside the point, because an approval is only as good as the judgment behind it.
The arithmetic nobody shows you
Let me make the cost concrete, because vague warnings about "the supervision burden" are exactly the kind of thing people nod at and ignore. Take a real weekly routine: refill two prescriptions, move one doctor's appointment, pay one bill. Doing it the human way costs about an hour a week. Phone trees, hold music, and the drive to the pharmacy. Call it 60 minutes.
Hand it to the agent and the execution collapses to roughly five minutes. Muse runs on its own computer, works through the calls and forms, and reports back. Now add the verification that the Berkeley result says you need. Two checkpoints: one after the money moved, one after the appointments changed. Somebody opens the pharmacy site, somebody checks the calendar, somebody reads the agent's summary with actual attention instead of the skim we all give push notifications.
For a fluent verifier, say the adult child who set the whole thing up, each checkpoint takes about three minutes. Total verification cost: six minutes against 55 minutes saved, leaving a net benefit of 49 minutes. That makes the supervision tax 11 percent of the gross savings, which is a genuine bargain.
Now run the same arithmetic for the 78-year-old himself. He re-reads the agent's summary twice because the wording is unfamiliar. He opens the pharmacy website and squints at a confirmation page designed by people half his age. He calls his granddaughter to interpret the approval dialog. Twelve minutes per checkpoint is not an insult; it is a description of unfamiliarity compounded by caution. Total verification cost: 24 minutes, leaving a net benefit of 31 minutes. Now the supervision tax is 44 percent of the gross savings. Same agent, same task. Four times the tax, for the person who needed the help most.
That is the optimistic version, because it assumes every error gets caught. Princeton says errors arrive wearing certainty, and catching them is itself a skill. If the fluent verifier spots nine errors in ten and the low-fluency user spots three, the residual risk, the wrong refill nobody notices, concentrates exactly where the user is least able to absorb it. Speed is not the scarce resource in this household. Judgment is.
A note on honesty before anyone quotes these numbers at a conference: no study has measured verification time or error-detection rates by age or digital fluency, so those per-checkpoint minutes and detection rates are my assumptions, chosen to be plausible rather than empirical. The arithmetic is shown so your family can substitute its own numbers, and what the cited research actually establishes is only the direction of the effect, which is where the discomfort lives: verification cost scales with fluency, agent speed does not, and a cost that scales with fluency while the savings stay fixed is regressive by construction.
The best argument against everything I just wrote
It comes from the strongest study in this whole debate, and it deserves to be stated at full strength. Economists at Stanford and MIT tracked 5,179 customer support agents as their employer rolled out a generative AI assistant, and the results are the opposite of a counsel of despair. Productivity rose 14 percent on average, with novice and low-skilled workers improving 34 percent while experienced workers barely moved. The AI took the best workers' best practices and spread them to everyone else, lifting customer sentiment and retention along the way. If that pattern holds for personal agents, the least fluent users are precisely who gains the most, and hand-wringing about their supervision burden is paternalism with footnotes.
Here is the distinction the paper forces, and it matters more than the headline number. In that study, the AI suggested and the human decided, with the customer on the line to correct course in real time, so every action was reviewed by construction. In that study the novice kept the decision seat, and her judgment about when a suggestion looked wrong had somewhere to sit. That is where the 34 percent came from. An autonomous personal agent inverts the loop. Flip the loop and the autonomous agent decides and acts while the human sees only the report. Her biggest asset, the sense that something looks off, has no lever left to pull, because the lever fired four minutes ago while she was eating salad.
So the finding is real and the regime is different, and the honest conclusion is not "don't give the agent to the novice." It is this: the paper's gains came with a human in the loop, and your parent's agent has no loop. Somebody has to build one, on purpose, before the prescriptions move.
What we don't know
The blind spots first, because this article makes a quantitative claim from qualitative parts. Nobody has measured how long verification takes by age or digital fluency, or how often different users catch agent errors; my tax arithmetic is illustrative, and the four-times figure moves if your family's numbers differ. Pew measured chatbot use, not personal-agent use, and the fieldwork predates Muse's September launch, so agent adoption among older adults could look different in either direction. That NBER study was customer support with a human in the loop, not autonomous agents acting alone. And no published study tracks what actually happens when a family delegates a parent's bills and prescriptions to an agent for a year. The data that would tell us whether families converge on workable supervision routines or quietly absorb the errors will not exist for years, because nobody is running the study, and until then every family trying this is both the subject and the lab. That household experiment is running right now, in millions of homes, without a control group.
The family playbook
None of this is an argument against the agent. Daniel's instinct was right; the pharmacy hold music is a genuine theft of human afternoons, and an agent that kills it is worth having. The argument is that the supervision has to be designed like the agent was designed: deliberately, up front, with named owners. Treat the agent the way a good operations manager treats a talented but unsupervised temp: a short brief, a defined lane, and somebody named on the paperwork who reads the output before it becomes a bill, a booking, or a pill bottle.
Start with reversible work. Summaries, lists, drafts, research. Let the agent earn the irreversible jobs, payments and bookings and refills, with a track record you have actually read. A month of correct grocery lists is worth more than any demo.
Name the verifier. Receipts go to the fluent family member, not whoever issued the command. If Daniel sets up the agent on his father's phone, the confirmations route to Daniel until his father can read them critically. If nobody in the family can do that job, the agent is not ready for irreversible tasks in that household; say it plainly and wait.
Treat the approval tap as the whole policy. Meta says Muse asks before sensitive actions, and the family rule should be stricter than the platform's: individual approvals for anything involving money or medicine, never "allow always" on someone else's account. One careless always-allow is how a helpful agent becomes an unsupervised one.
Brief it like a babysitter. One job, one page, what "done" looks like, what to do when stuck. Research on this site found that standing instructions only half-stick, so keep the brief short enough to survive contact with reality, and put the non-negotiables, the pharmacy name, the "never move cardiology off Tuesday," where they cannot be missed.
Budget the supervision like labor, because it is. The $20-a-month Power tier is the cheapest line item in this arrangement. What it actually costs is somebody's Tuesday evening, spent reading an agent's homework. Price it honestly when you decide the agent is "saving time." For Daniel's father, the honest price still leaves 31 minutes of savings a week. That is a good trade. It is just not the trade the demo advertised, and the difference between the two is where families actually live.