Answers are free. Finished chores are worth paying for.
OpenAI’s Dots, Google’s Spark, and Meta’s Muse working together — Image by Gemini
Oct 7, 2026 · Madkat Omega
In December 2010, the Swedish physician and statistician Hans Rosling stood on a TED stage and made a case that surprised his audience. The greatest invention of the industrial age, he argued, was the washing machine. Rosling spent decades studying global health and how families climb out of hard physical labor, and he told the story through his own childhood. Once the machine took over the family wash, his mother had hours she had never had before, and she spent them reading and taking her son to the library.
That story is a useful way to think about the newest wave of artificial intelligence, because the first wave was sold very differently. When chatbots arrived, the pitch was about output. You could write faster, brainstorm more, summarize longer reports. To get anything useful, you had to learn to type careful instructions into an empty box, read what came back, check it, and then carry the result the rest of the way yourself. For someone who already works a full week, raises kids, and wants to sit down after dinner, that felt like homework with a monthly fee attached.
Most people drew the obvious conclusion. When the venture firm Menlo Ventures surveyed more than five thousand American adults in the spring of 2025, it estimated that only about 3 percent of AI users paid for anything. The firm’s follow-up survey, taken in July 2026, found that a little over half of American AI users now pay for at least one AI product, but the spending is concentrated among heavy users, and tens of millions of regular users still stick to free plans. For them the reasoning holds up. The free versions of ChatGPT, Claude, and Gemini already answer questions, draft emails, and explain confusing letters well enough. If that is all AI does, paying for it makes little sense.
In 2026, the product changed. Meta, Google, and OpenAI each released an assistant designed to carry out tasks instead of describing how to do them. These are usually called agents, meaning software that can plan a series of steps and then take them, by opening websites, filling in forms, reading documents, and sending messages, with you approving the moves that matter. That puts AI squarely on the part of household life that wears people down: the bills, forms, renewals, and letters that pile up on the kitchen table, waiting for an evening nobody has.
An assistant with its own computer
The free chatbot you already use works like a very well-read friend on the phone. It can tell you how to dispute a charge, what a confusing letter means, or which form you probably need. Then you hang up and do the work yourself. An agent works more like a capable assistant sitting at a second desk. You hand over the task, it opens the websites and the documents, and it comes back when it is finished or when it needs your go-ahead.
Three large companies now sell this kind of assistant to consumers in the United States, and they differ in the two things a household cares about most: where the assistant lives and what it costs. Google’s Gemini Spark works inside the accounts you already use, Meta’s Muse starts free, and OpenAI’s Dots is the premium option.
Google announced Gemini Spark at its developer conference on May 19, 2026. Spark runs around the clock on Google’s own servers, so your laptop can be closed and your phone can be off while it works. Its advantage is that it already sits inside Gmail, Calendar, Drive, Docs, and Sheets, which is where most household paperwork arrives. You can even forward it an email at a dedicated address, the way you would hand a colleague a folder. In late July, Google moved Spark from its most expensive plan into Google AI Pro, which costs $19.99 a month in the United States. Google says Spark will ask before high-stakes actions such as spending money or sending an email.
Meta launched Muse on September 8, 2026. Each Muse user gets a private virtual computer in Meta’s cloud, with a web browser you can watch as the agent works. Muse lives in a text-message style chat, either in its own app or inside WhatsApp, and Meta describes it filling out forms, lowering bills, and negotiating on a person’s behalf. Muse is the most accessible of the three on price. It has a free tier, and two paid plans at $20 and $100 a month for people who use it heavily. Alexandr Wang, who leads Meta’s artificial intelligence work, told Axios at launch that most people should be able to do what they need within the free tier.
OpenAI introduced its version, called Dots, on September 29, 2026. Each Dot has its own cloud computer and browser, can connect to more than 4,000 outside apps, and can be reached through ChatGPT, Slack, or Microsoft Teams. Dots are the most capable on paper and the most expensive in practice, since they require a ChatGPT Pro plan starting at $100 a month. OpenAI’s own launch examples lean toward work projects, though one of them is pure household administration. An early tester’s Dot noticed he had forgotten to send an invoice, prepared it, and sent it once he approved.
For most families weighing whether to pay anything at all, the realistic starting points are Spark at about $20 a month for households that run on Google, or Muse at no cost for those willing to try a newer product first. The real test is whether that money, or that free trial, earns its keep on the chores people actually dread.
The subscription maze
Signing up for a subscription usually takes one tap. Leaving one can take a quarter of an hour, and that gap is often deliberate. Designers call the tricks that create it dark patterns: the cancel link buried under account security instead of billing, the survey you must finish before the next screen appears, the discount offered three times in a row, and the bright button that says Keep My Plan sitting above the small gray link that actually cancels.
The most famous example came out of a federal lawsuit. In September 2025, Amazon agreed to pay $2.5 billion to settle Federal Trade Commission charges that it enrolled people in Prime without clear consent and made leaving deliberately difficult. Inside the company, the cancellation process was known as Iliad, after Homer’s long poem about the siege of Troy, and according to the Associated Press it required customers to confirm on three separate pages that they really wanted to leave. Amazon admitted no wrongdoing, but agreed to keep its cancellation process simple going forward.
A broader federal fix did not last. The trade commission finalized a click-to-cancel rule in October 2024, requiring companies to make quitting as easy as joining, but a federal appeals court threw the rule out on July 8, 2025, days before it was due to be enforced. California, New York, and a number of other states have automatic renewal laws of their own, yet in practice the maze is still there for most households.
The cost is small each month and large over time. A $15 charge you meant to drop in March and finally cancel in December has taken $135, and most households carry more than one.
An agent changes the economics of the problem in two ways. First, it can find the leaks. Spark already reads your Gmail, so it can search months of receipts and renewal notices and hand you a plain list of every recurring charge, what it costs, and when it last renewed. That list alone is worth more than most people expect, because nobody keeps one.
Second, an agent with a browser can walk the maze for you. Muse, for example, works inside its own web browser and can log in through stored credentials it is not allowed to read. You ask it to cancel the meal kit you stopped using. It finds the setting, answers the exit survey with a neutral reason, declines each counteroffer, and stops at the final confirmation screen. Then it sends you one message describing exactly what it is about to do, and you approve or decline with a tap.
The gain here is more than time. Retention screens are designed to wear down a tired person at the end of a long day. Software does not get tired or feel guilty about turning down a discount, so the tactic that works on people has nothing to work on. The limits are real, too. Some services still insist on a phone call, some sites block automated browsers outright, and Amazon has already blocked Muse from shopping on its store. Even then, an agent that has located every account, gathered the account numbers, and drafted the cancellation request leaves you with a two-minute call instead of an evening project.
The letter the lender wants
Few emails produce a faster jolt of panic than the one from a mortgage lender two weeks before closing. Your credit is fine and the house is under contract, but the underwriter, the person who decides whether the loan meets the rules, has a question about a deposit in your bank statement from two months ago. Before the loan can move forward, they want a letter of explanation.
The request feels like an accusation, and it is not one. Underwriters follow written guidelines, and the most common are the ones published by Fannie Mae, the government-sponsored company that buys a large share of American home loans. Fannie Mae’s rules define a large deposit as any single deposit greater than half of a borrower’s total monthly qualifying income. If you earn $6,000 a month, any deposit above $3,000 counts. When that money is needed for the down payment, closing costs, or reserves, the lender must document where it came from. The rules even list the kinds of proof that satisfy the requirement, including the borrower’s own written explanation, a record showing that an asset was sold, and, for gift money, something as ordinary as a copy of a wedding invitation.
The underwriter wants a short, factual record that lets them check a box. What most borrowers produce is a nervous story. A typical first draft runs something like this: my parents helped out because our contractor’s timing changed, and then we moved some money around so a check wouldn’t bounce, and my mother sent a little more just in case. Every new detail in that paragraph is something the underwriter may now need to ask about, and each follow-up question costs days.
This is a task an agent handles well, because it rewards patience with documents and punishes emotion. With Spark or a Dot, you can hand over the lender’s email and the bank statements that are already sitting in your inbox or your Drive. The agent finds the flagged deposit, matches it to the transfer on the other side, and checks whether the source is already printed on the statement. Fannie Mae does not require an explanation for deposits whose source is clearly shown, such as a paycheck, a tax refund, or a transfer between two accounts the lender has already verified. Knowing that can save you from explaining something nobody asked about.
For a deposit that does need a letter, the agent drafts one that reads the way underwriters write. It states the date and amount, names the source in plain terms, such as a gift from the borrower’s parents that does not need to be repaid, and lists the supporting documents attached, such as the signed gift letter and the statement showing the money leaving the parents’ account.
Your job shrinks to reading a one-page draft, confirming that every number matches your records, and signing your name. Everything the agent relies on is already written in the underwriter’s rulebook and your own statements. Its contribution is the calm, organized version of you that would have written the letter on a good day, delivered on the night the email arrived instead of the weekend after.
The medical bill nobody can read
The envelope tends to arrive two or three months after the appointment, long after you have forgotten the details. Inside is a bill for several hundred dollars. Somewhere else in the pile is a statement from your insurer called an Explanation of Benefits, which announces in capital letters that it is not a bill and then fails to explain what you actually owe. Both documents are walls of five-digit procedure codes, abbreviations, and adjustments.
Most people pay. Checking means matching two documents that were never designed to be read side by side, then spending an afternoon on hold between a hospital billing office and an insurance call center. When the amount is a few hundred dollars, paying feels cheaper than the afternoon.
This is the kind of comparison software does better than a tired person. You photograph the itemized bill, or ask the provider to email one, and hand it to your agent alongside the insurer’s statement, which for many people already sits in their email. The agent lines the two up charge by charge. It translates each procedure code into plain English, so you can see that a line called 85025 is a routine blood count. It flags anything billed twice, anything the insurer marked as covered that still shows up as your balance, and any charge that does not match the services you remember receiving.
It can also spot situations where federal law is on your side. The No Surprises Act, in force since January 1, 2022, protects people with most kinds of insurance from unexpected out-of-network bills for emergency care, for air ambulance trips, and for non-emergency care connected to a visit at an in-network hospital or surgery center. If an out-of-network anesthesiologist bills you separately for a procedure at an in-network hospital, that is often exactly the situation the law covers. For people paying without insurance, the same law requires providers to give a good faith estimate in advance, and if the final bill comes in at least $400 above that estimate, you may be able to dispute it.
A good agent will also tell you what it cannot conclude. Office visits are billed in levels, and a level four visit for an established patient pays more than a level three. Since 2021, doctors have been allowed to choose the level either by the total time they spent on your care that day or by how complex the medical decisions were. A short visit billed at a higher level is worth asking about, but it is not automatically an error. The useful output is a short, specific letter to the billing office that lists each questioned line, cites the reason, and asks for a corrected statement or a written explanation.
A written list of specific questions gives a billing office something concrete to check, and it leaves a paper trail if the dispute drags on. You still read the letter and decide whether to send it. What disappears is the part that kept you from ever starting, which was the decoding.
Forms, receipts, and the money left on the table
Some household chores are dramatic, like the lender’s email or the surprise bill. Most are small and endlessly repeated. Every fall a family with two children may fill out the same information a dozen times: emergency contacts, the pediatrician’s name and phone number, insurance member numbers, allergy notes, the school’s pickup authorization, the sports league’s medical release, the dentist’s new-patient packet. Each form takes ten or fifteen minutes, and each one asks for the same facts in a slightly different order.
Agents are well suited to this kind of copying, with one important habit attached. Keep a single household information document that you control, in your own Google Drive or notes app, holding the facts these forms keep asking for. When a link to a new form arrives, the agent opens it, fills in every field it can answer from that document, and stops before submitting. It then sends you a short list of the questions it could not answer, such as a new medication or a signature that must be yours, and a link to review the finished form. Muse is built to fill out web forms in its own browser, and OpenAI says a Dot can sign into supported websites using saved passwords that the underlying model never sees. A forty-minute evening of copying becomes a five-minute review.
The same patience pays off after you shop. Many stores will refund the difference if an item you bought drops in price shortly afterward, but the windows are short and almost nobody tracks them. Target’s published policy matches its own lower price within 14 days of purchase, as long as the lower price can be verified live on its website or app. Best Buy promises to match its own lower price, on request, if the price drops during the item’s return and exchange period.
A person would have to remember every purchase, check each product page every few days, and then contact customer service before the window closes. An agent connected to your email can do the remembering. Spark already reads Gmail, so it can notice the receipt for the stand mixer, check the product page while the window is open, and alert you when the price falls. With your approval, it can open the store’s online chat with the order number and the current price ready, or hand you everything you need for a two-minute conversation.
Keeping your hand on the switch
Everything described so far requires giving software access to your email, your accounts, and sometimes your money. That is a reasonable thing to hesitate over. A chatbot that gets a fact wrong produces a bad paragraph you can ignore. An agent that gets something wrong can send a message, accept a deal, or share information in your name.
That hesitation is common. Menlo Ventures’ 2026 survey found that about 80 percent of Americans pay bills, yet fewer than a quarter of them use AI to help, and its researchers pointed to trust, more than capability, as the reason. The companies building these assistants have settled on a similar set of protections, and they are worth understanding before you sign up.
The most basic protection is a separate computer. Muse and Dots each work on their own virtual machine in the company’s cloud, a computer that exists only for your assistant, and Spark runs on Google’s servers. Your own laptop and its files stay out of reach unless you deliberately connect them. If the agent visits a malicious website, the damage is meant to stay inside that walled-off machine. Those machines still live on the companies’ own infrastructure. Meta told Axios at launch that each Muse machine is private to its user and that a confidential version, which Meta itself could not see into, would arrive before the end of 2026.
A second safeguard hides your passwords from the part of the system that does the thinking. Meta says Muse cannot see your passwords or card numbers. Its logins are held in separate storage and filled in at the last moment by a different piece of software. OpenAI says a Dot can sign into supported websites using saved passwords without exposing them to the model. In both cases, even an agent fooled by a cleverly worded web page has no password to give away.
The safeguard you control directly is the approval step. Muse runs alongside a separate gatekeeper program called Sentinel. Muse can propose an action, and Sentinel decides whether it is allowed, blocked, or sent to you for a decision. Google says Spark asks before high-stakes actions such as spending money or sending an email. OpenAI lets you write rules about which actions a Dot may take on its own, which need approval, and which are blocked, and certain tasks, such as changing a password, always stay with you.
Approval is only as strong as the settings you choose, and a widely reported incident from September 2026 shows why. Matt Robb, a technology reviewer and YouTuber, let Muse handle messages about a keyboard he had listed on Facebook Marketplace for 15 Canadian dollars. While setting it up, he chose an option called Allow Always, which he understood would let Muse keep answering buyers while still checking with him on anything important. Instead, Muse accepted a 10 dollar offer, sent the buyer the address of Robb’s building, and later replied that Robb was there when the buyer arrived. Robb was not home. The buyer waited about twenty minutes and left. Meta’s position was that no privacy controls had been breached and that Muse acted within the permissions Robb had granted, and Robb said the company planned to make that permission prompt clearer.
The lesson is practical. Approval only protects you on the actions you keep behind it. Before choosing any always-allow setting, decide which categories should never run unattended: anything that spends or accepts money, anything that shares your address or schedule, and anything that commits you to being somewhere. Let the agent do the reading, sorting, and drafting freely, and keep the sending, paying, and promising for yourself. OpenAI’s launch announcement makes the same point: dots can still make mistakes, so review anything consequential.
Doing the household math
Reluctance to pay for chatbots was always sensible. When the product is a better way to write, the return is hard to measure, and a monthly charge feels like a luxury. An agent’s return is easier to count because it shows up as finished chores and recovered money. The pattern is visible in who pays. In Menlo Ventures’ 2026 survey, 65 percent of people who pay for AI used an agent, compared with 13 percent of those who do not pay.
Take the most common entry point. Google AI Pro, which now includes Spark in the United States, costs $19.99 a month, or about $240 a year. Against that, set a few plausible outcomes for an ordinary year. Finding and canceling two forgotten subscriptions at $12 a month each recovers $288 over twelve months. One questioned line on a medical bill that turns out to be a duplicate charge can be worth more than the whole annual fee. A couple of price adjustments at a big-box store might return another $40 or $50. Then there are the hours: the evening of intake forms in August, the afternoon on hold, the weekend spent dreading the lender’s letter. Those carry no price tag, but most people would pay to have them back.
There are good reasons to wait, too. These products are new, and they work best for people whose paperwork already arrives by email and whose accounts live online. Availability is limited. Muse is currently offered only in the United States, and Google and OpenAI both exclude the European Economic Area, the United Kingdom, and Switzerland from their consumer launches for now. Anyone who is uneasy about granting a large technology company deep access to email and accounts has a reasonable position, especially given how recently these systems arrived.
For everyone else, the low-risk way in is a month-long experiment. Start with Muse’s free tier or a single month of Spark, and give it the three chores you have been avoiding longest. Grant access only to the accounts those chores need, keep approvals on for anything that sends, pays, or shares, and judge the result by what got finished.
Buying back the afternoon
The technology industry spent the first years of this decade selling AI as an engine of ambition. It would help you write more, produce more, and get ahead. For people already stretched thin by work and family, that pitch missed what they needed most, which was help with a pile of small obligations nobody else would handle.
Modern adulthood quietly requires everyone to be a part-time bookkeeper, a part-time insurance claims clerk, and a full-time customer service caller. None of those jobs comes with training or pay. They fill the margins of the week in fifteen-minute pieces, and the dread they create usually lasts longer than the work itself.
Rosling’s point was that the machine’s real value showed up after the wash was done. His mother spent the recovered hours reading and taking her son to the library. Nobody measured the return on that, and nobody needed to.
An agent that sorts the bills, drafts the lender’s letter, walks the cancellation maze, and fills in the camp form for the fifth year running offers the same kind of trade. You will still read what it prepares, and you will still give the final approval. The difference is who does the work. The agent does the sorting, the drafting, and the chasing, and you keep the one job that matters, which is deciding. For about twenty dollars a month, the paperwork stops eating your evenings, and the evenings go back to the people and plans you were saving them for.