The first job people give an agent is getting money back

Of the 70 published sightings this site holds of an agent you can text doing something for somebody, 26 are the person getting money back: a bill talked down, a refund claimed, unclaimed property found, a subscription nobody remembered. Buying, the job the vendors build for and the payment companies partner on, is 11. Counted by what the agent did rather than by the job, 22 sightings have it chasing money the person was owed and 19 have it spending; where one of those was the main event, the chase wins 17 to 12.

What a sighting is

A sighting here is a public post in which somebody says an agent did a specific thing for them, filed with the job it belongs to and what happened, in the author's own words, with the post linked. It is not a review, a feature list or a prediction; those are voices. The use-cases wall renders all of them and each listing carries its own. On 17 September there are 70, on 6 products, 55 of them Muse's and 10 Instinct's, which is the first thing to know about the set and is returned to below.

What getting money back looks like

The carrier bill is the single most repeated errand. Instinct's users posted it first, $120 a year off Verizon negotiated overnight; Muse's followed in a week with AT&T fibre taken from $80 to $40 under a ceiling the user set in advance, a Verizon bill from $516 to $386 after the user went to bed, and $300 off a WiFi bill in under ten minutes. Unclaimed property is the second: three separate people had an agent search every state they had lived in and file the forms, for $97, for over $800, and $1,750 for a father-in-law, after Instinct's users had posted the same job the week before.

Then the refunds and claims that need a form and a fight: an Alaska Airlines claim filed past an upload widget that lied about its file limit, $200 back from an airline by email after a delay, a year of a forgotten book subscription refunded after the agent read the return policy, a $100-a-month subscription found in six months of receipts that duplicated a licence the company already held.

The two largest figures on the record are insurance, not bills: a policy re-shopped, bought and the old one cancelled for a claimed $3,500 a year, and the same job copied by a second person the next week for $1,156.

Every number in those three paragraphs is the user's. This site records them as claims with the post beside them and adds none of them up, because a sum of other people's figures is a headline nobody can check.

What the vendors sell

The products sell spending. Muse launched with Shopify's catalogue and Stripe's Link inside it, both announced by the partners on launch day. Instinct pays through a wallet that hands it a one-time card. Three of the tools on this site's builder page exist to give an agent a card without showing it the number, and the piece on buying counts which listings can spend at all. Agentic commerce is the phrase the industry chose, and it is a phrase about the agent as a customer.

The sightings say the agent's first use is as a debt collector for its owner. That is the other half of the commerce story: the same connections that let an agent pay, the inbox, the cards through Plaid, the browser that gets through a phone tree, are the connections that let it notice what you are being charged for and go and get it back. One user's Muse, connected to a bank account, found a charge for somebody else's Adobe account on his card that the bank had never flagged; it could not fix it, because the account was not his, and the sighting ends where the dispute begins.

Why the chase comes first

Three reasons are visible in the posts themselves. The job is bounded: a bill has a number on it and a lower number is the win, where a purchase has taste in it.

The job is one people have already decided to do and not done: the insurance user says he had abandoned several attempts because the aggregator sites were painful, the book subscription had run for a year, the unclaimed property had sat in other states for as long as the person had lived away from them. And the job rewards persistence rather than judgement, which is what the products are good at: seven minutes of trying, a phone tree, a form that needs three attempts.

The failures on the record point the same way. 3 of the 70 went wrong, and the clearest of them is a judgement call, Muse recommending a restaurant that had closed two years earlier and then trying for seven minutes to book it. Nobody has yet posted an agent that chased a refund it was not owed.

The shape of the evidence

55 of the 70 are Muse's, and that is partly the product and partly two things about the week. Meta's chief AI officer ran a challenge asking users to post what Muse had made or saved them, which is a request for exactly the kind of post this piece counts; those entries are filed with the frame stated. And a crowd-sourced collection of Muse posts was read in full on 17 September, which is why Muse's set grew by 42 in a day. 6 of the 70 come from people whose profile puts them inside the company that makes the product; they are filed as the company's and never quoted.

So the finding is about what people post, in the first fortnight of two products, one of which asked them to. No public data measures what agents are used for; this is the best public record of what people say an agent did, and in it the money comes back before it goes out.

What is not known

Whether any of the claimed savings held: a negotiated bill can be re-raised, an insurance policy bought in five minutes can turn out to differ from the one it replaced, and no sighting reports a month later. What the agent said to the carrier, which only a transcript would show, and only one person has described one. And whether the balance changes once the challenge ends and the novelty does, which is what the watches below are for.

What to watch

Whether spending overtakes chasing as the main thing the agent did, which would say the commerce story has caught up with the collections one. Whether a second product's sightings reach twenty, so the finding rests on more than one company's users. Whether the sightings that went wrong pass ten, at which point they are a piece of their own. And the first post that says a saving did not last.