Suppose you need a suitcase.
The old one has a broken wheel. You go online, compare a few replacements, read reviews and check the overhead-compartment dimensions. Somewhere along the way, the store shows you packing cubes, a travel pillow and a luggage tracker.
You came for a suitcase. The store has other ideas.
Now give the job to an AI assistant: find a carry-on within this budget, check the dimensions and return policy, and buy it once I approve.
For you, that could save an afternoon. For the store, it could change what happens during that afternoon.
For the people whose work sits between wanting the suitcase and receiving it, the question gets bigger.
The appeal is obvious
The personal-agent market feels like a gold rush. Everyone wants to show what their assistant can do: manage the inbox, arrange the trip, make the call, complete the purchase.
These agents use websites and other services to carry out tasks with the permissions you give them. How much they can finish varies.
Meta’s description of Muse includes shopping and other tasks on someone’s behalf, with approval for sensitive actions. That tells us what Meta is offering. It doesn’t establish how reliably every task gets done.
Still, I understand the appeal. A lot of work we would gladly hand over consists of finding information, comparing it, entering it somewhere else and checking whether anything happened.
Those steps also support businesses. They create opportunities to sell something, charge for help or persuade a customer to stay.
If people start delegating regularly, some of those opportunities could shrink. Others could move.
What happens to “you might also like”?
I’m interested in the purchase that never gets suggested because the customer never browses.
An assistant sent to buy a suitcase may have no reason to look at travel pillows. It might finish the job without following the route the store designed for a person.
That could affect impulse sales. It could also remove useful discovery. Sometimes the thing beside the thing you came for is something you need.
But we shouldn’t announce the end of impulse buying—or assume an agent chooses on facts alone.
In controlled marketplace experiments, researchers found that AI buyers responded to product position, platform endorsements and changes in descriptions. Different models made different choices.
That doesn’t predict next season’s sales at a real shop. It does make the idea of an entirely unpersuadable shopper hard to defend.
The commercial effort may shift toward influencing the assistant: getting into its sources, its shortlist or its preferred purchasing route.
Before celebrating the end of advertising, ask this: if the customer no longer sees the persuasion, how will they know it happened?
The charge you meant to cancel
Andrew Yang wrote about a colleague who used Muse to cancel a gym membership without visiting or talking to someone. His broader argument is that AI can remove inefficiencies that businesses and jobs depend on.
It’s one anecdote, but the appeal is easy to understand.
Giving software responsibility for your life is a big proposition. Getting an unwanted charge off your card is a smaller one.
An assistant could check renewals, compare bills or work out how to leave a service. Some businesses might benefit: a clearer offer could win a customer who finally has help comparing it.
Others may discover that continued payment was doing more work than continued satisfaction.
That is good reason to improve the offer. But the employee processing the cancellation may have had nothing to do with making it difficult.
The employee didn’t invent the paperwork
Some friction exists to discourage a customer from acting. Some exists because systems don’t fit together, records are wrong or a request needs judgment.
Someone reconciles the bill. Someone corrects the booking. Someone checks the documents and chases what’s missing.
We can want those tasks to become easier without pretending the people doing them are unnecessary. They may be keeping a badly designed process working.
If an agent takes on part of the work, the employee might handle more cases or spend more time on difficult ones. The employer might reduce staffing. A service might become affordable to people who couldn’t previously use it.
We need evidence to know which is happening.
Anthropic’s Economic Index found AI use clustered in particular tasks within occupations. Only about 4% of occupations used it across most of their tasks.
That describes how people use Claude. It doesn’t tell us how many positions a valley employer will cut.
But it gives us a useful discipline: ask what work is changing before announcing that an occupation is disappearing.
Then ask where the savings go.
Here, the savings have an address
In the Coachella Valley, this could reach a hotel reservation office, a property manager, a clinic’s billing desk or the person handling a small retailer’s orders.
Those are places to investigate, not jobs we can declare lost.
A visitor’s assistant could make finding and booking a local business easier. The business would need accurate information and a usable way to complete the request. Being found, understood and booked are different problems.
One more thing you should know: my company, SunshineFM LLC, which is separate from this newsroom, sells AI visibility to businesses through a program called Get Agent Ready.
Getting the booking is only part of the change. What work did the assistant do along the way? Did it bypass a service someone used to pay for? Did it compare offers differently? Who chose which businesses it considered?
And whose assistant is it, in practice?
A recommendation can reflect the customer’s needs, a commercial arrangement or simply the services the agent can reach. A friendly conversation won’t make those distinctions obvious.
I want customers to spend less time fighting forms and chasing answers. I want our businesses to benefit from better systems. I also want the people doing the work in this conversation before their role appears in a spreadsheet as a saving.
Update, October 15: Shortened the headline and edited for readability with Sat’s approval. The argument, sources, archive links and business disclosure are unchanged.
“Optimization” is an easy word to put in a presentation.
Before we use it to describe what agents are doing here, name the task being removed, the person doing it and where the money goes afterward.