We spent decades designing for people. Now people are sending agents in their place, and the systems on the other end aren’t sure what to do.
The call nobody wanted to take
A Meta employee asked the company’s new AI agent, Muse, to call his insurance company. The insurer kept hanging up. The agent wasn’t rude or confused. It had simply done the honest thing and said it was an AI.
According to internal posts Reuters reported on last week, Meta tested a fix with half its employees: hand some of those calls to trained human contractors, without proper disclosure. Staff raised privacy concerns and Meta rolled it back. Think about that for a second. The AI that was honest about what it was got hung up on, and the workaround was a human with less disclosure, not more.
The story isn’t really about Meta. Google is testing Call for Me with some Pixel 11 owners with a Google AI subscription: the assistant phones a business, says it’s an AI calling on your behalf, waits on hold, and handles the conversation while you follow a transcript. Anyone who has spent forty minutes listening to a looping jingle while trying to change a flight will see the appeal. OpenAI, meanwhile, pitches GPT-6 Astra on filling out online forms and updating CRM records. Nearly all the attention is on the caller. Very little of it is on whoever answers.
Picture a restaurant host taking one of those calls this week. The voice says it’s an AI assistant calling for a customer, asks for a table for four on Friday, and checks whether the kitchen can handle a nut allergy. The host confirms, writes it in the book, and the booking takes the same table as any other. Nothing about that phone line was set up with a machine in mind. Like the checkout page, the support queue and the contact form, it rests on an assumption so basic nobody wrote it down, which is that a human would be the one showing up. That assumption is starting to crack, and most teams haven’t noticed yet.
Speed bumps nobody slows down for
I’ve spent close to two decades working with people who build things for the web, across engineering, QA, design, and search. Add up every design review, usability test, and heated argument about button colours I’ve sat through, and nearly all of it comes back to one question: what will a person do here, and where will they give up?
We got good at answering it. People skim, skip instructions, and will sit through a long form if a progress bar promises the end is near. We also learned to add friction on purpose, like the “are you sure you want to cancel?” dialog, the CAPTCHA, or the second tap before a payment, placed exactly where we wanted someone to slow down and think.
So what happens to those speed bumps when the visitor is an agent? Some turn into obstacles. A CAPTCHA built to stop bots will now stop the bot your customer sent on purpose. That’s a real problem, but it’s the loud one, and people are already arguing about it. The quieter issue is what happens when the speed bumps work exactly as designed and still do nothing. An agent clicks “yes, I’m sure” without hesitating, because hesitation was never part of how it works.
The confirmation step is still on the screen, but the pause it was meant to create no longer happens there. It moved upstream, into whatever the user typed into WhatsApp three hours earlier while half-watching a cricket match, and into whichever moments the agent decides are worth checking back about, like Muse asking before it makes a purchase. The interface used to be where the decision happened. Increasingly, it’s just where the decision gets carried out. The website decided where you had to stop and think. Now the agent does, and you probably never saw where it drew that line.
Accountability gets murkier too. If the agent books the wrong night, whose mistake is it: the customer who typed the request, the agent that interpreted it, or the restaurant whose phone menu it misread? There’s no settled answer yet.
Reading tea leaves someone else left
The part I find most underrated is what this does to the numbers we rely on. Every product team I know reads user behaviour like tea leaves. A drop-off on step three of the signup flow means something is confusing there. A spike in time-on-page means people are engaged, or stuck, depending on who is presenting that quarter. We’ve built careers on reading bounce rates, scroll depth and rage clicks as if they told us what was going on in someone’s head.
Agents leave traces too, but they mean something else entirely. An agent that fills a form in four seconds isn’t delighted. An agent that abandons a flow isn’t frustrated; it may simply have hit an ambiguous label and gone back to its user to ask. As this traffic grows, gradually rather than all at once, our dashboards will quietly start blending two very different populations into one set of charts. Teams will eventually run A/B tests where some of the “users” have no feelings about the colour of the button. We’ll ship redesigns to fix confusion that a model experienced and no human did.
To be fair, agent-mediated journeys are still early enough that this is easy to dismiss, and plenty of agent activity will be identifiable if you look for it. But the operative words are “if you look.” Measurement problems rarely announce themselves. They show up as a slow drift in what the numbers seem to say, and by the time someone questions it, a few roadmaps have already been built on it. Garbage in, garbage out is as old as data itself, but this is a different problem: the data may be perfectly accurate, and what breaks is our sense of what the behaviour behind it means.
The unglamorous fix
If there’s a practical takeaway here, it’s a surprisingly old one. Some browser agents already navigate a page through its accessibility tree, the same structure a screen reader relies on, while others work from screenshots much as a sighted person would. So the sites agents handle best are likely to be the ones already built with care for people who can’t rely on visuals: clear labels, buttons that say what they do instead of “Continue,” error messages in plain words. Accessibility advocates have been asking for exactly this for years, usually while getting politely deprioritized in sprint planning. It may turn out to be one of the cheapest ways to get ready for agents.
Search works the same way. An agent choosing a plumber on someone’s behalf won’t be won over by a hero image if it can’t find your opening hours.
A wall that moved
None of this alarms me. It feels more like noticing that a wall in a house you’ve lived in for years has shifted a few inches.
For most of my working life, “who is the user?” was the most useful question you could ask in any room where something was being built. It cut through arguments and ego. I still think it’s the right question. But for the first time, the honest answer can be “there isn’t a person in front of the screen at all,” and very few of us have started designing for that answer.
Sooner or later, that insurer will have to stop hanging up. Some calls will come from people and some from agents acting for them, and the claim still has to get processed either way. Maybe the question worth asking now isn’t only “who is the user?” It’s “where is the human?”
