Your Bot Answered. Nobody Helped.
New York City wants to call that deception
Someone in New York would rather sit for three hours in standstill traffic on the Cross Bronx, in 95-degree heat, in a car without air conditioning, than return a missed call from their pharmacy.
They said so in a public comment on the city's rulemaking website, signed "Anonymous." Returning that call meant more than four minutes of yelling at an AI phone system just to leave a message with the information the pharmacy staff had asked for to fill a prescription. If there was an insurance or dosage problem, the loop left two options: show up in person, or give up on the prescription.
They were clear that they didn't need a direct line to a human being. A voicemail inbox would do.
Read that again. It is about the most modest thing anyone can ask of a business, and someone felt they had to ask the government for it. We have built phone systems that never stop talking and still can't take a message. The machine keeps responding. The person trying to get their medication goes nowhere.
Press 0 for a person
On October 20, New York City's Department of Consumer and Worker Protection proposed rules aimed at customer service "doom loops."
Nobody is banning chatbots. The city wants the service behind them to work.
Businesses would have to tell you at the start of a conversation that you are dealing with a machine. Promising customer service and then repeatedly and unreasonably sending people into menus that don't address their problem, or answering them with unresponsive, irrelevant, or erroneous information, would become a deceptive and unconscionable trade practice.
The route to a human is where the rule gets clever, because it isn't a flat mandate. If a business runs an automated system and offers no way to reach a person, the city presumes it is breaking the rule. The presumption is rebuttable, but the business starts from behind.
Old phone trees are covered too. The rule defines an automated system by who is doing the talking, not by whether it carries an AI label, so a phone menu older than some of the people stuck in it gets the same treatment as this year's shiny agent.
The way out is described in almost comically plain terms: a single step, such as "Press 0" or "say agent," that puts you in line for a human, along with an estimated wait.
If nobody is available or the line is unreasonably long, the business can offer an alternative instead. The hours when humans are on. A phone number. A live chat, a contact form, a text line, an email address. Or the ability to leave a message and request a callback, with a stated timeframe for the answer. That last one is the pharmacy customer's voicemail. It made it into the text, as one of seven options a business is free to choose from.
None of this guarantees an immediate conversation with a person whenever you ask. Even so, it's embarrassing that it needs writing down. A consumer-protection agency is having to explain that offering assistance should come with a plausible chance of receiving some.
A customer who disappears looks like a win
This is the part I'd want every executive buying a support bot to sit with.
Picture a bot judged mainly by how many conversations never reach a paid employee. The industry calls it containment. A customer gets a response, closes the chat, never asks for an agent, and the dashboard counts a win.
Now look at who else produces that exact result. The customer who gave up after the third irrelevant answer also closed the chat and never reached an agent. On that dashboard, the solved problem and the abandoned one look identical.
Count avoided human contact without checking resolution, and you reward failure as generously as success.
That's the trap in the efficiency pitch. If the company saves employee time by spending more of the customer's afternoon, the work didn't disappear. It moved across the counter. The customer is now doing unpaid troubleshooting for a service they already paid for.
Not everyone has an afternoon to give. Someone with time to spare keeps trying. Someone squeezing a refund dispute into a lunch break gives up. A process that rewards persistence makes your access to a remedy depend on how much free time you can surrender. Nobody has to design it to wear people down. It's enough to accept "fewer tickets reached a human" as proof the automation works.
The draft suggests the city has seen that dashboard. Among the figures businesses would have to compile at least once a year: how many automated interactions were resolved without a transfer to a human, and how many involved a customer who had already been through the system in the previous 72 hours. The first is the number the dashboard loves. The second is the one that checks it. It won't catch the customer who gave up for good, but it catches the one who came back angrier. And the city asks for "resolved," a word a containment rate never has to earn.
Your refund needs someone who can approve it
This isn't only about irritation. The city's own justification for the rule lists real financial damage: unresolved billing errors, unauthorized charges, refunds people can't obtain, accounts they can't get into. It also cites a 2023 Gartner survey in which customers who used a chatbot got returns and cancellations resolved 58% of the time, and billing disputes only 17%. The moment a customer is arguing about money already taken is exactly where the bots did worst.
That changes the value of what you bought. Buy a defective product, fail to get a refund, and the cheerful chat window hasn't returned a cent. The draft makes the same argument in legal language, as a gross disparity between what you paid and what you received.
Violators would owe restitution to the people they harmed, plus fines starting at $525 for a first violation and climbing to $3,500 for repeat offenders. The fines aren't what I'd watch.
What I'd watch is a provision that has nothing to do with customer service. Tucked into the proposal is a new general rule: having used or relied on a technology is not a defense to any violation of the laws and rules DCWP enforces. Not just chatbots. Everything the agency polices.
The city had already said as much in September, in a joint statement with its Commission on Human Rights and its Taxi and Limousine Commission: businesses can't escape responsibility for a deceptive representation "because it was generated by a chatbot."
That statement announced no new legal standard. The proposal would write the principle into the rulebook.
Good. A company that presents its bot as a helpful representative while the bot is closing a sale can't treat the same bot as an unfortunate stranger once its answers cause trouble.
Somebody has to own the handoff
Reaching a human only helps if that person can do something. Picture a representative who takes your complaint but can't reverse the charge and has no real way to reach anyone who can.
You escaped the bot. The company simply staffed the dead end.
This is where the draft stops short, and it's worth saying plainly. The records it demands are useful. Businesses would keep them for at least three years, in machine-readable form, and hand them over on request. For every automated interaction, the logs would show whether a transfer was offered, whether the customer used it, and whether they ended up with a human, in a queue, or pointed to an alternative. That gives an investigator something better than a vendor demo in which the bot behaves beautifully. But the logs record whether the door opened. They don't record what was behind it.
Whether the person on the other end could approve the refund, whether the callback came within the stated timeframe, whether anyone owned the case: none of that is a field. No software purchase settles it either. Somebody inside the company has to own what happens after a customer asks to leave the automation: who answers the callback, who can reverse a wrong decision, who notices when the timeframe slips. Leave those questions open, and the customer becomes the one who finds the gaps, one wasted afternoon at a time.
If a company can't walk one ordinary complaint from "I want a person" all the way to an outcome, it should think hard about what it's calling customer service.
New York's rule is still a proposal, and the text can still change. Public comments are open until December 18, so if you have ever shouted "agent" into a phone, the city is asking. But the test doesn't need a final rule. When the chatbot has finished being helpful, can the customer actually get help?