Chatbot Doom Loops Under Fire: What NYC's New AI Customer Service Rules Mean for Your Business

Most of us know the feeling. You type a question into a support chat, get a cheerful answer that has nothing to do with your problem, rephrase it, get the same answer again, and somewhere around the fifth message you realise there is no human behind the curtain and no obvious way to reach one. New York City has decided this experience now deserves its own set of AI customer service rules, and businesses that ignore them could soon be paying for it.

On 20 October 2026 the city's Department of Consumer and Worker Protection proposed a rule that targets what officials call automated "doom loops". If adopted, it would make New York the first city in the United States with the power to fine companies whose chatbots trap customers in circles. Penalties start at 525 dollars per violation, and businesses would also owe restitution to the people their bots failed.

The proposal is not yet law, and a 30-day comment period is still ahead. But the direction is hard to miss, because regulators on both sides of the Atlantic have been moving the same way all year. In this article we look at what New York actually proposed, how it fits the wider picture, and what any business running a chatbot should do about it.

What New York Is Actually Proposing

The rule was published in the City Record on 20 October and announced by DCWP Commissioner Samuel Levine, who previously ran the FTC's Bureau of Consumer Protection. It covers automated customer service in a broad sense, so both website chatbots and automated phone systems fall under it.

The requirements themselves are short enough to fit on a sticky note. A business must tell customers when they are talking to an automated system rather than a person. It must give them an easy way to reach a human representative. And it cannot dodge responsibility for unlawful things its AI says or does, which closes the tempting loophole of blaming the bot.

There is one more piece that deserves attention. The rule would make it a deceptive trade practice to advertise customer service and then repeatedly fail to address inquiries, or to respond only with irrelevant or erroneous information. In other words, a support channel that exists mostly for show becomes a legal risk in itself, not just an annoyance.

After the comment period and a public hearing, DCWP will review submissions and decide on the final text. Nothing is enforceable today. Still, companies that waited for final text on similar consumer rules in the past usually ended up doing the same work later, in a hurry.

Why Regulators Picked On the Doom Loop

Customer service was one of the first places where businesses deployed AI at scale, mostly because the savings are so easy to count. A bot answers at 3 a.m., never calls in sick and costs a fraction of a support team. The trouble is that the quality often lags far behind the cost savings, and customers are the ones who absorb the difference.

The city's press release cites a Gartner study which found that chatbots fully resolve between 17 percent of cases for billing disputes and 58 percent for simple returns or cancellations. Flip those numbers around and the problem is obvious: for the hardest issues, more than eight out of ten people hit a wall. When there is no path to a human, that wall becomes the whole experience.

There is politics in it too. Mayor Zohran Mamdani framed the rule as a response to corporations "stealing your time, money, and patience", and it builds on the city's earlier subscription-trap ban that took effect in 2026. Consumer protection agencies have discovered that bad AI support is a complaint category of its own, and complaint categories tend to become rules.

AI Customer Service Rules Are Spreading Beyond New York

New York's proposal gets the headlines this week, but it joins a map that has been filling in for a while. The most consequential entry is the EU AI Act, whose Article 50 transparency obligations became enforceable on 2 August 2026. Any system that chats with people in the EU must make clear it is an AI, no matter where the company behind it sits, and fines can reach 15 million euros or 3 percent of global turnover.

In the United States the picture is a patchwork, which is exactly what makes it dangerous to ignore. California has required bot disclosure in commercial contexts since 2019 under its B.O.T. Act. Maine passed a Chatbot Disclosure Act in 2025 that applies whenever a reasonable consumer might mistake the bot for a human. New Jersey ties disclosure to sales and advertising, and Utah requires it in consumer transactions, with stricter rules for sensitive sectors. Above all of it sits the FTC Act, which already treats a bot pretending to be human as plain old deception.

Taken together, the rules differ in detail but agree on the core: say it is a bot, do not let it impersonate a person, and stay accountable for what it tells your customers. A business that builds to that standard is in decent shape almost everywhere.

Who Should Pay Attention

The instinctive reaction outside New York is that this is someone else's problem. That is usually wrong, for a simple reason: consumer protection rules follow the customer, not your office address. If your online store sells to people in New York, or your SaaS product has users there, the city's rules can reach you the same way the EU's rules already reach any site with European visitors.

Small businesses are not exempt either. Most chatbots today are not custom builds but off-the-shelf widgets wired to a language model, set up in an afternoon. The legal responsibility still lands on the business that deployed the bot, not the vendor that sold the widget. If the bot quotes a wrong price or invents a refund policy, that promise is yours to honour or to litigate.

Agencies and developers who build chat experiences for clients should pay attention for a different reason. Disclosure banners, escalation flows and conversation logging are about to become standard requirements in briefs, the same way cookie consent did. Teams that can design those well will have an easier conversation with clients than teams bolting them on after a complaint.

How to Build a Chatbot That Passes the Test

The good news is that compliance here mostly overlaps with what a well-designed chatbot should do anyway. The first step is honest labelling. Say clearly, in the first message, that the customer is talking to an automated assistant. Burying it in the terms of service does not count, and a human name with a stock photo avatar is the opposite of disclosure.

The second step is the off-ramp. There should be a visible, working way to reach a person, and it should not require winning an argument with the bot first. Good escalation triggers include a direct request for a human, two or three failed attempts at the same question, and signs of frustration in the customer's messages. Test the path yourself once a month; off-ramps have a habit of silently breaking after platform updates.

Third, treat the bot's words as your words. Constrain it to answer from your actual policies and product data instead of letting a general-purpose model improvise. Log conversations and review a sample regularly, because that is how you find the places where the bot confidently misleads people before a regulator or an angry thread on social media finds them for you.

Finally, measure resolution honestly. A conversation the customer abandoned is not a resolved ticket. If your dashboard says the bot handles 80 percent of inquiries but refund complaints keep climbing, you already have a doom loop, whatever the metrics claim.

The Quiet Upside

It is tempting to read all this as yet another compliance cost, but there is a competitive angle to it. Customers do not actually hate chatbots. They hate being lied to and being trapped. A bot that introduces itself honestly, solves the simple cases fast and hands over the hard ones gracefully leaves people more willing to buy again, not less.

Escalation data is also an underrated source of product insight. The questions your bot cannot answer are a ranked list of what confuses your customers, whether that is unclear pricing, a hidden cancel button or a shipping policy nobody can find. Companies that review that list monthly end up fixing the website problems that created the support tickets in the first place.

Final Notes

New York's rule still has to survive a comment period, and its final shape may change. What will not change is the direction: regulators now treat the chatbot on your website as part of your customer service promise, with the same accountability that comes with it.

The sensible move this quarter is a short audit. Check whether your bot discloses itself, whether a human is genuinely reachable, and whether anyone on your team reads the transcripts. It is a few days of work now, and it is a lot cheaper than retrofitting all of it under a deadline with fines attached.