Anyone who has typed “agent” into a chat window five times already knows the problem New York City now wants to write into its rules.
On October 20, the city’s Department of Consumer and Worker Protection proposed rules for what it calls automated customer service systems. The proposed text defines one as any technology where “all or substantially all the communication to the consumer is made by such technology rather than by a natural person.” That covers AI chatbots, and it also covers the phone menu that asks you to say your account number a third time. Comments are due by November 19, 2026, the same day as the public hearing. DCWP says that, once adopted, New York would be the first city in the country able to enforce this kind of protection.
What the proposal requires
The rule does four things. A business has to tell you at the start, in the language you’re using, that you’re talking to an automated system and not a person. It becomes a deceptive and unconscionable trade practice to promise customer service and then, “repeatedly and unreasonably,” send people to menus that don’t address their question or answer them with “unresponsive, irrelevant, or erroneous information.” Listing a phone number or email under “contact us” counts as a promise that service is available there.
The third piece carries the weight. If a business uses an automated system and doesn’t offer a way to reach a human, the rule presumes the business is in violation, and the business would have to rebut that presumption. The cleanest way to meet it is a single-step transfer, such as “Say or type ’agent’” or “Press 0,” with an estimated wait time if there’s a queue. If no human is available, the business has to give the customer another route to one, such as a phone number with hours or a contact form with a stated response time.
The fourth piece reaches past customer service. A new section, 5-02, says a business can’t use a technology as cover for breaking any law DCWP enforces, whether the technology made the claim, left out the information, carried out the act itself, or helped carry it out. A footnote in the notice says plainly that using a specific technology “has never provided a safe harbor” under the city’s Consumer Protection Law. Penalties in the proposed schedule run from $525 for a first violation to $3,500 for a third, and DCWP says violators would also owe restitution to the consumers they harmed.
The records are the real test
The section that gets the least attention is section 5-313. Any business using one of these systems would keep three years of records in a machine-readable format and hand them over when the department asks. For every automated interaction, the log has to show how long it lasted, whether the one-step transfer was offered, whether the customer used it, and whether the customer reached a person, waited in a queue, or got another route to help.
Once a year, the business would also compile totals: how many interactions, how many were resolved without a transfer, how many customers used the transfer, the average length, and how many came from a customer who had already contacted the system in the previous 72 hours. That last number is the one to watch. A customer who comes back within three days is telling you the first conversation didn’t fix anything, no matter what the bot’s own dashboard says.
Where the machine’s own checks run out
A chatbot vendor can test whether the bot answers quickly and flags replies it isn’t confident about. That’s Responsible AI, a machine checking a machine, and it has real value. What it cannot do is tell the business whether the customer who closed the chat window got the refund. The 72-hour count can, because it is taken from the customer’s side of the conversation.
The human transfer belongs to a different layer. AI Governance needs a person with authority to accept the machine’s answer, change it, or overrule it, and a record of what that person did. Basil’s Checkpoint-Based Governance is built around that requirement, and the proposal puts one version of it in the customer’s hands: one step to leave the machine. The text defines the transfer. It doesn’t say what the person who picks up must be able to do, so a business could meet the rule with a representative who can only read the bot’s script back.
The responsibility clause follows a line courts are already drawing. When a Munich court ruled that Google’s AI Overviews are Google’s own statements, Basil wrote in A Munich Court Rejected the AI Disclaimer Defense that the disclaimer didn’t transfer the duty and the company that deploys the system answers for what it says. New York’s draft reaches the same place through consumer protection law, and it names the business as the party that answers.
The complaint itself goes back years. In 2023 the Consumer Financial Protection Bureau described chatbot “doom loops” that leave customers “without an offramp to a human customer service representative.” On October 1, DCWP, the Commission on Human Rights, and the Taxi and Limousine Commission issued a joint statement that New Yorkers “do not lose their rights when a decision is made or facilitated by technology,” while noting it announced no new legal standards. The October 20 proposal is where the standard shows up.
Five steps before November 19
The rule is a proposal and may change after comments. Businesses serving New York City customers still have good reason to start now.
First, list every channel your site labels “customer service” or “contact us,” and confirm a person can actually be reached through each one.
Second, put the disclosure and the one-step exit at the very start of every automated conversation, then test it yourself as a customer would, on the phone and in chat.
Third, check whether your system can log the four interaction fields today: duration, transfer offered, transfer used, and the outcome. If your vendor can’t produce them, you’ll learn that before the city asks.
Fourth, pull your 72-hour repeat-contact rate now. If it’s high, the bot’s resolution numbers were counting closed chats as solved problems.
Fifth, decide in writing what the human who takes the transfer can approve, whether that’s a refund or a billing dispute. A transfer to someone with no authority just moves the doom loop to a person. Anyone who wants to weigh in can comment through the NYC Rules page until November 19.
Every chatbot already grades itself. Under this proposal, the city would grade it by a different measure: how many people had to come back.
Sources
- City of New York Department of Consumer and Worker Protection. (2026a, October 1). DCWP, TLC, and CCHR issue enforcement policy statement on artificial intelligence [Press release]. https://www.nyc.gov/site/dca/news/067-26/dcwp-tlc-cchr-issue-enforcement-policy-on-artificial-intelligence
- City of New York Department of Consumer and Worker Protection. (2026b, October 20). Mamdani administration proposes ban on AI customer service doom loops [Press release]. https://www.nyc.gov/site/dca/news/072-26/mamdani-administration-proposes-ban-ai-customer-service-doom-loops
- City of New York Department of Consumer and Worker Protection. (2026c). Notice of public hearing and opportunity to comment on proposed rules: Rules regarding automated customer service systems. NYC Rules. https://rules.cityofnewyork.us/wp-content/uploads/2026/10/DCWP-NOH-Proposed-Rules-Regarding-Automated-Customer-Service-Systems.pdf
- City of New York Department of Consumer and Worker Protection, Commission on Human Rights, & Taxi and Limousine Commission. (2026). Joint statement on artificial intelligence, algorithmic tools, and automated systems, and New Yorkers’ rights. https://www.nyc.gov/assets/dca/downloads/pdf/media/Interagency-AI-Enforcement-Letter.pdf
- Consumer Financial Protection Bureau. (2023, June 6). Chatbots in consumer finance. https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/
FAQ
What did New York City propose for AI customer service chatbots?
On October 20, 2026, the NYC Department of Consumer and Worker Protection proposed rules for automated customer service systems, including AI chatbots and phone voice-response menus. Businesses would have to disclose at the start that the customer is dealing with an automated system, provide real customer service where they advertise it, and give customers a way to reach a human representative. Comments are due by November 19, 2026, the date of the public hearing.
What happens if a business’s chatbot offers no way to reach a human?
The proposal creates a rebuttable presumption that the business engaged in a deceptive and unconscionable trade practice. The business can meet the rule with a single-step transfer, such as “Press 0” or saying “agent,” plus an estimated wait time, or, when no human is available, by giving another route to one, such as a phone number with hours or a contact form with a stated response time.
Can a business blame its AI system for a violation under the proposal?
No. Proposed section 5-02 says it is no defense to a violation of laws DCWP enforces that a technology made the claim, omitted information, carried out the act, or helped carry it out. DCWP’s notice says using a specific technology has never provided a safe harbor under the city’s Consumer Protection Law.
What records would businesses have to keep?
Businesses would keep three years of machine-readable records, including logs showing each automated interaction’s duration, whether a one-step transfer was offered and used, and whether the customer reached a person, waited in a queue, or got another route. They would also compile annual totals, including interactions from customers who had contacted the system in the previous 72 hours.
What are the penalties in the proposed NYC chatbot rule?
The proposed penalty schedule sets $525 for a first violation, $1,050 for a second, and $3,500 for a third or later violation of the customer service, disclosure, and recordkeeping sections. DCWP also says violators would be liable for restitution to harmed consumers.
How is a human transfer different from the chatbot’s own quality checks?
A vendor’s automated tests check the software, which is Responsible AI. The human transfer and the interaction records support AI Governance, because they let a person overrule the machine and leave evidence of whether customers were actually helped. The proposal does not define what the human representative must be able to approve, so businesses should set that authority in writing.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply