New York City built a chatbot for business leaders like you.
Restaurant owners. Landlords. Shopkeepers. In the fall of 2023, the city put an AI chatbot on its MyCity website so a small business owner could ask how city law applied to their business and get an answer on the spot. 🏙️
In March 2024, reporters at The Markup and THE CITY asked it the questions you would ask.
💰 On tips, it said an owner could take a cut of their workers’ tips.
💵 On cash, it said a store could refuse it. A city law from 2020 says otherwise.
🏠 On housing, it suggested landlords could turn away tenants who pay with Section 8 vouchers.
Followed, every one of those answers puts a business on the wrong side of the law.
The mayor at the time, Eric Adams, promised a fix: “We’re identifying what the problems are, we’re gonna fix them, and we’re going to have the best chatbot system on the globe.”
The city added a disclaimer telling users to “not use its responses as legal or professional advice.” It improved some answers. It narrowed the questions the chatbot would take.
Then it kept the chatbot running.
So the restaurant owner now had a tool for checking the law that told them not to trust it on the law. 🤔
It stayed up for almost 2 more years. On January 30, 2026, the new mayor, Zohran Mamdani, ended it during a budget announcement: “The previous administration had an AI chatbot that was functionally unusable. It was costing the administration around half a million dollars.”
Who was going to turn it off?
Ending the chatbot needed no new law and no vendor’s permission, because the city owned it.
I call this AI decision debt: the cost of authority that exists on paper but never gets exercised. When no one knows who can decide what, decisions stall, or they get made by default.
With AI, the default keeps going. The tool answers the next question while your team is still working out whose job it is to check. 💬
Which of these does your AI decide?
Maybe it’s the chatbot on your website answering a refund question. The scheduling tool deciding who works Saturday. The screening tool moving some applicants forward and leaving others behind.
3 questions fit any role:
- What does the AI say in your name?
- What does it let through without a person checking?
- What does it decide about people?
Your team approves what comes out because it is usually right.
But the team didn’t decide which results need a person’s eyes before they go out.
The owner thinks the team is checking. The team thinks the owner is.
Goldman Sachs surveyed 1,256 small businesses in its 10,000 Small Businesses program in early 2026. 76% use AI. Only 14% say it is fully embedded in their core operations. 📊
I read that gap as the place this debt builds.
Sunk cost keeps the tool running
Ending something you paid for can feel like admitting a mistake in public.
A tool the business has paid for carries a specific fear: we spent all this money, and now we’re going to throw it away?
The money is spent either way. If the tool isn’t doing the job, ending it stops the spending and any issues it causes from here on.
The dividend: every AI decision has an owner
Same tool, decisions owned. When a tripwire fires, the decision goes straight to the person who owns it, and nobody has to work out whose job it is first.
You know who can sign a check in your business, without looking it up. The AI Decision Board gives every AI decision that same kind of answer. ✍️
Activity: Bring everyone who uses the tools, or will, and whoever answers for the results. Then work through 4 steps, in order.
👉 If you’re using AI now: start with the AI output your team approves most often.
👉 If you’re new to AI: put the first tool on your AI Decision Board while you set it up, so its decisions have owners from the start.
The decisions and the levels
- List the decisions. Write down each decision your team makes with an AI tool’s help, again and again: the answer a customer gets, the invoice that gets flagged, the applicant who moves forward, the subscription that renews. Your AI inventory from Edition 07 gives you the list to work from.
New to AI: list the decisions the first tool will touch, before it arrives. - Mark who decides today. Each person marks every decision on their own, then the team compares:
- A person decides. The AI can help draft or research, and a person makes the call.
- A person approves. The AI recommends, and a person approves each result before it goes out.
- A person reviews afterward. The AI’s results go out first, and a person checks them after.
Mark what happens in practice. Where the marks differ, the decision has no owner yet. Where no person sees the results at all, the team takes that decision first.
New to AI: a person decides all of them today. Each person marks the level they would give the tool, and the team compares during setup.
The owner and the tripwire
- Choose the level and name the owner. The team chooses how much of each decision a person keeps, and names the one person who answers for it.
- Write the tripwire and the review date. The tripwire is the result that sends a decision back to the people who own it. The review date is when they look again, even if nothing trips. At every review, the team answers one question: knowing what we know now, would we choose this tool today?
New to AI: set the first review date during setup, when the team names the owner.
I put my work through the same 4 steps
My AI writing assistant has an AI Decision Board too. Fixes with a written rule behind them, like a banned word, sit at “a person reviews afterward,” and I can reverse any of them in one reply. Anything that changes what a piece says sits at “a person approves.”
My tripwire is an edition that takes more than 5 drafts. My last one took 6, plus 2 patch versions. 😅 So yes, the tripwire has tripped once, and the review comes after this edition and the next 2.
Which decision is yours?
With a tripwire of any answer shown to break a city law, New York’s decision comes back to the city on March 29, 2024, the day the reporting ran.
So here is the question in the title, about your business: what does your AI decide today that your team didn’t agree it could?
P.S. The AI Decision Board is becoming a web page, with starter decisions for 7 roles and a reminder on your review date. I’ll share it in a follow-up post.
Every edition of this series is at digitalbard.co/perspectives.