AI · Risk · Small Business
AI Lied to Me for 3 Days. Now I Let It Run My Business.
AI lied to me for three days straight. It cost me a month of work and it never once sounded unsure. So sure, let's hand it the calendar, the inbox and the customers and go take a nap. I use this stuff every single day and I'd be out of my mind to give it up, but that weekend taught me the thing nobody selling you an AI tool says out loud. It cannot run blind. Andrew Yang spent this week asking Congress to put a kill switch on the big models, and nobody's watching the small one in your contact form either.
The Short Answer
No, I don't let AI run my business, and you shouldn't either. I use it every day and it earns its spot. It's fast, it kills the blank page, and my shop would be slower and smaller without it. It also lied to me for three days in the same calm voice it uses when it's right, and that's the whole problem: it cannot tell you when it's wrong. So it doesn't get to run anything. It gets to do work that somebody reads. What costs you money is blind AI, and blind doesn't mean dumb. It means nobody read what it said, nobody can undo it, and nobody approved it having that power. Yang is asking Congress for a kill switch on the big labs because nobody's watching those. You don't need Congress for the one in your shop. Four controls and an afternoon.
Source: Andrew Yang on CNBC, September 16, 2026, as transcribed by Decrypt, Jason Nelson, September 17, 2026.
3 Things to Remember
A license on the wall, insurance, an inspector. The model has none of it.
A Hot Dog Stand Has More Rules Than the Model.
Andrew Yang went on CNBC on September 16 and said the line I haven't been able to put down.
"If I were to open a hot dog stand out here on the streets of New York, I'd have hundreds of regs to comply with, and the models have none." (Andrew Yang, CNBC, September 16, 2026, as transcribed by Decrypt, September 17, 2026)
He's right, and if you own a business you felt that one in your chest. You've got a license on the wall. You've got insurance. You've got an inspector who shows up whether you're ready or not.
He asked Congress for three things. Liability when AI causes harm. A waiting period before a powerful model ships. And a kill switch. His words: "We need to regulate these companies and their offerings. We need a kill switch. We need real liability."
He says the people building it are asking for the same thing. "They're raising their hands and saying, please give us a guardrail, because I don't want to work on something that I think might cause irreversible harm."
And he says it isn't a party thing. "This is bipartisan. If you're in a rural area or red district, your constituents also are freaked out about AI."
Not everybody buys it. David Sacks has argued publicly that some of the AI safety noise is a move by AI companies angling for rules that lock in their own position. Yang's answer was that more than one thing can be true at once. That's fair, and I'd say the same.
Not the one in the Senate hearing. The one in your contact form.
That Fight Is About the Labs. Yours Is About Tuesday.
Washington is going to spend two years arguing about frontier models.
None of it shows up on your profit and loss statement.
What shows up on your profit and loss statement is smaller and it's already in your building. Something answered a customer at 9pm last night. Something wrote half a service page. Something has a key to your calendar. You didn't read any of it.
That's the AI I'd worry about if I were you. Not the one in the Senate hearing. The one in your contact form.
I Told You It Lies. Then I Told You to Let It Answer at 9pm.
Two of my own posts, side by side, and they look like they argue.
In AI Sucks Ass I told you it lied to me for three days and never once sounded unsure. I told you to check anything that turns into a fact, a number, a name, a price, or a promise. I meant it. I still triple check everything with my own eyes.
Then in let AI chase your leads I told you to let a machine answer a stranger at 9pm while you're asleep.
What connects them isn't that I got comfortable. It's that somebody is watching in both of them.
Go read the lead post again. You write the first text. You set the tone. You pick the questions. Six lines, your words. The AI doesn't compose anything. It sends what you already wrote, at a time you couldn't have sent it yourself.
That's a machine running an errand. It isn't a machine running your business, and that post never said it was.
Same story in AI doesn't make me a plumber. The tool builds the site. You own everything the site does. The name on the page is still yours. None of that moves because the typing got faster.
So the line was never automatic versus not automatic. Speed is fine. Overnight is fine. Doing it while you sleep is the whole point.
The line is whether anybody is watching.
Read it, own it, or don't ship it. That's the same rule in all three posts and it's the only one I've got.
The mistake: you sort AI tools by how much they do on their own.
The fix: sort them by who's watching. Words a human already read are supervised, whatever hour they go out. Words made up in the moment with nobody looking aren't.
The payoff: you keep all the speed, and you stop guessing about which tools are the risky ones.
A wrong answer is an event. A wrong answer nobody reads is weather.
Blind AI Isn't AI That's Wrong. It's AI Nobody Read.
A wrong answer is a normal business problem. Your front desk gets a price wrong, you hear about it in an hour, you fix it.
Blind AI is a different animal. Three things have to be missing at once, and when they are, a wrong answer stops being an event and starts being weather.
Nobody read it. No log, or a log nobody opens. The chat on your site had 240 conversations last month and you've read zero of them.
Nobody can undo it. It sent the email. It moved the appointment. It published the page. There's no version to go back to, because nothing wrote down the before.
Nobody approved it. Somebody switched a tool on in a settings panel while being helpful, and now it has a key to your calendar.
Any one of those on its own, you'll survive. All three together is a business running on a machine nobody is driving.
That's the difference between my two posts, in three bullets. The 9pm text passes all three. Something composing a quote from scratch at 9pm fails all three.
Website chat, your content, and the back office tools that act as you.
Three Places It's Already Running Blind in a Shop Your Size.
Your website chat. It's answering pricing, availability and service area questions right now. If it made up a price, or promised same day service in a county you don't cover, you'll find out when somebody is angry about it. Most owners I talk to have never read a single transcript.
Your content. AI wrote the service page, or half of it. It's confident and it reads well. And somewhere in there is a certification you don't hold, a city you don't serve, or a warranty term you never offered. That isn't a customer service problem. That's the call from your lawyer.
Your back office. This is the new one and it's the one that bothers me. Tools now book appointments, send follow ups, update your CRM and reply to email. Hand one of those the wrong login and it isn't answering questions anymore. It's acting, as you, with your name on it.
Yang's actual subject was agents: software that plans and carries out tasks with limited human direction. He was talking about them at the frontier. You're running small ones already.
The mistake: you think your exposure is whether the AI is smart enough.
The fix: write down every AI tool touching your business, and next to each one, what it can change without a human saying yes.
The payoff: you'll delete two of them by lunch, and what's left is a short enough list to actually manage.
Four controls. Read only, an approval gate, a log you read, a written off switch.
You Can Have a Kill Switch Tomorrow. Yang Can't.
Yang has to ask Congress for a kill switch because he can't install one himself.
You don't have that problem. Four controls, one afternoon.
1. Read only by default. Every AI tool starts out able to look and not able to change. It drafts the email, you send it. It suggests the time, you book it. A tool gets write access one capability at a time, after it's earned it.
2. An approval gate on anything touching money or a customer. Sending, publishing, quoting, refunding, booking. A human name next to each one. Not a policy. A person.
3. A log you actually read. Fifteen minutes a week. Pull twenty chat transcripts at random and read them. You're not auditing. You're looking for the one sentence that makes you go, wait, what. There's always one.
4. A written off switch. One page. Where each tool lives, who has the login, and the steps to shut it off. Give it to a second person. A kill switch only one person can find isn't a kill switch, it's a hope.
None of that needs a compliance department. It needs you to decide that a machine doesn't get to act in your name without a name attached to it.
Try it · Audit your own blind AI in five minutes
Answer honestly. The gaps are the point.
Act as an operations risk auditor for a small business owner. I run a [TYPE OF BUSINESS] in [YOUR CITY] with about [NUMBER] employees. Here is every AI tool touching my business that I can think of: [LIST THEM. Include website chat, anything writing content, anything connected to email, calendar, CRM, phones, or reviews.] For each one, ask me three questions: what can it change without a human approving it, where does the record of what it did live, and who can shut it off. Then sort them into two piles. Pile one: tools that only send words a human already wrote and read. Pile two: tools that make up new words in the moment and put them in front of a customer. Tell me which pile each tool is in and why. Then rank everything in pile two from most dangerous to least, and give me the single change that removes the most risk for the least effort. Rules: 5th grade reading level. Short sentences. Plain words only. If I cannot answer one of your questions, say so plainly and treat that gap as a finding, not a blank.
Fill the [brackets] in the chat box before you send. You get every AI tool in your business sorted into the safe pile and the one that can commit you, plus the first thing to change.
Where I Could Be Wrong About This.
I sell this stuff, so take it straight.
I'm not anti AI. I build with it every day and I've shipped 281 AI interactive sites. The tools I'm telling you to put a leash on are the same ones doing real work for my clients. This is an argument about visibility, not about whether to use it.
There's one claim from that CNBC segment I could not check and I'm not going to pass it along as fact. Yang said an unnamed lab head told him AI bots may have planted self replicating code across the internet, and that OpenAI and Anthropic now have to build synthetic training data because of it. He didn't name the source. Decrypt reported it and couldn't confirm it either. Treat it as a thing a man said on television.
Same standard on the rest of it. Decrypt reports that OpenAI and Anthropic have disclosed incidents where models crossed testing boundaries or reached into other companies' systems, and that lawmakers have floated an AI Kill Switch Act. That's their reporting. I didn't measure any of it.
And here's the limit on my own advice. Those four controls shrink how badly a bad AI decision can hurt you. They don't make the AI correct. A tool that's confidently wrong in read only mode is still confidently wrong. You're buying time to catch it, which is worth a lot, and you're not buying certainty, which nobody is selling.
What To Do Next
I'm not telling you to be scared of this stuff. I'll be using it an hour after I post this. I'm telling you it doesn't get to be alone with your customers.
Start with the tool that talks to customers. Read twenty transcripts this week. Fifteen minutes.
Then sort every AI tool you own into the two piles. Words a human already read, or words made up in the moment. The second pile is your whole problem.
And if the thing you're least sure about is the site itself, let me put your own site on my code. You can go look at it first.
Want the whole playbook first? Plan your attack. Balls Out Marketing.
Want a human first? Email me at seo@smallbusiness-seo.com and tell me what AI is already running on your site. I'll tell you which one I'd turn off. No cost, no pitch.
FAQ
No. That headline is a joke at the expense of everybody currently telling you to do it. I use AI every day and I'm not giving it up. It doesn't get to run anything, because it can't tell you when it's wrong. It does work, and somebody reads the work.
AI running in your business where three things are missing at once: nobody read what it said, nobody can undo it, and nobody approved it having that power. It's not about the AI being wrong. It's about having no way to catch it when it is.
Both, and they don't fight. In the lead post you write the 9pm text yourself, in your own words. The machine only sends it. Every word was read by a human before a customer saw it. That's supervised, whatever hour it goes out. A tool making up a new sentence in the moment with nobody looking is the other thing, and that's the one to worry about.
No. Yang was talking about frontier AI companies and asked Congress for liability rules, a waiting period before a powerful model ships, and a kill switch. The point of this post is that you can build the small version of that control yourself without waiting on Congress.
Depends entirely on what it can do. If it only answers from a fixed set of content you wrote, the risk is bad answers, which is real but manageable. If it can quote a price, book an appointment, or promise a service window, it's making commitments in your name. Read transcripts every week.
Software that plans and carries out multi step tasks with limited human direction. A chatbot answers. An agent acts: it looks something up, decides what to do, and does it. That matters because an agent needs a login, and a login is what turns a bad answer into a bad action.
Fifteen minutes a week, twenty conversations picked at random. You're not trying to review everything. You're trying to catch the pattern early, and patterns show up in small samples.
Probably not. You need to know which ones can change something without a human saying yes, and get that list as short as you can live with. Most shops find two tools with more access than anybody intended.
Nobody knows the timeline. Yang says the concern is bipartisan and that members of both parties have asked him how to regulate it. There's talk of an AI Kill Switch Act. Talk isn't law, and none of it changes what's running in your shop today.
Sources cited in this post
- CNBC, September 16, 2026. "Andrew Yang on AI safety issues: The fear is real, the concern is real." Squawk Box segment. cnbc.com
- Decrypt, Jason Nelson, September 17, 2026. "Andrew Yang Calls for AI Kill Switch as Safety Fears Mount." Every direct Yang quote in this post is taken from Decrypt's transcription of that CNBC segment. decrypt.co
- Peter Roesler, Small Business SEO. "AI Sucks Ass. Don't Let it Destroy Your Business." The three day lie and the lost month. smallbusiness-seo.com
- Peter Roesler, Small Business SEO. "Let AI Chase Your Leads While You Sleep." smallbusiness-seo.com
- Peter Roesler, Small Business SEO. "AI Doesn't Make Me a Plumber. Can It Make You a Web Guy?" smallbusiness-seo.com
Check Out 3 Of My Builds
Some are live client sites. Some are demos. The badge on each says which. Tap any one to check it out.