Only 3,462 of 52,881 August Job Cuts Blamed AI. So Why the Panic?

The internet can make it sound as if every empty desk has a chatbot sitting in it. The August 2026 numbers tell a more useful story. Of 52,881 announced U.S. job cuts that month, employers attributed 3,462 to artificial intelligence. That is about 6.5 percent, putting AI fourth among the cited reasons for August.

That does not mean AI is harmless. It means the headline needs a denominator. This episode follows the numbers back to the primary source, separates an employer's label from proof of causation, and turns the story into a practical operating plan for small businesses in Santa Clarita, Greater Los Angeles, and anywhere an owner is being pushed to buy an AI agent before fixing the basics.

TL;DR

AI accounted for 3,462 of 52,881 announced cuts in August. Restructuring accounted for 16,173, market and economic conditions for 15,260, and closings for 6,743. Year to date, AI was cited in 116,175 of 529,914 announced cuts, so the longer trend still matters. The useful conclusion is neither "AI is harmless" nor "AI took every job." Ask what work changed, who made the decision, and what evidence supports the label.

What the layoff number actually says

A press release is evidence of what a company said. It is not automatic proof of why every decision happened. At least three different stories can hide inside an AI layoff announcement:

1. The machine may now perform the work directly.
2. The company may already need to restructure and use AI as the most respectable label for the change.
3. Leadership may cut payroll to pay for models, data, chips, and implementation.

Different companies can be doing different things at the same time. That is why viral totals need their time period, denominator, definition, and source attached.

The August number was 3,462. The year-to-date number was 116,175. Both are real, but they answer different questions. One describes a single month. The other describes the first eight months of the year. Blending them produces heat instead of insight.

The threat report needs both eyes open

The episode also examines Anthropic's September 2026 threat intelligence report. It describes, among other activity, a commercial influence operation involving roughly 70 fabricated news sites and more than 250 inauthentic commenting accounts across six continents.

That is a serious misuse story. It is also a vendor describing defensive work involving its own product. Both facts can be true. Read the original report, note the company's incentives, and distinguish documented activity from a dramatic prediction about what happens next.

Most small businesses need a switch before an agent

Consider a six-truck heating and air company receiving about 40 calls a week. Nine calls go to voicemail around lunch. Website forms are checked twice a day. Google messages sit unread. The owner thinks the problem is lead volume because that is what the advertising industry keeps selling him.

The proposed cure may be a $2,000-per-month autonomous agent. The first repairs require almost no intelligence: send an immediate text, route the lead, and log in.

A switch performs one action. Every new lead gets a text inside 60 seconds.

A lever applies a condition. An after-hours form goes to one route while a yard-sign lead goes to another.

An agent receives a goal and chooses the steps. That can be valuable where real judgment is required, but only after the owner defines the philosophy, positioning, permissions, and lines the system may never cross.

Buying the third when the business needs the first is how owners spend more and solve less.

The 15-minute privacy check

Open the settings of the AI product used for work. Find its current retention and training policy. If the account offers a training control, set it intentionally rather than assuming a default.

Then write down three categories that never go into an unapproved public AI tool:

• Client financial information
• Social Security information
• Anything a competitor would pay to obtain

Local models can be appropriate for highly sensitive workflows, but local does not mean free or effortless. Hardware costs money, draws electricity, needs updates, and eventually stops answering at 11 PM on a Thursday. The job and the sensitivity of the data should justify that responsibility.

Questions people are asking

Is AI causing layoffs?

Yes, in some companies and roles. The August report attributed 3,462 announced cuts to AI. That is material, but it is not all 52,881.

Why do AI layoff numbers conflict?

Reports can use different windows, definitions, sources, and employer labels. Always inspect the denominator and the original report.

Does every small business need an AI agent?

No. Start with the leak. Fix switches and conditional rules first. Use an agent only where judgment is required and after its authority is written down.

Should private business data go into public AI tools?

Retention and training policies are operating controls, not fine print. Know what the vendor retains, disable training where appropriate, and define information that never leaves an approved system.

What to do next

Walk one customer journey this weekend. Ask four questions: How does the lead come in? What happens in the first 10 minutes? Who touches it? Where does it die?

That small audit may find more money than a complicated agent demonstration.

Continue at https://santaclaritaartificialintelligence.com/, https://honorelevate.com/, and https://hireaivoice.com/. AI for everyone, not just the wealthy.

Sources: https://www.challengergray.com/blog/august-2026-job-cuts-fall-47-from-july-down-14-from-august-2025/ and https://www.anthropic.com/news/detecting-and-countering-misuse-september-2026.

Full episode transcript and field notes

Somebody in your life read a news article this week that was written by a machine on a website that doesn't employ a single human being. And you know they believed it. And I know you got those people like I got those people. They believe it. And then they told you about it at work.

And there were about 70 of those websites inside of one operation. 70 fabricated news sites plus more than 250 inauthentic commenting accounts underneath them running across six continents, sold as a service to whomever paid the bill. And that's not a prediction. It's not a movie. It's mass disinformation used to create unease across every platform, every societal structure, everywhere at the same time, paid by people who apparently will remain nameless, at least for now.

And it was in a report published yesterday by the company whose own model was being used to do it. You see these large language models. They're really powerful. They are not only able to potentially, as long as the fail-safes are there, they can't, but they're capable, if left unlocked, to seek out biological destruction. They could do that.

Well, that's not the only thing they could do. They can master a mega media campaign and put, if you have the money, put 10,000 agents with all PhD-level educations in misinformation and societal manipulation, and, you know, we're easy. I like the little cats that dance around that really upbeat music. I like that. That's easy.

Okay, I want you to hold a number for me, 3,462. I'm going to come back to it in about nine minutes. And when I do, I think some weight off your chest. I think it'll take a little bit. Now, you've been carrying around all year this weight on your chest.

You haven't been asked to carry it around, so I'll get to it. Good morning, good afternoon, good evening, of course, whenever you happen to be tuning in. This is AI with Honor, the daily download. It lives at Santa Clarita Artificial Intelligence, the practical AI porter for Santa Clarita business owners. Everything in this read came off a primary source.

So if a company says something about its own product, I'm going to tell you it's a company claim. If a number came from the government, I'm going to tell you that too. There are different kinds of things, and the difference is most of the argument. And it's all very self-interest driven. I mean, I don't know how many folks are out there without trying to drive some kind of an agenda.

I don't monetize this. There's no way. I mean, three followers, one watch. I think it's me reviewing it when I'm done. So don't like it.

Don't subscribe. Don't do anything. I don't want to get caught up in all that mess. But please comment if that helps. I'm not trying to help the algorithm, but I would love to at least get a conversation going with you.

But you're not required to do anything. I didn't talk to my employees and they keep telling me something about my particular way. Most of you got unsubscribed. Yeah, nothing like that. Just I hope it helps.

I hope it helps. And don't believe anything. You got to really look into it. You got to do a little bit more heavy lifting these days. and the system you go to look it up, it might be lying like the original source. All right, let's get back to this.

Anthropic published its threat intelligence report yesterday, September 10th, that covers what their team found. We're 2026, by the way. Just, God, I hate that. Don't you? You're watching that video and you think, God, this sounds, kind of sounds familiar.

And you look, it's a year ago and they just republicized it or published it. Yeah, that's also sucks. Okay, it covers what their team found and shut down between December 2025 and August 2026 across seven categories. Cyber operations, surveillance, influence, sorry about that, influence operations, I forgot I had it, conventional weapons, biological misuse, scams and fraud, and distillation, like the water, but different, which is stealing a model by asking it a very large number of questions and training your own thing on the answers. Distillation, yeah, I guess they all kind of do it. what's the word?

Hasn't been proven. I forgot what the word is real quick. All right. The influence operation I opened with is one case of many out there. Roughly 70 fabricated news websites, more than 250 inauthentic commenting accounts run as a paid commercial service across six continents.

Somebody sent an invoice for that. Somebody paid it. Label this one carefully with me. This is a company reporting on misuse of its own product, and that's not a reason to throw it out. It is a reason to notice that they hold two incentives at once, though, to show you that they're catching things.

Good job. And to show you that the thing is powerful. Good job. We all like gangster. Come on, let's get over this.

Got to be careful, though. Now, both can be true, can be true, while the findings are also true. The line in that report that actually matters is not the fake news count. It is this one. Anthropic says the models roll in cyber operations has become, in their words, increasingly autonomous with multi-agent frameworks running reconnaissance, exploitation, and data exfiltration, and humans staying in the loop mainly to set the targets.

Read that slowly. Humans mainly in loop to set the targets because it's a description of a workflow, not a monster. A person picks up a goal. The software surveys the ground, writes its own script, runs them, reads what came back, and goes again until the job has completed. Now here's uncomfortable part and it's uncomfortable for the sales pitch, not for you.

This is the identical loop being sold to your plumbing company as an agent. Same architecture, same autonomy. The only difference in the whole picture is what the human pointed at. Same day, September 10th, OpenAI shipped two things. The first is an agent's API with the company says, let's develop or build agents reliable enough to put into production.

This is a company claim and the company's own words and reliability is precisely what has been missing from every agent demo for two years. It's what you watch for now, somebody outside the company measuring it. Now, the second is more interesting to a business owner. ChatGPT for financial services, it pairs their GPT-6 Astra model, which is, yeah, got to say, I wasn't a big fan at first, but yeah, I get it. Astra model with the licensed financial data, but with Morgan Stanley and Evercore's design partners using data sets from Dalupa, plus PitchBook, LSEG News, and Crunchbase.

OpenAI hosts and indexes that data itself so the thing can cite a specific table on a specific page instead of telling you a beautiful story. And read the boring part because the boring part is the actual product in this case. Business data is not used to train models by default. It's encrypted at rest and in transit. Administrators configure retention.

Compliance teams can export audit logs, look at the shape of that thing, because the shape is the whole lesson, and the shape is free. They did not build something that knows everything. They picked one narrow job, investment, banking, and equity research. They licensed real data instead of hoping the model remembered it correctly. They made it cite the exact table it pulled from, and they put retention and access controls in the box while they have sold it to anybody before they did that.

Narrow job, real data, citations, controls. That's what development looks like. And it would be awesome if we could take these instead of trying to jump into this general intelligence, which might be okay, but maybe needs to be a little bit safer. We just poke these things. We focus them right on all the cancer research, all of the disease, all of the longevity problems, all of the aging issues, every kind of sickness out there and world peace and oh my god but no no so narrow job real data citations controls that's what deployment looks like now everything else you've been shown this year is a demo you don't need morgan stanley budget to copy that list you need to pick one job the list is the expensive part and somebody just published it for free 3 462 that's how many announced job cuts in the United States in the month of August were attributed to artificial intelligence.

According to Challenger, Gray and Christmas, the firm that has been counting announced layoffs in this country since the 1990s, 3,462 in a month that had 52,881 announced cuts in total. It's the lowest monthly AI figure since December of 2025. My question to all of those people that were laid off, if it really is true that they were laid off because of AI, isn't AI coming into the business, isn't it going to make that person's position better? And isn't it going to make more money for the company than that person was capable of? Now, if the human being is the least able person on the team, why don't they just keep the human?

AI is working. AI is doing all this fancy stuff. Just keep the human. AI is going to double, triple their production at least, that's what it would be. Humans, eight hours.

AI, you got a couple more of those shifts. 24 hours a day, seven days a week, no days off, no vacation, no sick time, no nothing. And it's there and it's working. Let's keep them, folks. Come on, talk to somebody.

We shouldn't be getting rid of people for AI because AI is, that's the savior, right? That's the one that fixes everything. So just keep the employee. No, they don't have to keep their hands in the cookie jar. It's okay. you know, have them do something else.

Have them, I don't know, wash the car. I don't know what the other thing would be. You want me to figure it all out? No, but seriously, though, keep them there. Maybe there's something they can do with AI without creating the biases with it.

Maybe they can do something about slowing it down or interfering. Maybe there's some additional attribute that the human being can do. There's a lot of things. I'm sure it would be awesome. But again, so we got that unemployment.

Let me get back to it. And it ends a five-month run. This was the lowest monthly AI figure since December 2025, when it was 142, and it ends a five-month run starting back in March, where artificial intelligence was the single most cited reason every month. In August, it fell all the way to fourth. Here's what beat it.

Restructuring beat it. 16,173 cuts. Market and economic conditions, 15,260. Closings, 6,743. Then artificial intelligence at 3,462.

3,462 out of 52,881 is about six and a half percent. Now here's why I wanted you to hold that number. Search the words AI layoffs and you will find much larger claims. Go to the firm actually doing the counting and inspect the year-to-date figures: 116,175 out of 529,914. That's about 22%. And it's still the number one cited reason for this year.

So this is genuinely not nothing, but the year-to-date share is about 22 percent. And last month it was six and a half. Now I'm going to call this a respectable word. When a company has to cut people, artificial intelligence is the most respectable word available to it. It says we're advancing.

Restructuring says the committee got it wrong. Market conditions said we misread the year. AI says we are the future. So in one of those memos lands, run the three possibilities before you panic. The machine genuinely does the work now.

Well, they say it does. And if they're replacing human beings with artificial intelligence, yeah, it's got to be at least as good as the human, at least as good as the human. so keep the human. AI, well, it's free. I mean, it should be free, right? I mean, that's where this is.

And if it's not free, the extra that it's going to be making because it is super intelligent, so awesome and magical, hell, keep the human. I don't see why we should get rid of them. Or the company needed to cut anyway and AI is the polite label on the door. Or they're cutting people to pay for the AI because the hardware is not free. Well, it's going to be much more productive.

It gets a two extra shifts. You can at least keep the human and pay for the AI because it gets to work 24 hours, not eight. And every day, all three are happening in real companies right now. Only one of them is the story you get told and take the rest of this counterweight because I'm sitting here. I don't want you to hand it to you in a half picture.

Announced cuts in the first eight months of 2026 total 529,914. That's down 41% against the same eight months of 2025, which ran 892,362. Announced hiring plans set at 119,825 for the year. That's up 37% over last year. And Andy Challenger's own caution on that.

In his words, while companies are making plans to hire more workers than last year, it doesn't appear those positions are being filled quickly. So put the three together. Fewer cuts than last year, move more hiring announced than last year and the jobs are slow to actually get filled. That is a strange, sluggish, frustrating labor market and it's not a collapsing one and it does not fit either sales pitch that you're being handed. Which brings me to the thing I say every day and will keep saying until it stops being true.

Two groups are selling you fear and they're selling it in opposite directions.

Common questions

Connor T. MacIvor · CalDRE #01238257 · Sync Brokerage, Inc. · DRE #02031490