AI is still a tool, and companies keep cutting the people who can aim it
More than half of the employers who cut jobs for AI reasons now regret it. Forrester puts it at 55% in its 2026 workforce predictions and expects about half of those layoffs to get quietly reversed, with the work reappearing offshore or at lower pay. Roughly a third of the companies that reversed spent more restaffing than they saved. Nobody's model hallucinated a reorg. The technology performed about as well as anyone should have expected and the staffing decision built on top of it did not.
AI fails at the edges of the work rather than the middle. Harvard and BCG gave 758 consultants tasks inside what the model was good at, and the ones using AI finished 12% more of them about 25% faster and graded roughly 40% higher on quality. Then the researchers moved the tasks just outside that boundary, close enough to look identical, and the consultants using AI were about 19 percentage points more likely to be wrong. The output didn't come back empty. It came back confident and plausible and near enough to right that reasonable people signed off on it.
That's a critical thinking problem, not a knowledge problem. The model already knows more facts than anybody on the payroll and it will keep pulling ahead there. It can't read its own answer, see that it landed six inches off, and work out which part is still worth keeping. Somebody has to look at output that's close and decide whether close is useful here or dangerous here, then aim it somewhere better and run it again. Nobody sells that in a license.
The productivity gains don't go to the strongest people. Brynjolfsson studied 5,179 customer support agents and found novices improved about 34% while the most experienced gained almost nothing, and by some measures got slightly worse. Read fast, that's a case for cutting the expensive people, and I'd bet money it's been used that way in at least one deck. But that study measured support work where a correct answer existed and the system could route somebody to it. AI is excellent at that. It says nothing about the work where no answer is sitting in the corpus and somebody has to decide what to do with an approximation.
The cuts are landing on middle management. Gartner expects about a fifth of organizations to use AI to flatten structure and cut more than half of that layer through 2026. Korn Ferry surveyed 15,000 professionals and 41% said their company trimmed management layers last year. Oracle cut somewhere between 20,000 and 30,000 roles, weighted heavily toward middle managers. That layer's job is sorting what's actually going wrong from what only looks like it, which is the judgment the 19 point gap measures. Critical thinking doesn't track with title or salary band, so a reduction ranked by compensation can't find it on purpose.
The public explanation is performance. Companies say they're trimming low performers and raising the bar, which is safe to say because nobody outside can check it. I know a senior director who got recruited into a company and laid off inside a year. Both of those decisions can't be right. Either the search was wrong and somebody should be looking hard at the people who ran it, or performance is cover for a number that came from somewhere else.
The tool keeps getting bought as a replacement for people and keeps behaving like a tool. Those are two different purchases with different staffing attached. Nobody sensible is against the tool. Digging was done by hand until somebody built a machine that did it better, and that was obviously right, because breaking your back on a shovel is not the work anybody was trying to protect.
Once the machine does the digging, the reasoning goes that the crew can run it themselves, and the geologist and the civil engineer who were telling everybody where to dig start looking like overhead. They cost more than the crew and they never touch a shovel, so on a spreadsheet they're the obvious cut.
But the geologist was never paid to move dirt. The pay is for knowing where a hole is a bad idea, because there's a gas line under one spot and the soil in another won't hold what you're planning to put on top of it. That kind of knowledge only announces itself when somebody skips it. The machine makes digging cheaper and changes nothing about where the hole belongs, which is the part that ends up in front of a judge.
Cut the people who decide where it's a bad idea to put a hole and nothing has been saved. You've bought a faster machine and aimed it at the ground with nobody left to say wait. The companies that got there first are paying market rate to hire back the judgment they cut, which is an expensive way to learn that thinking was never the overhead.
Touring
Panic buttons often get inspected the way fire extinguishers get inspected. Somebody walks the building, confirms the device is there and powered, initials a tag, and moves on. That's a device check. All it proves is the button exists.
Activating one and following the alert to a person is a drill. HSEEP, the FEMA program most emergency managers build to, defines a drill as an operations based exercise validating a single function. A panic button test qualifies. It has to happen anyway, and an opportunity to train should never go to waste.
Run it on the paperwork you'd use for any operational exercise. You get a rep on the whole cycle, objectives through after action report, with nothing riding on the outcome. The officers get one too, on a real activation instead of a tabletop. Next time you build an exercise for a regulator or an insurer, the format isn't new.
The drill is twenty minutes. Tell the monitoring center it's a test, press one for real, and follow the alert until a person answers out loud. Quarterly covers it. Log how long that took, and whether the alert told them where it came from.
The button is almost never the problem. It's the account that stopped getting paid when a contract changed hands, or the queue labeled with a name nobody at the console recognizes, or the alarm nobody was ever told the meaning of. Those are records problems, and each one becomes a finding with a name and a date.
Signals
Four integrators changed hands in five days
Between August 3 and August 7, Everon bought Scarsdale Security Systems, a New York commercial security and fire integrator founded in 1982 with about 100 employees, and said outright the point was feeding its National Account program. SAGE Integration picked up Vital Installs in Columbus and sudoVision Consulting in Charlotte, an SDVOSB certified firm working government accounts. RapidFire took Comsec Associates in Flintridge, California. And Resideo finished spinning off ADI Global Distribution, which reshuffles the parts channel underneath all of them.
Four in five days is a market moving, and SDM's Top Systems Integrators report from last month shows which direction. The hundred largest integrators booked $9.72 billion in North American systems integration revenue for 2025, up 15 percent, on 263,333 new systems started against 300,667 the year before. More revenue on fewer jobs, with fewer people, out of fewer locations.
Where the big work went shows up in the comments. Data centers come up more than any other vertical, and the executive calling it the most dynamic and fast expanding segment in the industry is Don Young, CEO of Everon, which sits third on that list and just bought Scarsdale. One hyperscale campus is worth what a hundred mid-size buildings are worth. Tariffs are in there too, but the integrators describe those as hitting margins between proposal and install rather than invoices.
A national account program is a standardization play, and it's worth being honest about who it's standardizing for. The acquirer buys regional firms so a multi-site customer can sign one contract, one invoice, one service level, one preferred product line, which is worth real money if you're running forty locations across twelve states. If you're one building on the acquired book, the average job on that list got materially bigger in a year while your camera refresh stayed the size it always was. What you get is a dispatch pool instead of the technician who knows which door is on which panel.
Nobody's doing anything wrong here, and you won't see it in the quarter it happens, because the contract keeps getting honored and the invoices keep arriving on time. You see it a year later when something breaks and the person who shows up has never been in the building.
Passdown
dictated by Ricky Portezzo, Senior Security Supervisor in Center City, Philadelphia
Nothing on the log worth waking anybody up over.
Mikey sent down a list, we now press every panic button for real once a quarter and write down what happens. My first reaction, and this one's on me, was that this is somebody at a desk inventing work for people who are already busy. I check those buttons every month. They work. I initialed the tag on every one of them personally.
So I go do the bank branch figuring I'm in and out. Press it. Call comes back and the operator reads me a suite number for a travel agency. There hasn't been a travel agency in that building since before I started on this loop, and I used to get a Tab every day at lunch.
Three weeks of emails to get that fixed. Three weeks! I had to find out who owns the account, and the answer was nobody for about a year and a half.
Then Tuesday the manager over there hits it for real, guy in the vestibule going at the ATM with something, and it came through clean with the right address on it and Joey was in the lot before she finished telling me about it.
So Mikey was right but don't tell him I said that. The buttons were fine. I was testing the wrong end of the thing the whole time.
SCC Spectrum Security Almanacs. www.gsoc911.com/products