Last Monday, Mark Zuckerberg published a 6,000-word essay about superintelligence. Somewhere in the middle of it is a sentence I've now read maybe ten times. He says the outcome for jobs depends on which AI labs "lead" — if the labs building automation win, workers have a hard decade; if the labs building personal empowerment win, we all come out ahead.
I want to tell you why that sentence is wrong. Not wrong in a philosophical way. Wrong in a way you can check.
But first, you should know who's talking.
I run GoZupees. We build AI agents — voice agents that answer phones for fiber ISPs, intake systems for law firms, support automation for e-commerce brands. When a company decides to automate a function, I'm the guy on the vendor side of that call. I have been in the room, dozens of times, when a business decides what happens to work that people used to do.
So yes: I sell the thing this whole debate is about. Read everything I write here knowing that. I'd rather be the biased person who's been in the room than the confident person who never has. That's the deal with this newsletter. You get the bias declared and the room described honestly.
Here's what the room is actually like.
The words that never come up
A mid-market operator gets on a call with us. Say a regional ISP — a few hundred thousand subscribers, a support team of forty, fifty seats. The conversation is about numbers within the first ten minutes, every single time. What's your cost per contact. What's your abandonment rate after 6pm. How many calls are "where's my technician" — questions the customer could answer themselves if anyone had built the thing that answers it.
Those are the words that come up: containment, deflection, cost per resolution, after-hours coverage.
You know what words never come up? Empowerment. Augmentation. Capability expansion.
Not once. Not in a single deal I've done. And not because operators are villains — they're some of the most decent, practical people I deal with. It's because the business case that gets a budget signed is a cost case or a revenue case. "My team becomes more capable" is neither, until someone does the work of translating it into one. Most weeks, nobody does.
This is the thing Zuckerberg's essay skips entirely. Labs supply capability. Capability doesn't decide its own use. Demand decides. And demand has a spreadsheet, and the spreadsheet doesn't have an empowerment column.
Meta can build the most empowerment-flavored model in history. The day it ships, a thousand procurement teams will evaluate it on cost per contact. The lab doesn't get a vote on that. It never did.
Where the jobs actually go
Now the part that surprised me, and it took me a couple of years of deployments to see it clearly.
When people picture AI taking jobs, they picture the Amazon headlines. Sixteen thousand people, one announcement, one bad Tuesday. And those headlines are real — Challenger, Gray & Christmas counted over 112,000 US job cuts citing AI just through July of this year.
But that's not what I see in the mid-market, and the mid-market is most of the economy. What I see is quieter and honestly stranger.
Nobody gets fired.
The open req just... closes. Someone resigns and doesn't get backfilled. The night shift the ops director was going to hire never gets hired, because the agent covers 2am now. The seasonal ramp that used to mean six temp contracts becomes zero temp contracts. At one of our clients, the conversation after the pilot wasn't "who do we let go" — it was "we can finally stop trying to hire for the role we've failed to fill for eight months."
Attrition and non-hiring do the work. Silently.
And here's why that matters for how you read the news: Challenger counts announcements. There is no announcement for a job that was never posted. The Amazon and Oracle layoffs are the visible tip. The mass of the iceberg is thousands of mid-market companies each quietly not hiring one or two people, and no tracker on earth can see it.
Which, by the way, is exactly why the damage shows up where it does in the data — not in the unemployment rate, which is sitting near 4%, but in the hiring rate for 22-to-25-year-olds, which is falling hard in AI-exposed roles. The jobs aren't being destroyed. They're not being created. Different mechanism, same missing paycheck. I'll spend a whole issue on that one, because I think it's the biggest story in this entire subject and almost nobody is telling it right.
The part where I defend the other side
Now, honesty cuts both ways, and here's the honest part that complicates my own argument.
Empowerment is real. I've watched it happen.
At one deployment, the receptionist whose calls the agent now answers became the person who runs the agent — she reviews transcripts, catches the weird edge cases, decides what escalates. It is, no exaggeration, a better job than the one she had. More judgment, less repetition, and she's visibly prouder of it. I've watched an operations lead go from drowning in a queue to actually managing one.
So Zuckerberg isn't wrong that the empowerment future exists. I've stood in it.
But notice when it happens. It happens after the automation case closes. The ROI math gets the project approved; what happens to the freed-up human is a management decision made afterward, and it's made by the client, not by us, and definitely not by whichever lab trained the model.
That's the correction to the essay. Automation and empowerment aren't rival directions that competing labs push the world in. On the floor, they're sequential. Automation pays for the project. Empowerment is what a good leader chooses to do with the dividend. And plenty of leaders choose to just bank the dividend, and no model card in the world stops them.
The only question that matters
So if you strip everything else away, the whole jobs debate comes down to one question, and it isn't asked in San Francisco. It's asked in a conference room in Leeds or Tampa or Croydon, usually about six weeks after go-live:
What do we do with the freed hour?
There are only three answers. You bank it — margin goes up, headcount drifts down through attrition, and you've chosen the automation future. You reinvest it — the freed people go work the churn list, or fix the onboarding that's been broken for two years, or finally build the thing customers keep asking for — and you've chosen the empowerment future. Or you split it, which is what most companies actually do, without ever noticing they made a choice.
The future everyone's arguing about gets decided in that meeting. Thousands of times a week. By people who will never read a superintelligence essay.
I don't have a tidy answer for how you tilt those thousands of meetings toward door number two. I genuinely don't. What I have is a seat in a lot of those rooms, and a commitment to report what I see from them — including when what I see makes my own industry look bad.
That's what The Runbook is.
Here's what I'd ask you to do this week, if you run anything at all — a company, a team, a support desk. Find the last process you automated, or the next one you're about to. Ask yourself the freed-hour question before your CFO answers it for you by default. Then tell me what you decided. I read every reply.
Sandeep Bansal is the founder of GoZupees, which builds and deploys AI agents for mid-market companies — which means he profits from the thing this essay is about, and thinks you should know that.
Next issue: The strangest fact in this year's layoff data — the first people AI removed weren't the people doing the work.
Sources worth your time:
- Challenger, Gray & Christmas, July 2026 Job Cut Report — challengergray.com
- Mark Zuckerberg, "The Future is for Everyone" (Aug 10, 2026)
- TechCrunch running tracker of AI-cited layoffs — techcrunch.com
- Payroll research on 22–25-year-old employment in AI-exposed occupations, as reported by Yahoo Finance, July 2026 (verify primary study attribution before publish)