About twenty years ago, when the TV program “Mad Men” aired for the first time, I was enamored. I had always found previous eras more interesting than my own, and I had the fedoras and spats to prove it. To put it mildly, I was a unique young man. To suggest that much has changed would be disingenuous. I still wear a straw summertime fedora, and my kids know the difference between Audrey and Katherine.
Anyway, back to Mad Men. Before the show was produced, I’d always been curious about what people did for work before my time. I have had a computer since 1988, to be exact, so my entire academic and professional career has been connected. I still own a typewriter (see above), but I’ve been marketing and selling electronically since the dawn of electronic go-to-market ubiquity.
Watching Mad Men didn’t disabuse me of my hypothesis: before technology, lots of people just sat around, and what got done was done by pools of typists, creative folks, laborers, secretaries, and more. And everything that got done took time. Lots of it.
The Slow Start
Let’s consider the changes in operational efficiency and productivity between the 1960s and today. It was night and day, monumental, transformative. But it happened in stages.
After Don Draper’s story ended, we entered a period that can easily be considered one of unfulfilled promise. That 1988 computer I mentioned was incredible, but it was a word processor, not a connected device. The phone in my pocket today is 10,000 times more powerful and useful than the Mac Plus that sat in my bedroom, and it barely resembles it. Computers replaced typewriters throughout this stretch, but the nature of the work stayed the same. People composed and edited, but still printed everything for consumption.
It wasn’t until the mid-1990s that connectivity began in earnest. Gradually, over the course of the 1990s and early 2000s, the world became more connected, and the promise of technology gained the infrastructure to make good on itself, letting people work in ways they hadn’t in the 1960s, the 1950s, or before. That’s less than two decades of progress, and yet by almost any measure the impact has been incredible.
The IT sector was less than one percent of US GDP in 1980, and by 2015 it had grown to 5.2 percent. Employment in that same sector barely moved the whole time. Same number of people running the machines, six times the output twenty-five years later. That gap between headcount and output is the unfulfilled promise, cashing out. Economists even gave the earlier stall a name: the Solow Paradox, after Robert Solow’s crack that you could see the computer age everywhere except in the productivity statistics. Fifteen years of buying machines before the machines earned their keep.
The Payoff
Once connectivity arrived, the payoff arrived fast. Zoom into the peak of the dot-com run and the numbers get almost absurd: IT industries made up 8 to 9 percent of the US economy between 1996 and 2000, and they produced 1.4 points of the 4.6 percent GDP growth over that same stretch. Less than a tenth of the economy generated more than a third of the growth. E-commerce tells the same story from the consumer side. It was 3.86 percent of retail sales in 2000. By the end of 2025 it had crossed 18 percent. The catalog didn’t disappear, it moved onto a server, and the server got faster every year for three decades straight.
The Part Nobody Talks About
Here’s the part of the story that doesn’t get told enough, and it’s the part I care about most.
None of that growth held together on its own. Somewhere between the Mac Plus in my bedroom and the phone in my pocket, an entire discipline got built to keep watching the machine while the machine did the work. Network operations centers. Application performance monitoring. Site reliability engineering, a job title that did not exist when my grandparents bought me that first computer, now sitting on every tech org chart in the country. Dashboards, alerts, on-call rotations, postmortems. Nobody voted on this. It got built because the alternative, systems failing in the dark with nobody watching, was not survivable at scale. We learned, the hard way, that you cannot run infrastructure you cannot see. Observability was not a product category before it was a necessity. It became a product category because it was one.
That’s the thesis. Fifty-six years of computing history, and the throughline is not the chips or the bandwidth. It’s that every leap in capability got paired, eventually, with an equal investment in watching that capability. The eyes came with the growth. They had to.
Right Now
Which brings me, with the timing only a blog post can manufacture, to right now.
I won’t pretend I don’t see where this is going, and I won’t insult you by dressing it up as a coincidence. I run an observability company. I have a stake in this argument. Read the rest with that in your pocket. I still think it’s true, and here’s why.
What we’re watching happen with AI is the Mac Plus to smartphone transition, except fifty years got compressed into about three, and almost nobody built the watching function first this time, either. Everyone is racing on capability and skipping the part where you earn the right to trust it. The headlines from just this week make the case better than I can.
On September 3rd, ChatGPT, Claude, and Grok all went down within the same ninety-minute window. Three companies, three separate engineering organizations, three competing products, and they all buckled at once because a regional failure at one shared cloud provider rippled through infrastructure none of them fully controlled. That’s not a bug in one company’s code. That’s an industry that scaled capability faster than it built resilience into the plumbing underneath it.
Five days later, CIO ran a piece on Article 73 of the EU AI Act, and it should worry anyone running AI in production. The old breach clocks, the ones every security team has memorized, start when you know data got accessed. Article 73 starts when you suspect an AI system caused harm downstream, even if nothing was ever hacked. The example in the piece: a flawed AI matching system at a benefits agency misflags a run of different people over several weeks, and nobody notices the pattern until people stop getting paid. No intrusion. No alert. Just an AI system doing exactly what it was told, wrong, in the dark, until the damage surfaced somewhere else entirely. The clock on that kind of incident is 15 days by default, and most companies do not yet have anyone assigned to notice it’s running.
Then there’s Unit 42’s writeup on an actual breach. A threat actor turned a set of AI agents loose on an enterprise network, and the agents did in ten hours what would normally take a human red team two weeks: reconnaissance, credential harvesting, privilege escalation, more than fifty distinct attack techniques, executed and adapted in real time with almost no human in the loop. The attacker left behind an eighty-page report on the way out, documenting every hole it found. Palo Alto’s takeaway wasn’t “patch faster.” It was that the old model of a human analyst investigating each step no longer works once the thing on the other side moves at machine speed.
The aggregate numbers back it up. One in four malicious breaches over the past year involved AI, up 56 percent from the year before, according to IBM’s latest breach report. Those AI-involved breaches cost companies about a million dollars more on average than a typical one. Ninety-two percent of the companies hit that way had no real AI access controls in place at all.
What We Can’t Afford to Skip
Put those four stories next to each other and you get a clean picture. The capability got built. The watching function did not keep pace. We are living through the Solow Paradox again, except this time it isn’t a twenty-five-year lag buried in a productivity chart. It’s outages and breaches and regulatory deadlines showing up in real time, because the systems in question move too fast to wait a quarter century for someone to notice the gap.
The 1970s and 80s taught us the tools alone were never the story. The fifty years since taught us the watching was never optional, just slow to arrive. This time, slow to arrive might be the one thing we can’t afford.