Everyone Bought the Computers. Six Sectors Got the Gains. Will AI Be Different?
The computer revolution was in the 1990s. The AI transformation is playing out now, and hardly anyone has kept evidence showing which side they are on.
TL;DR: Your board will ask what the AI spending bought, and most companies can no longer answer, because nobody measured the work the deployments were about to change. The lag in returns is real, and it sorts the buyers from the beneficiaries. Six sectors out of sixty took 99% of the 1990s gain, and the heaviest spenders missed it. You cannot locate yourself in a distribution without a “before”, and almost nobody took one. Engineering and support kept timestamps, so those baselines can still be recovered.
Your board is on your back, asking what the AI money bought. The answer went missing in 2023, when the first deployments went in and nobody wrote down how the work ran.
The comfort on offer is that returns lag, and it is well founded. Robert Solow wrote, in the New York Times Book Review in July 1987: “You can see the computer age everywhere but in the productivity statistics.” The statistics caught up about a decade later. A general-purpose technology pays off only once processes are redesigned around it and the people retrained, and the accounts book all of that as cost until it pays. Re-tooling alone changes the tool and leaves the process where it was. The average company that read Solow as a promise spent the decade buying computers and worked out later what to do with them.
The catch-up was narrower than almost anyone remembers. Six sectors out of sixty produced 99% of the gains. The sectors that bought the bulk of the new computing grew productivity 0.3% a year. Retail was one of the six, and inside it the acceleration happened because Wal-Mart forced it: rivals copied the big-box format and the barcode scanning after a decade of losing share. So the sorting ran between companies inside one industry, as well as between industries. No competitor announces that it has finished reorganizing. The market share moves first, and that is the notice.
So the lag is real, and it sorts the buyers from the beneficiaries. Productivity is rising while the share of executives who credit AI with 5% or more of their company’s earnings, and call the effect significant, holds at about 6%. An average can rise while the group capturing the gain stays the same size.
Which side are you on? The answer changes your next two years, and you will have a hard time getting it. AI went in at scale from 2023 onward, and I still haven’t seen a company measure the work it was about to change. Not one. The freed hours get eaten by approval steps and handoffs that are still sitting there, and the gain never reaches the P&L. The most expensive thing companies did with AI wasn’t the spend. It was deploying without a “before”.
Engineering and support are the exceptions. Count resolved tickets per support person, and lead time from commit to production in engineering, with median time to resolution beside the ticket count. The tracking systems were already stamping times when the first models arrived, so you can recover those baselines instead of starting from scratch. Everywhere else, the record is gone. Start the count now; in 2028 you will be arguing from data instead of from memory.
One objection I can already hear loud and clear. Baselines on knowledge work are gameable, and the two I just named are as gameable as any: cycle time moves for six reasons a quarter, and none of them is the model you deployed. Good leaders have judged a process change by whether the work got visibly better, for a century, and it has served them. It never once told them whether they were gaining on the field or losing to it.
Solow’s paradox resolved itself in the end. The sectors that had paid for the computers watched the gains land somewhere else. Will you re-tool or re-think your processes with AI? Only one of them earns anything back.
Sources
Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies.” American Economic Journal: Macroeconomics 13, no. 1 (January 2021): 333-72. https://www.aeaweb.org/articles?id=10.1257/mac.20180386
McKinsey Global Institute. US Productivity Growth 1995-2000. October 2001. https://www.mckinsey.com/~/media/McKinsey/Featured%20Insights/Americas/US%20productivity%20growth%201995%202000/usprod.pdf Accessed 2026-09-08
Solow, Robert M. “We’d Better Watch Out.” New York Times Book Review, 12 July 1987, 36.
Tinkoff, Dan, Lieven Van der Veken, and Michael Chui. “The state of AI in 2026: On the road to ROI.” McKinsey Global Survey, 25 August 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai Accessed 2026-09-08
US Bureau of Labor Statistics. Productivity and Costs, Second Quarter 2026, Revised. Released 3 September 2026. https://www.bls.gov/news.release/archives/prod2_09032026.htm Accessed 2026-09-08
Originally published at www.orionplaybook.com.
The Oldest Skill in Your AI Budget
Briefing a person and briefing an agent are the same act. Most people were never taught to do either, and the second one shows up in minutes.
TL;DR: Mid-market companies now put more investment into AI than into anything else, and most organizations credit it with under 5% of EBIT. What sits between the spend and the result is an ordinary management skill: stating the outcome, supplying the context, defining what done looks like. Only one in three workers has received any employer-provided AI training in the past six months. The existing training covers prompting and stops where the work gets useful. Nobody in your company has read anybody else’s prompt, so nothing compounds. Ask three people for the prompt each used on one task.
Continue reading “The Oldest Skill in Your AI Budget”Authority Delegates. Accountability Doesn’t.
Hand work to a person, and part of the risk goes with them. Hand it to an agent, and none of it does. The fix everyone reaches for targets the wrong variable.
TL;DR: A human delegate absorbs part of your risk because they have assets, a career, and legal standing of their own on the line. An agent has none of that, so authorizing one concentrates your exposure at the moment it feels like you are sharing it. Liability attaches regardless of how closely anyone was watching the work. People correct AI errors less when correcting costs effort, and money does not help. Skipping every approval prompt is itself the decision, taken once, by one person. Write down what it can reach and what it could destroy before you authorize the next one.
Continue reading “Authority Delegates. Accountability Doesn’t.”Directionally Correct, Specifically Utterly Wrong
The errors in a generated research report are not spread evenly. They cluster where the decision is, and the rule you already use to catch them is aimed at the wrong variable.
TL;DR: A generated research report is right only if every link in it holds, so accuracy falls away as the steps pile up. In one benchmark’s public filings, the same models score 83% on answering what a document says and 35% on answering what follows from it. Verifying the recent and the obscure misses it; decisions rest on public material. The mode that most resembles diligence fabricates the most citations. Mark every claim retrieved or composed, and re-source only the composed ones before money moves.
Continue reading “Directionally Correct, Specifically Utterly Wrong”An Ode to Electronic Communication (Now That Words Are Free)
Writing got free, reading did not, which makes every message you send a withdrawal from somebody else’s day.
TL;DR: AI made writing free and left reading as expensive as it always was, so every inflated message moves work from the person sending it to the person receiving it. The meter is bolted to the side that got cheaper, and nothing counts what lands on the other. The padding a model adds is billed twice, once in tokens and once in attention. Five daily AI users voting together now spot machine prose 299 times out of 300. Change what you ask the tool for: cut the draft in half every time. Send nothing that came back on the first prompt.
Continue reading “An Ode to Electronic Communication (Now That Words Are Free)”The Diagnostic You Already Ran
How much of your work an agent can take depends on how completely you can specify it, and you are paying for the gap on every single run.
TL;DR: Everyone has the same models, and nobody has the same space around them, which is where the real variable sits. The framework for measuring that gap ran in a management journal back in 2005, written for an entirely different purpose. Two questions decide it: can you codify the work, can you measure the result. What ruled work out then, untransferable team knowledge, is exactly what you can now supply. A colleague is onboarded once; an agent without memory is onboarded every time. Specify once, where the agent reads by default; unclear thinking now runs a meter.
Continue reading “The Diagnostic You Already Ran”An Ode to (Valuable) Meetings
Almost everything that ruins a meeting is decided before anyone walks into the room.
TL;DR: In January 2023, Shopify deleted every recurring meeting with more than two people, and the instructive part was watching which ones people fought to get back. A meeting is a symptom: the damage was done days earlier, by someone dragging a block onto a calendar. Overstuffed invite lists are what an organization that writes nothing down looks like. If your AI avatar could attend without loss, that is a document with catering. The undelivered action item costs more than the silent attendee; it books the next meeting. Judge your AI by the meetings it prevents, then cancel one and watch.
Continue reading “An Ode to (Valuable) Meetings”This Time, AI Doesn’t Have an Alibi
The work AI is absorbing now is the work we offshored twenty years ago, for the same reason. The advantage won’t go to whoever swaps a human for a machine. It goes to whoever redesigns the work.
AI had an alibi for the vanishing entry-level job until 2025. It won’t have one for what comes next. The work most exposed to it in the short term is the one we spent the 2000s offshoring, from call centers to entry-level coding to the back office, because both waves seek the same thing: how much of a job is written down and how little of it is judgment. That work already left once. The mistake now would be to automate it in place, swapping a machine for a person and changing nothing else. The larger prize, including for the economies that first took this work, belongs to whoever is willing to rethink how it gets done at all.
Continue reading “This Time, AI Doesn’t Have an Alibi”The SaaS Bill Just Split Into Two Meters
