The Tool Won't Decide What Your Company Becomes. The Mindset Holding It Will.

In the 1860s, an economist named William Stanley Jevons noticed something strange about steam engines. As engines became more efficient, England didn't burn less coal. It burned far more. Efficiency made steam power cheap enough to put in places it had never made sense before, and the number of profitable uses exploded. Economists still call it the Jevons Paradox: when something powerful gets cheaper, demand for it doesn't shrink. The market for what it can do grows.

Thinking is now that something.

The cost of a first draft, a market analysis, a modeled schedule, a researched answer has collapsed. And every company deciding what to do about it is standing where the mill owners stood: holding a dramatically more efficient engine, choosing between two readings of what it's for.

Spend less on coal.

Open three more seams.

One reading says: spend less on coal. Run the same operation with fewer inputs. Cut.

The other says: open three more seams. All the work that was never economical before, the markets never entered, the clients never pursued, the questions never asked because asking was too expensive, just became affordable.

Same engine. Two entirely different companies ten years later.

Same engine. Two entirely different companies ten years later.

Who holds the engine matters more than the engine

Here is our observation from inside mid-market AEC, energy, manufacturing, and industrial companies: the reading a company chooses is rarely a strategy decision. It's an accident of org chart. Hand AI to the functions built to prevent bad outcomes, and it gets deployed to prevent cost. Those departments are doing their jobs. They are rewarded for savings they can document and punished for risks that go wrong, so they will always read a new engine as a way to burn less coal.

The entrepreneurial parts of a business read the same engine differently, because they're paid on upside. A business developer sees pursuits the firm used to walk past. An operator sees capacity that used to require a hiring cycle. An owner sees the market share of slower competitors.

The behavioral scientist Rory Sutherland jokes that there are two groups of people he'd rather not see get their hands on AI first, and one of them is the finance department. The serious point underneath: family-owned and founder-led companies, the exact profile of the American mid-market, are structurally positioned to win this transition, precisely because they can choose the opportunity reading without asking a committee. Most of their larger competitors can't.

The obstacle was never the technology

When adoption stalls, companies almost always diagnose a technical problem. Our data isn't ready. Security isn't settled. The tools aren't mature. Real issues, all solvable, and almost never the actual blocker.

Apple didn't win by building faster computers. It won, in part, by making a computer you could put in a room without turning the room into an office. American Express once discovered that people weren't declining to apply for the Gold Card because they lacked reasons to want it. They were afraid of being turned down. One reassurance multiplied applications several times over. In both cases the breakthrough wasn't engineering. It was locating the psychological obstacle everyone else had walked past.

AI adoption in industrial companies has the same shape. The stated obstacle is data readiness. The real obstacle is a project manager who doesn't want to be the name attached to the failed pilot, a senior estimator quietly wondering if the tool makes thirty years of judgment worth less, a leadership team with no cheap way to be wrong. No software purchase fixes those. Design does: experiments sized so failure costs nothing but the lesson, deployment that makes veteran judgment more valuable rather than less, and wins sequenced so confidence compounds.

Jeff Bezos calls decisions you can walk back two-way doors. AI has made almost every experiment a two-way door. The companies pulling ahead aren't the boldest. They're the ones who noticed the doors now swing both ways, and started walking through more of them.

The stated obstacle is data readiness. The real obstacle is a project manager who doesn't want to be the name attached to the failed pilot.

Better questions beat better answers

The American Express turnaround didn't come from answering the old question better. It came from replacing the question. "How do we give people more reasons to apply" became "how do we remove their fear of rejection," and the second question contained the answer the first one never could.

This is the discipline we think AI actually rewards. For decades, analysis was expensive, so companies rationed their questions and asked only the safe ones: the what-is questions. What did we sell last quarter. What does the schedule say. Now that the cost of exploring a question has collapsed, the advantage moves to companies with the imagination to ask what-if. What if we could see every project our ideal client will bid before they announce it. What if proposals took a day instead of three weeks. What if the knowledge in our best operator's head outlived his tenure.

A company's AI ceiling isn't set by its software budget. It's set by the quality of its questions. Which is why the winners of this transition will look less like early adopters and more like good askers.

Stop measuring with a borrowed ruler

There's a trap waiting inside all of this, and most companies are already in it: benchmarking. Measuring your AI progress against your competitors' AI progress, using metrics your software vendors handed you, in a race everyone entered on the same day.

The strategist Roger Martin puts it bluntly: benchmarking is for losers. If the whole market optimizes the same numbers, the whole market converges, and margin goes to whoever owns the platform everyone is bidding on. A great company isn't marginally better than its rivals on shared metrics. It's incomparable: unreasonably good at something adjacent to what everyone else is chasing.

AI is the cheapest incomparability has ever been. When every competitor uses it to write the same proposals faster, the opening is to use it for something none of them thought to measure: answering clients before competitors have opened the email, seeing a market the day it starts moving instead of the quarter after, keeping four decades of institutional knowledge alive past the retirement party. Fit in, and you disappear.

The human work is about to get more valuable

The fear underneath most AI conversations is that people become worth less. We read the evidence the other way.

When the routine layer of any job gets automated, what remains is the hard layer: the problem that requires breaking a rule, reading a room, making a judgment call no model has seen before. Scarce work commands a premium. Sutherland predicts that in a world of automated service, the remaining human specialists should be the best-paid people in the building, and we think mid-market companies should take that prediction literally. The point of AI in an industrial business isn't a smaller team. It's a team whose hours migrate up the value curve, from producing information to exercising judgment, and a company that gets more valuable per person as it grows.

A tool nobody trusts is shelfware. A team that's been taught to aim the tool at opportunity is a compounding asset.

Mindsets don't install. They're taught.

Here is the uncomfortable truth sitting under everything above: you cannot purchase a mindset. The software installs in a week. The opportunity reading of that software has to be built person by person, from the field to the front office, and this is the gap where most AI initiatives quietly die. The tool arrives, three enthusiasts adopt it, and everyone else waits politely for it to pass, the way they waited out the last system that was going to change everything.

In Jevons's England, the mill owners learned this the hard way. The efficient engine wasn't the constraint for long. The constraint became the number of people who knew how to run one. The same math governs AI: a company's real capacity isn't the tools it has licensed, it's the share of its workforce that can wield them with confidence. Twenty capable people open more seams than two hundred licenses.

So we treat training as half the work, never an afterthought. Estimators, coordinators, superintendents, the front office, the field: the whole organization learns to aim the engine, because a capability that lives in three heads is a bottleneck wearing a badge that says progress. And there's a second reason we insist on it, one we'd want stated plainly if we were the client. A capability your own people run is an asset that compounds after the engagement ends. A capability only your consultant can run is a dependency with an invoice attached. We build the first kind, on purpose, because the firms we admire most are the ones that eventually stop needing us for yesterday's problem and call us about tomorrow's.