LISTEN
Listen to this perspective
Written by Oleg Cohen. Companion narration uses the Brian voice.
Here is a question for any company two years into its AI spending. What can the company do now that it couldn’t do before, and what does it know that it didn’t know?
For most companies the answer has two parts. The tools got better, because vendors ship improved models to every customer, competitors included. And some employees got faster, with habits that leave when they do. Neither gain belongs to the company, and the institution itself has learned almost nothing.
In What to Commit To When You Can’t Predict, I argued that owners need institutional maneuverability, along with capabilities that competitors can’t easily match. This essay explains why the way most companies adopt AI can’t produce either one, however well each project goes.
I should be precise about the claim. The tools work, and most projects meet their business case. They do the job they’re aimed at, which is making a task faster. My argument is that the job is the wrong size.
Buy tools. Build capability.
“Implement an AI quoting assistant” specifies a tool.
“Reliably price unfamiliar work, recognize delivery risk, and learn from what jobs actually cost” specifies a capability.
The first can be bought, installed and declared a success on an adoption chart. The second pulls commercial judgment, operating knowledge, decision authority, delivery capacity and evidence into scope. It also gives a board something real to evaluate. The mistake isn’t buying software, because every company will keep buying software. The mistake is treating installation, deployment or adoption as the business outcome.
Four reasons the pieces don’t add up
1. AI is being bought along the org chart. The CRM adds AI, the ERP adds AI, marketing buys a product, and employees use an assistant. Each purchase has a budget owner and a view of one slice of the company. But meaning sits between the slices. A rise in demand looks like good news. It means something else if the demand is concentrated in a low-margin segment. The same is true if it needs technicians you don’t have, or threatens commitments to your best customers. Tools bought one function at a time inherit the fragmentation they would need to overcome.
2. Everyone can buy the same tools. Try a simple test. If your closest competitor buys the identical product, what do they get? If they get everything you got, you bought parity. Parity may be necessary, but it isn’t an advantage. As the first essay argued, gains that everyone has tend to pass through to customers.
3. Nothing accumulates. Institutional learning happens when experience improves what the company decides or does next. Examples are the reason behind a customer exception, why an estimate went wrong, and what the best manager does differently. A tool that drafts faster captures none of that. Much of it was never recorded in the first place. The system says the job was completed. It doesn’t show the supervisor who drove across town to make that true. Outcomes don’t flow back into anything, and location six relearns what locations one through five already discovered.
4. Each deployment hardens today’s process. Automation gets built around the way work is done now. Forty automations are forty commitments to the current design, and that is how an efficiency program makes a company less able to change.
Forty successful pilots do not add up to a capability.
We have seen this before
When factories electrified, starting in the 1890s, most of them replaced the steam engine with one large electric motor and kept everything else. The line shafts, the belts and the building layout all still centered on a single power source, and the gains were small. The economist Paul David showed that the large productivity gains arrived in the 1920s, about forty years after the first central power stations. They came when manufacturers gave each machine its own motor and rebuilt the factory around the flow of work.
The old layout was never a requirement of manufacturing. It was a constraint of how power moved through the building. Once that constraint went away, keeping the layout was a choice, and most owners took decades to see it.
Most AI adoption today is the large motor on the old line shaft. No hold period runs forty years.
The limits worth questioning
Every established company carries limits it treats as permanent.
- “We can’t serve smaller accounts profitably.”
- “We can’t open another location without stretching our best people too thin.”
- “We can’t offer that service, because the coordination would overwhelm us.”
Some of those limits are inherent in the business. Others only reflect what knowledge, judgment and coordination used to cost. Deciding which is which is the most valuable question modern intelligence lets a company ask, and a task-level tool will never raise it.
Established companies have the most to gain from asking. An AI-native startup has speed and no legacy. It doesn’t have customers, trust, history or operating experience. An established company has all four, and they are the one asset a competitor can’t buy. But that experience sits in people’s heads, in scattered systems and in yesterday’s reports, where point tools never touch it.
Much of it the company doesn’t know it knows. Lee McCabe of Claymore Partners made this point recently about facilities management. He has seen service businesses whose managers know revenue by customer but not true contribution, once travel, supervision, rework and overtime are counted. “That is where EBITDA hides,” he wrote. I agree, and I’d add one step. Seeing the leak is analysis. Repricing the contract, rerouting the crew and deciding whether to keep the account is capability, and it crosses every function in the company.
The opportunity is to keep what took years to build and gain what used to require starting over.
So companies should use AI, and use it aggressively. What matters is what they’re building with it.
What an answer would have to do
I don’t think an off-the-shelf answer exists, and I’d distrust anyone who claims to have one. The requirements are clear enough to test any approach.
- It starts from how the business earns money and what it needs to get better at, and tool features come second.
- It works across functions, so consequences show up before a commitment and not after.
- It puts the company’s own experience to work, not only a model’s general knowledge.
- It keeps reasoning, assumptions and decision rights visible and changeable, so the company can unlearn a practice as well as learn one.
- It feeds outcomes back, so the company itself learns. That learning stays with the company through a change of vendor or owner.
- It has one accountable owner across business, technology and operations.
That is an integrated design, in which strategy, operations, technology and organization are decided together. It’s harder than buying software. The return is also different in kind. The company gains something it could not reliably do before, and the ability to build its next capability without starting over.
Where I sit: I advise through Kainora, and I’m building technology shaped by these views.
Two questions for you. Take your three largest AI investments. For each one, was it specified as a tool or as a capability? And what would a competitor get by buying the same product?