LISTEN
Listen to this perspective
Written by Oleg Cohen. Companion narration uses the Brian voice.
On a recent call, the CEO of an established services company walked me through how his firm was using AI. He wasn’t describing a pilot someone else was running. He was describing his own work: the Markdown files that tell the model how the company quotes jobs, how he replicates those files across tools, which connectors reach which systems, what happens if the company changes model providers, and how to keep client data where it belongs.
He was thoughtful, technically capable and clearly ahead of most of his peers. He had done something this series has been arguing for. He had recognized that what the company knows is its most valuable asset, and he had started turning it into something AI could use.
He had also, without quite deciding to, become the company’s AI systems integrator.
I hear versions of this in many conversations with leaders of lower-middle-market and middle-market companies. Some try, and some don’t. The ones who try often end up here.
Who carries the work
In a large company, the work between an AI product and a working business capability is spread across a platform team, an integration partner, security, procurement and vendor management. It’s often done badly, as earlier essays in this series argued, but it’s someone’s job and it sits on a budget line.
In the middle market there is usually no such team. So the work lands on whoever cares most and understands enough, and often that’s the CEO. The person responsible for the company’s direction, people, clients and economics is now also maintaining its instruction files.
His engagement is an advantage. The problem is how much of the setup and continuing maintenance the business comes to depend on him personally carrying. Four things follow. I’d treat each as a question to investigate, not a conclusion, because some companies already have answers.
The company’s knowledge becomes configuration for a tool. A quoting file may hold years of judgment about scope, effort, risk and pricing. Its business significance is far greater than “instructions for the model,” yet it’s stored as setup for one product. Who owns it? Who approves a change? What happens to it if the company switches providers?
Buying the product still leaves the gap. A subscription doesn’t connect information, interpret rules, manage access, check results or handle exceptions. Someone has to. In this company, that someone is the CEO.
Writing it down doesn’t make anyone responsible for it. A quoting guide can describe how to estimate. It doesn’t settle who approves an exception, which revision is current, what happens when scope changes, or whether what the company learned delivering the last job improves the next estimate. In Busy With AI I wrote that a prototype isn’t a capability. A well-written instruction file isn’t one either.
The company may be trading one key-person dependency for another. In Underwriting the Heroics I argued that moving a dependency from the founder to a new CEO isn’t institutionalization. Neither is moving it into the CEO’s personal grasp of the instructions, integrations and failure modes. Apply the sabbatical test to the CEO himself. If he stepped away for three months, would the AI setup keep working, and would anyone notice when it stopped?
The larger cost is attention
The deeper problem isn’t maintenance. It’s what the maintenance displaces.
In The Rationed Company I argued that experienced attention is the constraint companies have always designed around, and that AI loosens it. There’s an irony here. The technology that lets a company stop rationing attention is, in many middle-market firms, consuming the scarcest attention in the building. Every hour the CEO spends on connectors is an hour not spent on what the company could become.
That shapes what gets built. A CEO doing the integration naturally aims the technology at the work in front of him: quoting faster, drafting proposals, summarizing calls. Those are real improvements. But the do-it-yourself path has a ceiling, and the ceiling is today’s business done somewhat better. It rarely reaches the question that matters more. With intelligence this available, what could this company now do that it never could, and how far could it go?
The middle market has a real advantage in answering that question. As I wrote in The Executive Reality Gap, the distance between insight and action is short. The person who sees the opportunity can usually change the company. That essay also argued that change depends on leaders who recognize what’s happening, have the authority to act, and own the consequences. This CEO has all three. The integrator role spends exactly that advantage.
Not a lesson. A different question.
I don’t think these leaders need educating. Many are ahead of their own organizations and most of their advisers. What helps is a different starting question.
“How do I get this tool to handle our quoting?” leads to instruction files and connectors.
“What should our quoting become, and what should I no longer have to carry personally?” leads somewhere else. It puts the company’s judgment at the center. It asks who owns each decision, and how each completed job makes the next estimate better. It also asks for something most AI programs never measure: how much of the leader’s attention the capability needs to stay healthy.
That second question is also the one that opens the larger future. A company that can rely on its own capabilities, without its CEO holding them together, is free to decide what to become next.
The test for any answer, including mine
Whatever a company chooses, whether it builds, buys, partners or hires, the test is the same. If the answer requires the CEO to become an ontology designer, an agent administrator or the permanent supervisor of the company’s AI, it has reproduced the same burden in a more sophisticated form.
A good answer should show two things. It should show how the company’s judgment becomes something the whole business can use, with clear responsibility and continuity. And it should show what the CEO no longer has to manage. Better quoting that needs less of the CEO’s attention to maintain is a capability. Better quoting that needs more is a new dependency with a nicer interface.
For owners and operating partners, a few questions make this concrete:
- How many of the CEO’s hours last month went to making AI work, and how many to deciding what the company should become?
- Who besides the CEO understands the instructions, connections and failure modes?
- Is the company’s operating knowledge stored in a form the company owns, or as configuration for one vendor?
- If the model provider changed tomorrow, what would have to be rebuilt, and by whom?
Where I sit: I advise through Kainora and am building technology aimed at this problem. That’s exactly why I hold it to the test above. If what we build adds to the CEO’s load, it has failed, however capable it looks.
The CEO who started building his company’s AI himself has shown that the demand is real. What remains is making sure the company, and not only its CEO, ends up holding the capability. Intelligence should give a CEO back the time to chart the company’s future. In too many companies right now, it’s taking that time instead.
Sources and further reading
- The Executive Reality Gap, The Rationed Company, Busy With AI and Underwriting the Heroics, earlier essays in this series.