At Ozoh, the board meeting was supposed to be a victory lap. Revenue was up, margins had expanded, response times had fallen from hours to seconds, and the company had made AI visible on its org chart. Each digital agent had a sponsor, a threshold, and a dashboard. The chart looked complete.
Then Priya Nair slid a page across the table. Returns on the Alta Junior Shell were above the reserve model. Parents described the jacket as scratchy, too stiff, or impossible to get a child to wear. A school outdoor program had canceled its spring order. No single comment looked like a crisis. Together, the comments had arrived too late.
The failure was not located in one department. Demand had been real. Inventory had moved as planned. The in-store system had recommended a well-rated, price-advantaged jacket. Customer service had processed exchanges. Finance had reported improved margin quality. Every local decision was defensible. The customer, however, had experienced one company recommending one jacket for one child.
“The humans were in the loop everywhere. The loop did not contain the whole choice.”
That is the problem with an incomplete org chart. It can show who supervises each system while hiding where the organization actually forms a recommendation, creates reliance, and becomes answerable for the result.
What does the org chart show?
An org chart shows formal authority. It tells us who reports to whom, which executive sponsors a system, and which function owns a process. That map remains useful. It tells a person where to take a question and gives leaders a way to assign resources.
But AI-shaped work does not necessarily form along those lines. A demand agent can move inventory based on a weather signal. A store system can turn product attributes into advice. A service system can classify a complaint. A finance system can turn operating signals into a confident narrative. The customer does not encounter those systems one at a time. The customer encounters the combined effect.
This is why adding AI to the org chart is not the same as designing accountability for AI. Naming a system and assigning a sponsor answers, “Who can we call when this agent misbehaves?” It does not answer, “Who can stop a profitable answer because it is not responsible yet?”
Why was every local decision reasonable?
The Alta case is useful because it was not a story about an obviously broken model. The weather pattern was read accurately. Demand in the relevant markets was real. The product met its formal specification. Toma’s recommendation was truthful: the shell had water resistance, durable construction, broad availability, and a promotional price.
The missing question was suitability. Value ranked the jacket against other jackets. Suitability asked whether this particular child would actually wear it. That distinction was not owned by Product, Retail, Customer Experience, Supply Chain, Finance, or Technology in isolation.
The organization had optimized each handoff for its local job. Product saw fit language without a defect. Service saw exchanges. Retail saw conversion. Supply Chain saw demand. Finance saw leverage. The signals existed, but the structure kept resolving them locally. By the time the pattern became visible, the company had already made supplier commitments, built a financial narrative, and persuaded customers in its own voice.
The old Ozoh was slower: a merchant could hear that the jacket was technically fine and practically useless, then carry that sentence into a meeting. The friction wasted time, but it also carried meaning across boundaries.
What is the Second Map?
The Second Map is a view of the organization that shows where work, reliance, and consequence actually form, beyond the boxes of formal authority. It supplements the org chart by tracing the moments when a signal becomes a recommendation, a recommendation becomes advice, and advice becomes something a person acts on.
A useful Second Map for an AI-shaped workflow should make four questions visible: where does a machine output become a company recommendation; where does someone begin to rely on it; where can someone interrupt the workflow when the answer is technically right but humanly wrong; and who owns the consequence when several individually reasonable systems agree?
These are design questions about the work itself, not requests for another committee or a thicker risk register. In the Alta case, the crucial point was not the agent sponsor. It was the moment value became advice in a store, and the moment a manager could see that “best value” did not mean suitable for the child standing in front of them.
The Second Map also changes what “human in the loop” means. A review step is not enough if the reviewer sees only a category, a green status, or a summarized signal. A person is meaningfully in the work only when they can see the relevant context, understand what the system has treated as decisive, and stop or redirect the output without being punished for interrupting a successful metric.
How can leaders draw the missing map?
Start with one consequential workflow, not the whole enterprise. Choose a place where AI already changes what a customer, worker, or supplier does next. Then follow the work in the order it is experienced, rather than the order the functions report into one another.
Mark each transition. Identify the raw signal, the system interpretation, the human review, the recommendation, and the point where someone begins to rely on the result. At each point, name the person who can say, “This is not enough,” and define what happens when that person does.
In practice, the first useful changes are often modest: preserve raw customer language alongside classifications; let frontline managers reopen a green workflow without a performance penalty; require a suitability question before a value ranking becomes advice; and pause a recommendation while its downstream effects are still reversible.
Each move costs something. It can slow conversion, create staffing pressure, unsettle a forecast, or complicate a quarter. That is not evidence that the design is failing. It is evidence that the organization has found a place where speed had been carrying a hidden decision.
What does an organization owe the person at the point of consequence?
It owes that person a visible choice. The customer should not have to know which systems agreed before receiving advice in the company’s voice. The store manager should not have to risk a metric penalty to act on what the system missed. The executive should be able to name who can stop a profitable answer before the answer becomes a promise.
The org chart still matters. It shows the formal workforce and the lines of authority around it. But it is incomplete whenever reliance forms somewhere the chart cannot see. The Second Map supplies the missing view: not another set of boxes, but the path by which meaning travels through the organization and reaches a human life.
The org chart was complete. The organization was not. A durable organization does not merely assign systems to boxes. It makes responsibility visible where the work actually forms.
This essay draws on Chapter 1, “The Incomplete Org Chart,” of AI in the Org Chart.