TSP has heard the same sentence in a dozen conference rooms, always after the damage is done. "Everyone knew what that system was for."
The scene that follows it is a composite, because it has happened in so many places. An AI agent built to route and approve routine supplier invoices pays a large invoice to a vendor the company had quietly stopped using. The amount is not ruinous. The question that follows is. General counsel asks who decided how much the agent was allowed to approve on its own.
The room goes quiet, and then the answers come in. The CFO believed the ceiling was five thousand dollars. The head of operations remembered twenty-five thousand. The vendor's default configuration, which no one had changed, allowed fifty thousand. Three capable people had each carried a private version of the job, and each version was reasonable. None of them had ever been said out loud to the person who was supposed to own the agent.
No one lied in that room. No one was negligent in a way a policy could have caught. The conversation that should have happened before the agent went live simply never took place, and everyone assumed someone else had held it.
TSP calls what was missing the Accountability Contract Model. Accountability is not a value posted on a wall. It is a conversation, and it has four parts: what needs to be done, what authority is granted, what success looks like, and what the timeline is. A leader has to say all four before asking anyone for results. Most leaders skip it with people, because they assume the job is understood. They skip it faster with software, because software feels like a configuration setting and not like a person who needs to be told the job.
An agent is not a setting. It acts, it spends, it commits the company to things, and it does so at a speed that leaves no time for anyone to notice that the job was never defined. So the four parts become four questions, and they belong in a one-on-one between the chief executive and the executive who will own that agent. What is the domain it acts in, and where does that domain end? What are the red flags that must stop it and send the decision to a person? How much authority does it hold, measured in dollars, counterparties, and whether the action can be undone? And when it escalates, to whom, how fast, and who answers when that person does not?
Here is the law underneath all of it.
TOUCH STONE LAW: The Law of the Unwritten Job
An agent will be held to a job no one wrote down, and the leader who never wrote it will be the one who answers for it.
A one-on-one is not a status update. It is a contract. The chief executive who tells the owner of an agent "keep me posted" has not made one. The chief executive who says "it may approve up to this amount, for these vendors, it must stop on these three conditions, and you and I will review its exceptions every Friday for the first ninety days" has. The second conversation takes eleven minutes. The first one costs a board a very long afternoon.
This is where the directors come in, and where they most often go wrong. When an AI failure surfaces, a board wants dashboards. It wants activity counts, exception rates, model versions. Activity is not accountability. A dashboard can tell a director what the agent did. It cannot tell a director whether anyone with authority ever stated what the agent was permitted to do. The board's question is simpler and harder. Ask the chief executive what was said to the person who owns this agent, and ask the owner to say it back without notes. If four people in the room give four different numbers, the conversation has not been held, and no reporting package will make up for it.
The board's job is not to write the agent's limits. That work belongs to management, and a board that takes it on has crossed a line of its own. The board's job is to insist that the chief executive has held the conversation, with a named owner, and that the owner can state the job in a paragraph.
Leaders who do this well tend to discover something they did not expect. The exercise of stating the job clearly enough for an agent also exposes the jobs that were never stated clearly for the people. The ceiling nobody had set was rarely only an agent problem.
A chief executive who holds this conversation before the agent goes live leaves its successors something specific to inherit: an owner who can say what the agent is for in a single paragraph, and an organization that no longer depends on anyone remembering what everyone knew. Built from conviction rather than crisis, this is not a paper trail. It is a company that keeps the standard correctly long after the people who set it have moved on.
Clarity has to come before results, and no agent has ever changed that order.