AI is moving from answering questions to exercising delegated authority
Start with the things a firm might let an AI system do, and ask of each: would you let it do this on its own? Drafting an email and recommending an offer sit low on the ladder. A person reads the draft and overrules the recommendation. So a wrong answer costs a minute. Approving a refund is different in kind. Money leaves within an agreed limit and nobody looks first. So a wrong answer is money out and a complaint back. Changing a production schedule and moving money sit higher still, and the last rung cannot be undone.
The dashed line between the second rung and the third is the point where the firm stops checking an answer and has delegated an act. Below it the firm asks whether the answer is good. Above it the firm asks what every delegation of authority has always had to answer: what may it do, for whom, and up to what amount. The rest of the set answers that question for an agent.
Capability and control are diverging
Two numbers pair across one year. AI incidents reported to one public database went from 233 in 2024 to 362 in 2025, up by more than half, while regular use of AI in at least one business function reached 88 percent of the organisations surveyed. And one firm in five among those planning agents within two years, 21 percent in a survey of 3,235 leaders in 24 countries, said it had a mature governance model for them.
Capability moves with every model release. The firm's control of delegated work moves with its policy cycle. So the gap widens by itself, and the third card says why a model error is different from a software bug: the output varies from run to run, the model is connected to data and tools, and one weak control repeats across thousands of decisions before anyone reads one. None of that is an argument against agents. It is the reason a model error becomes an enterprise event the moment the model is connected to authority, data, and action, which is where the upper rungs of the ladder put it.
Control is how a firm runs, and software is one part of it
The gap on the last slide is not closed by a product. Anyone who runs an organisation knows what control is made of, and the slide lists the things the firm already has. The law, as the firm reads it, owned by legal, compliance, and risk. Policy and limits, the risk appetite and the ceilings, owned by the board, the CRO, and the CFO. Procedures and processes, how the work is done and who may decide what, owned by the business and operations. The record of what was decided and on what basis, read by risk, internal audit, and the supervisor. People and their judgement, the tacit part the procedure leaves unsaid, held by the handlers and specialists who do the work. And the systems: identity, access, the gateway, the logs, the model and its harness, owned by the CTO and the CISO.
Only the sixth runs on a server, and it is the only one a product can sell. But every one of them is a piece of the control plane for an agent, and every owner named here has a seat at the table. So the map in part one gives each piece a box, at the altitude where its owner already works.
Control is where the value is
Turn the argument round. The value of an agent rises with how much of the work the firm lets it do alone: minutes saved when it recommends, a process run end to end when it acts. The bars show that shape from A.0, a tool that recommends, to A.4, all tasks in stated circumstances. A.5 stays off the chart: no evidence condition allows it. Each level needs a level of evidence, K0 to K3, and the set's first hard rule says autonomy may not rise above evidence; chapter 05 adds the second, on what the agent relies on.
Most firms today can prove K1: every act recorded and attributed. That licenses A.1 and A.2, and the value above the line is value the models could deliver and the firm may not take. But a record that someone who did not run the decision can rebuild is K2, and it licenses A.3. So the firm's evidence sets the ceiling on its value, and raising the ceiling is the work of the control plane. Part one draws the map of that work.