Four operating modes for AI, with a control boundary at each step

Use these four modes to choose what authority a system should have. They can coexist; the final mode is not a destination every company needs to reach.

Stage one is the moment of impact. Somebody saw a demo. Maybe it was you. Maybe it was your CTO after a conference. Maybe it was your youngest analyst who now sends Slack messages with tool links every three days.

The demo was impressive. An agent writing code in real time. A system that read a fifty-page contract and returned a clean summary in forty seconds. Something that analyzed six months of customer data and produced a report that would have taken your team two weeks.

Whatever it was, something shifted. You started understanding what this can actually do.

Discovery lets the team test capabilities on representative material without authorizing changes to live operations.

That is stage one. The discovery. It is real, it matters, and it is not enough.

Mode 2: bounded, supervised execution

Deterministic software can enforce thresholds, schemas and allowed transitions. An AI model classifying a contract remains probabilistic. Validate its classification and use software-enforced permissions before any downstream action.

The boundary belongs in executable controls as well as written instructions. Unknown or contradictory cases must reach a named person.

This sounds simple. It rarely is. Getting to stage two means someone in your company sat down and actually mapped what a process looks like, step by step, with enough precision that a machine could follow it. Most companies have never done that. The process exists in someone's head. Usually the person who has been there eleven years, knows where everything is, and would be impossible to replace if they left.

That is the first hard thing AI forces you to do. Make explicit what has always been implicit. Write down the thinking that lives in your people. Document the process that everyone understands until the moment they need to explain it.

Mode 3: delegation within enforced limits

An agent may choose intermediate steps toward a goal, inside a permitted tool set and action budget. Delegation does not replace rules. High-impact actions retain approval, and failures need monitoring and a recovery path.

The companies that try to jump to stage three without building stage two properly discover the problem quickly. The AI delegates to what it has. If your data is messy, if your processes were never documented, if permissions and policies are implicit instead of explicit, the AI will confidently execute on garbage. At scale. Faster than any person could.

OWASP’s excessive-agency guidance places authorization in downstream systems rather than leaving the model to decide whether an action is allowed.

Mode 4: propose improvements under change control

A system can flag recurring exceptions or propose a rule change. An authorized person reviews the evidence, tests the change and approves release. The agent does not grant itself new permissions or write binding governance.

Choose the mode that fits the task and its consequences.

What matters is the question you need to answer today: which stage are you actually in?

The real test is the stage you have in production, running on real data and changing how work gets done today.

If the answer is stage one, that is fine. Knowing where you are is the whole point of the map.

Companies that remain in stage one lose ground. Every month delays the data infrastructure, documented processes, and institutional knowledge that stage two requires.

The companies that moved from stage one to stage two did not do it by evaluating more vendors. They did it by picking something real, something specific, something that mattered to the business, and building it. A system that runs, that people use, that changes how a part of the company operates. In production. On real data.

Then they did it again somewhere else.

That is how you move through the stages. Not by leaping. By building one thing that works and using that as the foundation for the next thing.

Use the first-workflow method to prove and package one bounded capability before extending it.