Skip to article

How Much Freedom Should an AI Agent Have? Autonomy Against Consequence

A workflow holds many actions, and they do not all carry the same risk. PwC's Jenn Kosar on calibrating oversight per action, and the default BNEDai starts every AI agent with.

By Jacqueline V. TwillieNovember 20, 20264 min readGuide
Quick answer

Calibrate oversight to risk, one action at a time. At an online enterprise AI summit on November 12, 2026, PwC's Jenn Kosar described a two-by-two of autonomy and consequence applied to each action. Some actions need approval before they happen, and for others a log reviewed after is enough.

Jenn Kosar, PwC's US AI Assurance Leader, described one way to decide how much freedom an AI agent should have at an online enterprise AI summit on November 12, 2026, and spoke on a related question the same day.

The two-by-two of autonomy and consequence

Kosar pictures a two-by-two, with autonomy on one axis and consequence on the other. The point, she said, is calibrating oversight and governance to the risk, and she describes risk as that mix of autonomy and consequence.

Apply it to each action, not each use case

Kosar said to apply it at the action level, not at the use case level, because a mature use case has a lot of actions in it and they will not all land in the same spot on the grid. Some actions call for approval before they happen, and for others, evidence logging and review after the fact is sufficient. Either way, she said, a human still has to be accountable for what the AI agent does.

Good governance lets an AI agent go faster

Kosar sometimes uses the analogy of brakes on a car, which are there so you can go fast. She said good control and good governance fuel innovation.

DJ Sampath, Senior Vice President and General Manager of AI Software and Platform at Cisco, said security has historically been seen as something that slows adoption down. With AI agents, he said, "it's the other way around." People want to feel comfortable that an AI agent will not do something they do not want it to do.

Opening too slowly carries a risk

On which company worries her more, one that locks AI agents out of company data or one that opens everything up, Kosar said: "The bigger risk is the far end of the conservative spectrum at this point." She comes from financial services and said you need to manage it. But right now she sees a higher risk in not starting to open the door and getting confident in your own capabilities, platforms and people.

Where BNEDai starts

BNEDai's default for any AI agent it sets up is to draft and leave the decision to a person. Jacqueline V. Twillie, Founder of BNEDai, puts it this way: "Every AI agent I set up starts with the same rule: draft, never send." BNEDai's description of what follows: an AI agent that runs while you sleep still needs your review habits. Keep it drafting and preparing until you have watched its output long enough to trust it. As trust builds you can loosen deliberately, one action at a time, for one AI agent at a time.

That is the same unit Kosar used. The two-by-two decides how much oversight each action needs, and the draft-first default is where BNEDai starts before any action has earned more. The five places a person sits in an AI agent Jacqueline V. Twillie built live on October 7, 2026 are described in what human in the loop means for an AI agent. The page on every AI agent starts with the same rule shows where each limit is set.

Try it on one workflow

BNEDai suggests listing the actions in one workflow. For each action, mark whether it needs approval before it happens or whether a log reviewed after is enough, and name the person accountable. Start every AI agent drafting, and loosen one action at a time as the output earns it.

Frequently asked questions

What are the two axes in Kosar's framework?

Autonomy and consequence. She applies them at the action level and not the use case level.

Does every action need approval before it happens?

No. Kosar said some actions call for approval before, and for others evidence logging and review after the fact is sufficient. A human stays accountable either way.

What is BNEDai's default?

Every AI agent starts with the same rule, in Jacqueline V. Twillie's words: draft, never send.

Source: remarks by Jenn Kosar, PwC's US AI Assurance Leader, and DJ Sampath, Senior Vice President and General Manager of AI Software and Platform at Cisco, at an online enterprise AI summit, November 12, 2026. The draft-first default is Jacqueline V. Twillie's, and the recommendations are BNEDai's.

Reviewed on November 20, 2026.

More on Ownership for Corporate Executives