As AI agents take on more decision-making, the line between automation and autonomy determines who's accountable when things go wrong.
What they're actually buying
When organizations explain why they want agentic AI, the top answer is automation. In the 2026 SAS and IDC Data and AI Impact Report, 56% named it, ahead of competitive advantage and innovation. Companies are paying for the most autonomous systems on the market to get the thing software has sold them for decades.
Automation means work runs without a person doing it. Autonomy makes a much bigger claim: the system decides, and no person has to. The roots show the difference. Automatos was the Greek word for something self-acting, like a machine that moves on its own. Autonomos came from politics and described a city that made its own laws.
The same report shows the blur in its own numbers. About 89% of organizations say AI agents play a role in their decision-making, a figure that runs from informing a decision a human approves to acting on its own. One number captures both ends of the range, and most organizations talk about it that way, too.
Who hears what
Inside a company, autonomy already has an owner. Employees have it. It is one of the core human needs in self-determination theory, and at work people feel it as discretion: the trust to make the call on their own work.
So when leadership announces "autonomous agents," people hear something about their own discretion. A vendor means a capability. An executive hears speed. Risk and IT hear audit exposure. Frontline staff hear that something else will be making their calls.
Some of the blur is convenient. Vendors stretch automation into autonomy because autonomy sells. Some leaders shrink real autonomy back down to "just automation" because it sounds less threatening. That second version bites later, when people find out decision rights moved, and nobody told them.
Arguing for the correct word would be pedantic. The useful point is narrower: your word choice tells your workforce where decision rights are heading, whether you meant it to or not. "Delegate" tells people who handed the decision over and implies they can take it back. "Autonomous" leaves both questions open.
Where authority actually lives
At The Atlantic Festival this fall, Marinela Profi of SAS suggested leaders stop asking where they need an agent and start asking what they're willing to delegate to AI. For each process, she proposed mapping what AI decides, what a human decides, and what happens when neither is certain.
Decision rights aren't new. RACI charts and Bain's RAPID have been around for years. What AI exposes is how much authority was never on paper in the first place. Most decisions get made through judgment and a feel for when to kick something upstairs.
An agent has neither, so authority that lived in people's heads now has to be written down. That's as much a change problem as a technology one.
The third question, what happens when neither is certain, is already being answered. In the SAS report, 79% of organizations override AI in more than 10% of cases. Each override is someone answering that question on their own, one decision at a time, and mostly without anyone recording why.
The first exception
In The Rules Should Point Up, I argued that the first exception decides a transformation. Whatever leaders do the first time someone asks to bend the new rules becomes the actual rule.
Enterprise AI makes that literal. The first time an agent hits a case nobody planned for, somebody decides what happens next. That decision becomes policy whether or not anyone calls it policy, and the people watching learn where authority really sits.
What triggers the handoff
Manufacturing named this problem decades ago. Toyota called it jidoka, sometimes translated as "autonomation": a machine runs on its own, then stops and calls a person the moment something goes wrong. The machine builds the handoff in.
Jidoka works because defects are physical. A thread snaps and the loom stops. Knowledge work rarely sends that signal, and a wrong model sounds as sure as a right one. The abnormal case looks normal until someone pays for it.
So the real design question is what triggers the handoff. My answer is accountability: the handoff belongs wherever someone will have to answer for the outcome. Implement's Change Communication X-Ray 2026 points the same way for messages. In my reading of that data, people's reaction to AI-written communication hinged on whether they expected a specific person to stand behind it.
The same study found that 89% of people edit AI output and 57% disclose using it. That gap matters here. A line between human and AI decisions that nobody can see does about as much good as no line.
If I could ask every CISO one question, it would be this: if an agent took an action tomorrow that nobody intended, how long would it take you to find out, and whose name would be on it? Most can answer the first half with tooling. Far fewer can answer the second.
The second half is where the risk sits. Agents usually run on credentials sized for a person's judgment. When an agent uses them in a way nobody planned, the action is technically authorized, and nobody actually sanctioned it.
Putting a human in the loop only helps if the human is checking. Madeleine Clare Elish calls the failure case a moral crumple zone: a person positioned mainly to absorb the blame when the system breaks. A name in an approval log, attached to someone clearing 200 agent actions a day, is a crumple zone with a job title.
Understanding the decision
Every organization says it wants people using judgment around AI, and most say they're already funding it. The money usually goes to AI literacy: how the tools work and how to prompt them. That helps people use the tool. Judgment at a handoff point draws on something else, and it needs three conditions in place at once.
Understanding of the decision. People need to know what's at stake and why they are the one checking. That's sensemaking work, and tool training doesn't cover it.
Capacity to review. Someone approving 200 agent actions a day has stopped judging. Research on automation bias by Parasuraman and Manzey found that trained experts still over-rely on automated systems when volume is high, and errors are rare.
Permission to override. If overriding slows things down, challenges leadership-championed systems, or brings blame when the override turns out wrong, people defer.
Missing understanding cuts both ways. In the SAS data, lack of explanation is the top reason people override AI, at 34.5%. The same gap produces blanket rejection on one team and rubber-stamping on another, depending on stakes and incentives, which is why over-trust and under-trust show up inside the same organization.
Before the next agent goes live, write down who answers for it when it's wrong. Then check that the person can see the line, has the time to act on it, and won't be punished for using it.
***
Sources
SAS and IDC, Data and AI Impact Report: The New Economics of Trust (2026), n=2,699.
SAS Software, When AI Takes the Wheel: Debating the Promise of an Autonomous Future, recorded at The Atlantic Festival 2026.
Implement Consulting Group, Change Communication X-Ray 2026 (n=832).
Parasuraman, R. and Manzey, D. (2010). Complacency and bias in human use of automation. Human Factors.
Elish, M. C. (2019). Moral crumple zones. Engaging Science, Technology, and Society.
