Agents on the Org Chart: Who Owns Work Nobody Fully Did
When people and AI agents split the work, managing means owning outcomes nobody fully did: who reviews, who escalates, who takes blame and who gets credit.
Nobody has managed a team that is partly software for long enough to know how it goes. The tools are new. The job titles are older than the tools. And the org chart has no box for a thing that drafts four hundred replies before lunch and cannot be put on a performance plan.
That is the honest starting point. We don’t know yet what good management of a mixed team of people and AI agents looks like, because almost nobody has run one through a full review cycle.
Companies are reorganizing around the assumption anyway. On October 26, Microsoft offered voluntary buyouts to about 60,000 employees, roughly 27 percent of its workforce by the Associated Press’s count, according to an internal memo from its chief people officer that CNBC and others reported on. Days later, Comcast notified Pennsylvania labor regulators of 634 job cuts at its Philadelphia headquarters, after executives had described putting AI agents into customer service on the company’s October earnings call, as Fox News reported. Comcast did not say the agents caused the cuts. And Microsoft’s Ignite conference runs November 17 to 20 in San Francisco, the kind of week when agent products get demoed. A demo ends right before the question this post is about begins.
This post argues that when work is split between people and agents, the manager’s job becomes owning the outcomes that no single person or agent fully did: who checks, who escalates, who takes the blame, and who gets the credit.
One refund, four hands
Take a support operations lead; call her Dana. Her case is a composite, but every piece of it is ordinary. She runs eleven people and four agents. One week, one refund went wrong.
Nobody on the chain did anything unreasonable. The triage agent tagged the ticket correctly by the rules it had. The drafting agent quoted the policy it was given, which was three weeks stale. Jon approved a reply that looked like the thirty-nine before it. The customer got the wrong amount and posted the email.
Then comes the question every manager of a mixed team will eventually face: whose fault is it? The agents have no career to adjust. Jon read the reply for about ninety seconds. Updating the policy file was nobody’s ticket. The failure lived in the seams between the steps, and no one owned the seams.
The model: four roles, and only one needs a person
Here is the model in one sentence: every unit of work on a mixed team has four roles (the doer, the checker, the escalation path, and the owner), agents can increasingly fill the first two, and the manager’s job is the seams between them.
The first two rows are where the progress is. Agents draft, agents summarize, and agents can check other agents against a rubric for routine work. The third row is half filled: an agent can route a case upward, but deciding the exception is a judgment call with consequences. The fourth row is empty on the agent side, and the reason is structural, not technical. An owner can be asked why, can change the process, and carries the result into a review. A system can do none of those.
On a team made only of people, the four roles usually collapse into one name. The person who wrote the reply also checked it, escalated it if needed, and answered for it. That is why nobody ever wrote the roles down. Adding agents pulls them apart, and the three scenarios below are what happens as they pull further.
Scenario 1: the agent is a tool
The modest case. Each person uses agents the way they use a spreadsheet.
Ownership stays obvious because the doer is still a person who chose to use the tool. If the output is wrong, it is the analyst’s output made with a tool’s help, the same way a bad formula is still the analyst’s formula. Most teams think they are here. Many already aren’t, because the volume of agent work grew faster than the habits did.
The manager’s job barely changes in this scenario. The one new duty is knowing which tools the team uses for what, so a quality problem can be traced to a person or to a tool.
Scenario 2: the agent drafts, a person reviews
The substantial case, and where Dana actually lives. Agents produce the first version of most work. People become reviewers.
Review sounds like oversight. At volume it becomes a different job, one most people were never hired or trained for. A reviewer approving forty drafts an hour has ninety seconds for each, and the drafts are fluent enough that the mistakes sit in the facts, where a quick read can’t see them. The reviewer’s name ends up on work they did not write.
Credit breaks in the other direction. When the team’s resolution numbers double, who did it? The person who rewrote the drafting agent’s instructions moved the result more than anyone who approved replies, and in most performance systems that work is invisible, because it never shows up as anyone’s tickets closed.
So the manager owns two new things. The review has to be designed: what to check, at what sample rate, with how much time per item. And credit has to be assigned on purpose to the people who improved the system, because the metrics will hand it to whoever touched the last step.
Scenario 3: the agent is the shift
The dramatic case. Agents handle whole flows end to end, and people handle whatever the agents hand up.
This is the shape the customer-contact announcements point toward: agents take the volume and a smaller human team takes the exceptions. The risk already has a name. In a 2019 paper in Engaging Science, Technology, and Society, Madeleine Clare Elish described the moral crumple zone: when an automated system fails, responsibility gets pinned on the nearest human operator, who often had little control over what the system did. She built the idea from cases like Three Mile Island and the crash of Air France Flight 447. On a support team, the crumple zone is whoever took the escalation.
Escalation in this scenario is the whole job. If the rule for handing a case to a person is vague, people get the case late, after the agent has already made two promises. If the rule is too cautious, the humans drown and the economics of the shift fall apart. Somebody has to write those thresholds and answer for them. That somebody is a manager, and the work is closer to design than to supervision.
What this gets wrong
The rows are not fixed. Agents checking agents is already reasonable for low-stakes, high-volume work, and it will get better. A model that insists a person must review everything will be wrong within a year for a lot of work. The useful version moves the line deliberately, row by row, and writes down where it moved and why.
Owning without doing has its own decay. A manager who signs for outcomes that nobody on the team fully produced still has to know what good work looks like, and that knowledge came from doing the work. If every owner stops doing any of it, their sense of what the checker should catch drifts, and nobody notices until the rubric is wrong. Someone with ownership has to keep doing a small share of the real work by hand.
The model also assumes the organization will reward ownership. Most performance systems still score individuals on their own visible output. Until that changes, a manager who credits the person who quietly fixed the drafting agent is arguing with the review form, and the review form usually wins the first round.
The manager’s job, rewritten
Any one of these would have stopped Dana’s refund. A rubric line that says “check the amount against the current policy table” turns Jon’s ninety seconds into the right ninety seconds. An owner for the policy file makes the stale policy someone’s ticket. An escalation rule for refunds above a set amount moves the risky cases to a person before anything is sent. And a credit line in the review cycle gives someone a reason to build the other three.
None of it needs a conference keynote. It needs a manager to accept that the work they are paid for has moved from the tasks to the seams between them. Start with the four-names exercise on one flow this week, and write down what you find.
