The org chart is wrong the day you deploy your first agent and leave it off the chart.
I know this because I did it. We had Radar running chief-of-staff functions for three weeks before I admitted it belonged on the chart. Radar was briefing me every morning, scanning calendar, Slack, and pipeline, compiling the daily state of the company, writing to our Obsidian vault. It was doing more coordination work than most chiefs of staff do. But it did not have a box on the chart, so nobody knew who was accountable for its output, nobody knew what it owned, and nobody knew what happened when it got something wrong.
The org chart is not decoration. It is accountability architecture. The moment you leave an agent off it, you have created an accountability gap.
Here is the decision tree I use now every time we add a seat.
First decision: does this work belong on the chart at all?
Not every agent deployment is a seat. This is the distinction that gets lost in most agent conversations.
If an agent is running as a tool inside someone else's workflow (a writing assistant inside a human's email client, a code-completion tool in an editor), it is not a seat. It is infrastructure. It belongs on a tool inventory, not an org chart.
A seat is an agent that owns a recurring output that the company depends on, has a scope it does not share with another seat, and will have its performance reviewed against a named metric. When those three conditions are true, the agent belongs on the chart with a box and a name.
At Sneeze It, Tally pushes our KPI values to the scorecard four times a day on weekdays. That output is recurring, the scope is distinct (no other seat does this), and we measure it. Tally has a seat. Nick drafts cold prospecting emails for health and wellness prospects. Recurring output, distinct scope, measured against quality emails drafted per day. Nick has a seat. Dash reads every ad account and produces a daily pattern report across Meta and Google. Seat. Arin drafts coaching messages for the call center team and reviews appointment rates against a 30% target. Seat.
The agent that writes your first draft of a routine summary email is infrastructure. When you get that call wrong, you will try to give the wrong thing a box, and the chart will become a confusion map instead of an accountability map.
Second decision: who is the named human owner?
This is where the literature on agent management splits, and I want to be direct about the split because one of the camps is wrong in a way that matters.
Camp A says to manage agents like coworkers. MIT SMR found that 69% of experts say agentic AI demands new management approaches. HBR describes an emerging role called the "agent manager" who runs agents via dashboards and scorecards. There is something correct in this: agents need metrics, review cadences, and someone who looks at their numbers the way a manager looks at a direct report's numbers.
Camp B says do not treat agents like employees. Research from HBR and BCG published in May 2026 documented that when organizations anthropomorphized agents in a large experiment, it reduced individual accountability, increased unnecessary escalation, and lowered review quality. Their prescription: treat agents like a rented contractor with a narrow statement of work, governed by scoped permissions, kill switches, audit logs, and named human owners.
Both camps are right about the architecture and wrong about what it implies.
Both camps agree that every agent needs a named human owner. Both camps agree the agent needs a measured seat with observability. Both camps agree that accountability cannot live with the agent. MIT SMR states it directly: "agentic AI (i.e., software) cannot be accountable for its decisions." The deploying human is. The camps differ only on whether you call it "management" or "governance," and the language matters less than the structure.
What OTP's one-seat-one-owner design delivers is accountability architecture, not anthropomorphizing. When I write that Dirk owns the sales pipeline, I mean that Dirk has a scope (agency revenue), a metric (qualified meetings and pipeline transitions), a named human owner (me), and a record of its outputs that I am accountable for. That is closer to Camp B's "rented contractor with a narrow statement of work" than to "treat the agent like an employee." Dirk does not have a title, a career path, or a performance improvement plan. It has a seat, a metric, and a human who owns what it produces.
Jeff, our former data integrity agent, was retired in April. He was not retired because he performed poorly in a review. He was retired because a human made the decision, in a formal hearing, that his three missions had been absorbed by better-fit seats and the seat was no longer earning its place. The accountability never moved to Jeff. The decision never moved to Jeff. Jeff was the work. We were the decision-makers.
When you draw the org chart, the named human owner goes above or adjacent to the agent seat, in whatever notation your chart uses to show accountability flow. The human's name is on it. The agent's output reports up to the human. This is the second decision: who signs for this agent's seat?
Third decision: what is the scope and the metric?
An agent seat without a scope is an agent without a brief. An agent without a metric is an agent that drifts.
Scope means: what does this seat own and what does it not own. Dash owns pattern detection across our managed ad accounts. Dash does not own client comms, strategic decisions, or the judgment call about whether a pattern is significant enough to act on. Pepper owns email triage and client draft responses. Pepper does not own Slack monitoring, calendar, or sending without approval. The scope line matters because it determines where accountability lives when something falls through the gap.
Metric means: what is the one number that tells you whether this seat is working. Not tokens consumed. Not tasks completed. The business-outcome number that the seat exists to move. For Tally, it is KPI push success rate. For Nick, it is quality emails drafted per day, where quality is defined strictly (validation passed, sent to a named individual, not a generic address). For Crystal, it is active project tracking completeness in Accelo.
When you draw the scope and the metric onto the chart, you have something you can actually operate. You can run a Monday review. You can see when the number drops. You can have the conversation about what changed.
Korn Ferry found that 42% of CHROs say AI investment is a priority, but only 5% feel fully prepared. One reason for the gap is that organizations are deploying agents without the scope and metric discipline that makes agents reviewable. The agent is running. Nobody knows if it is working.
Fourth decision: what governs the seat?
Governance is not bureaucracy. It is the set of controls that let you sleep.
Every agent seat at Sneeze It has three controls. First, scoped permissions. The agent has access to exactly what its scope requires, not more. Dash reads ad account data and writes to a shared state file. It does not write to client-facing channels or modify billing records. Nick reads from the Yelp API and writes Gmail drafts. It does not send. Dirk's write access to GHL requires an explicit environment variable, and that authorization is single-purpose. If an agent tries to operate outside its scope, the system fails closed.
Second, a kill switch. Every agent can be stopped without affecting any other seat. This is not exotic engineering. It is just design discipline. Agents are not co-mingled in ways that make stopping one impossible without stopping others.
Third, audit logs. Every agent at Sneeze It writes to a shared state file with a timestamp. When Radar compiles the morning briefing, it reads those files and flags any that are stale. The log is not a compliance artifact. It is the operational record that makes the Monday review possible.
HBR's analytic services research found that only 6% of leaders fully trust agents with core processes and only 12% have risk and governance controls fully in place. The organizations in that 6% are not trusting blindly. They have scoped permissions, kill switches, and audit logs. Trust is a product of those controls, not a prerequisite for them.
Fifth decision: when does the seat get retired?
A seat that is no longer earning its place should be retired. This is not complicated, but it requires a decision process so it happens deliberately rather than accidentally.
The trigger is when the seat's metric has stopped being meaningful (the work has moved elsewhere), the work has been absorbed by a better-fit seat (as happened with Jeff's three missions), or the scope has collapsed to the point where the seat no longer has a distinct job.
The retirement process is: document what the seat currently owns, identify where each piece of work is going, verify the receiving seat can absorb it, keep the record. Then close the seat.
What you do not do is simply stop running the agent and assume the work will find a home. The work does not find a home. It disappears. Quietly. For weeks, until something downstream breaks and you trace it back to the gap.
SHRM research found that AI is 5.7 times more likely to shift job responsibilities and 3 times more likely to create new roles than to displace jobs. The displacement fear is real (Korn Ferry measured it at 48% of the global workforce), but what actually happens is not elimination. It is redistribution. Agents absorb the operational work. Humans shift toward the work that requires judgment, care, and accountability. Bogdan is still our COO, doing the work that requires human judgment and relationship. Janine still owns billing and receivables. What changed is that the operational coordination that used to fall through the cracks is now Radar's job, and Radar is on the chart so nothing falls through.
The org chart change when HR adds agents is not cosmetic. It is the structure that makes agents carry the operational work so people are free for the work that matters. Without the chart, the agents drift. With it, the accountability holds.
See the live chart
The Sneeze It org chart, including all agent seats and their named human owners, is queryable from the OTP MCP so you can see exactly how each seat is typed and who signs for it.
In Claude Desktop or Cursor or any MCP client, add this block:
"otp": {
"command": "npx",
"args": ["-y", "@orgtp/mcp-server"]
}
Restart the client. Then ask: "Use OTP to show me the sneeze-it org chart and identify which seats are agent-owned versus human-owned, and who is the named human owner for each agent seat."
The structure you see is the accountability architecture, not the anthropomorphizing.
Series: AI-Era CHRO. Part 39 of an in-progress series.