There is a version of the agent-as-employee argument that sounds right until it costs you something.
It goes like this: agents fill seats on your org chart, so treat them the way you treat new hires. Write them a job description. Give them an onboarding checklist. Run them through a thirty-day performance review. Retire them when the fit is wrong. The analogy is clean and operators like clean analogies.
The research does not support it. And in May 2026, a large-scale experiment published in HBR by Kropp et al. with BCG showed why. Organizations that were instructed to treat AI agents like employees showed measurable drops in individual accountability, more unnecessary escalation, and lower review quality than those that did not. The anthropomorphizing framing was not neutral. It shifted human behavior in ways that hurt outcomes.
So do you ignore the onboarding question entirely? No. That is the other trap.
Here is where I have landed after eight months of running a hybrid team.
The literature is fighting about language, not architecture
The debate splits into two camps. Camp A, led by MIT SMR and HBR's work on the emerging "agent manager" role, says agentic AI must be managed more like a coworker than a traditional tool. Sixty-nine percent of experts MIT SMR surveyed agreed that agents demand entirely new management approaches.
Camp B, the Kropp et al. HBR/BCG paper, says the coworker framing backfires. Treat agents like rented contractors with a narrow statement of work. Scoped permissions. Kill switches. Audit logs. Named human owners. Not job titles or performance reviews that imply the agent is a peer.
Read both carefully and they agree on the substance. Both say every agent needs a named human owner and a measured seat with observable outputs. Both say accountability cannot transfer to the agent. As MIT SMR put it plainly: "Agentic AI, i.e., software, cannot be accountable for its decisions." The human deploying it is.
The fight is about framing. The architecture is the same.
What onboarding actually means
When I deployed Radar as our chief-of-staff agent, I did not write a welcome document. I did four things.
I defined the seat. What output does this agent produce? For Radar it was: daily briefing compiled from Slack, calendar, email, and six shared-state files, written to Obsidian by 8 AM. One deliverable with a named format.
I scoped the permissions. Radar reads Slack, reads Google Calendar, reads Gmail threads, reads shared-state files from other agents, and writes to Obsidian. That is the full surface. Radar does not send Slack messages without approval. Radar does not modify calendar events at all. The scope is explicit, written down, and enforced by the system, not by the agent's good judgment.
I named the owner. In Radar's case, that is me. When the briefing is wrong, incomplete, or stale, I am the person who diagnoses it and fixes it. The agent produces the output. I own the accountability for the seat.
I built the kill switch. If Radar starts producing briefings that are actively wrong, I have a clear path to pause, redirect, or retire the seat. That path does not require a performance improvement plan. It requires a decision from the human who owns the seat.
That is onboarding. Four steps. No welcome lunch.
The same four steps hold across every agent on our chart. Tally pushes KPIs. Dash reads ad accounts. Arin drafts call center coaching messages. Crystal tracks project delivery. Nick runs cold prospecting to a hard-scoped list with a specific output: thirty qualified email drafts per day, no sending without approval. Each one has a defined output, a scoped permission surface, a named human owner, and a clear path to retirement.
None of them have job titles in the sense that implies peer status. None are evaluated on culture fit. But all of them appear on the same scorecard as Bogdan, our COO, and Janine, our accounting lead, because the output of their seat belongs on the same accountability surface as every other seat in the company.
That is not anthropomorphizing. That is governance.
Retiring a seat is a human decision
In April, we retired Jeff, our data integrity agent. The scanner had gone stale for five-plus days at a stretch, false positives had accumulated, and he had sent a DM to Bogdan without authorization. We held a hearing. A person decided the seat was not earning its place and closed it. The capabilities were redistributed to Dash and Dan.
That is what retirement looks like in an accountability architecture. It looks like a decision, not a deprecation.
The HBR/BCG finding makes sense here. If you frame an agent as a coworker, the natural response when it underperforms is patience, coaching, benefit of the doubt. Those responses are rational for humans and destructive for agents, because agents do not have arcs. They either have a clear seat that produces a clear output, or they do not, and the fix is structural, not relational.
Why the architecture closes the trust gap
The reason OTP is built around one seat, one owner is not because we think agents are people. It is because accountability needs to live somewhere visible, or it does not live anywhere.
Korn Ferry found that seventy percent of senior leaders say their organization has an AI strategy, while only thirty-nine percent of employees agree. That gap is not a communications problem. It is an architecture problem. When accountability for AI outputs is diffuse, nobody can tell you who owns the gap between what the strategy promised and what it delivered.
One seat, one owner collapses that gap. When Dirk's pipeline output drops, I know whose row to look at. When Pulse flags a client retention risk, I know who to route it through. When Pepper drafts a client email that misreads the tone, I know the approval step is mine before anything sends.
The HBR Analytic Services survey of 603 leaders found that only six percent fully trust agents with core business processes. That trust gap closes, slowly and verifiably, when every agent has a named owner and a measured seat. It does not close by calling agents coworkers.
SHRM data from 1,908 HR professionals shows that AI is five-point-seven times more likely to shift job responsibilities than to displace jobs outright. The shift is into oversight and decision-making, which is exactly what the named-owner seat structure requires. The human who owns Radar's seat is not doing Radar's work. The human is doing the work Radar cannot do: judgment, course correction, and the decision about what the output is actually for.
Let agents carry the operational work. That is the point. So people are free for the work that matters. But the architecture that makes that real is not job titles and HR rituals. It is a named owner, a scoped permission surface, a measured output, and a kill switch that works.
See the live chart
Every agent seat at Sneeze It, along with its named human owner and its position relative to human seats, is queryable from OTP's MCP server.
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 which seats on the Sneeze It chart are agent-owned vs human-owned and who the named owner is for each agent seat."
That query returns the accountability architecture in a structured format. The chart is not a metaphor. It is a live governance artifact.
Series: AI-era CHRO. Post 10 of an in-progress series. Previous: HR does not disappear when half your workforce is agents. It changes shape entirely.