The most contested question in the current literature on AI agents is not whether they are useful. It is how to govern them once they are.
One camp says manage agents like coworkers. Give them scorecards. MIT SMR found that 69 percent of experts agree agentic AI demands new management approaches. HBR has named an entirely new human role, the "agent manager," to own the relationship.
A different camp, grounded in a controlled experiment published in HBR in May 2026 by researchers from BCG, says the opposite. When people anthropomorphize agents, three things happen: individual accountability drops, unnecessary escalation rises, and review quality falls. Their model is a rented contractor with scoped permissions, a kill switch, audit logs, and a named human owner.
I have run agents at Sneeze It for over a year. Here is where I landed.
Both camps are right about the mechanism and wrong about the conclusion. The BCG experiment showed that treating agents as socially equivalent to people breaks accountability in the humans around them. The MIT SMR finding said agents need deliberate management, not casual deployment. Neither endorsed the other camp's framing.
The synthesis: an agent seat needs exactly what a well-governed contractor engagement needs, expressed as accountability architecture. A named human owner, a scoped mandate, a measured output, and a human-held kill switch. MIT SMR puts it plainly: agentic AI cannot be accountable for its decisions. The deploying human is.
Those four requirements are the structure. Each one has a failure mode when missing.
One: a named human owner
Every agent seat at Sneeze It has one human owner. Radar, my chief-of-staff agent, is my seat. Dirk, our sales agent, is my seat too, until someone is credentialed to own it. Arin, our call center manager agent, is jointly owned by me and Erica, our human call center lead. Bogdan, our COO, sits in the accountability chain for anything Crystal, our project management agent, misses.
The failure mode when an agent has no named owner is not that the agent misbehaves. It is that nobody is watching for drift. The agent runs, the numbers come out, and nobody asks whether the outputs are still accurate because nobody has been told that is their job. HBR Analytic Services surveyed 603 leaders in late 2025 and found that only six percent fully trust agents with core business processes. That distrust is mostly a structural problem, not a capability one. Without a named owner, trust never accumulates because there is no human accumulating it.
Two: a scoped mandate with hard edges
The BCG finding is most useful here. The agents that created accountability failures were not the ones with specific jobs. They were the ones whose scope was ambiguous enough that humans could project general authority onto them.
Tally, our KPI agent, pushes values from local data sources to our OTP scorecard. That is the full mandate. Tally does not interpret KPIs, set targets, or decide which seats get measured. The scope makes failure legible: if Tally is broken, the scorecard stops updating.
Dash, our analytics agent, covers Meta Ads, Google Ads, and call center data. Wide mandate, but hard edges. Dash reports patterns. Dash does not recommend changes. When Dash flags a client's cost per lead up 22 percent week over week, the observation routes to a human for a decision. Korn Ferry found that 70 percent of senior leaders say their organization has an AI strategy, but only 39 percent of employees agree. Agents with fuzzy mandates widen that gap because the people around them cannot understand what the agent actually owns.
Write the mandate before deployment. Enforce it at the permission layer where possible, not just in a document.
Three: a metric in business-outcome language
This is where most agent deployments fail fastest, because it is the easiest thing to fake.
Runtime metrics look like accountability. Messages processed per hour. Tasks completed per day. Tokens consumed. They tell you the agent is running, not whether the seat is doing anything for the business.
Dirk, our sales agent, carries three metrics: cold emails sent per week, qualified meetings booked per week, and pipeline stage transitions. The number that matters is not how many emails Dirk generated. It is how many produced a qualified meeting. Nick, our cold prospecting agent, has one KPI: quality emails drafted per day, where quality means validated, non-bouncing, sent to a named individual. Forty drafts with fifteen going to info@ addresses is not a good week. The precision is what makes the metric useful.
Bersin estimates that for each dollar spent on machine learning, organizations need to spend nine dollars on intangible human capital to realize the value. A business-outcome metric is part of that investment: someone has to define the outcome and commit to reviewing it. SHRM's 2026 research found that AI is 5.7 times more likely to shift job responsibilities than to displace jobs. That shift is into this kind of coordination work, and it only exists when the agent has a real metric.
Four: a human-held kill switch
Jeff was our data integrity agent. Named owner, defined mandate, measured output. Jeff was retired in April 2026.
The retirement was a human decision after a formal review. Jeff's metrics had not been earned, and the capabilities his seat was supposed to own had migrated to other seats. Nobody negotiated with Jeff. The seat was evaluated against one question: is this still earning its place on the chart? The answer was no. Work redistributed, record kept, chart cleaner.
That is what the kill switch looks like in practice. Not only a technical off button, though that matters. The governance-level kill switch is the structured human decision to retire a seat. The HBR Analytic Services survey found only twelve percent of organizations have risk and governance controls fully in place for agentic AI. The missing control is usually the governance one, not the technical one.
"Onboarding" an agent means establishing these four requirements before deployment. "Retiring" means a human deciding the seat is not earning its keep and closing it. Both actions are human. Accountability never moves to the agent.
The chart that holds both humans and agents
At Sneeze It, Bogdan, Janine, and Kristen hold human seats. Radar, Dash, Dirk, Pulse, Pepper, Crystal, Arin, Nick, and Tally hold agent seats. Each agent seat has a named human owner, a scoped mandate, and a business-outcome metric on the same weekly scorecard as the human seats. Jeff held a seat too, until a human decided it was not earning its place.
Korn Ferry found that 42 percent of CHROs are prioritizing AI investment but only five percent feel fully prepared. The gap is structural, not technological. The four requirements above are a structure problem, and most organizations can close it without waiting for better models.
The right framing for agent seats is accountability architecture, not digital teamwork. Accountability is a human responsibility that cannot be transferred. Let agents carry the operational work so people are free for the work that matters. Build the structure in before the agents run.
See the live chart
Every agent seat at Sneeze It is queryable from the OTP MCP server, including which seats are agent-owned versus human-owned, who the named human owner of each seat is, and what metric each agent seat carries.
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 the named human owner and business-outcome metric for each agent seat."
Accountability architecture is easier to see when the chart is live and queryable.
Series: AI-Era CHRO. Post 13 of an in-progress series.