The research literature on managing agents is split in a way worth naming before anything else.
One camp says treat agents like coworkers: onboard them, give them a performance review. MIT SMR found 69% of experts agreed agentic AI demands new management approaches, closer to human management than tool management. HBR followed with a profile of the emerging "agent manager" role.
The other camp, a May 2026 HBR and BCG paper, ran the experiment and found the opposite. Organizations that anthropomorphized agents saw reduced individual accountability, increased unnecessary escalation, and lower 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. The HBR/BCG finding is correct: accountability cannot live with the agent, and anthropomorphizing creates real problems. MIT SMR is also correct: managing an agent seat requires more discipline than managing a software subscription. The synthesis: every agent needs a named human owner, a measured seat, and human-retained accountability. The agent gets scoped permissions and a clear metric. The human owner answers for what the seat produces.
That is accountability architecture, not anthropomorphizing. Here is where the failures appear. I have made most of these mistakes myself.
Failure one: deploying an agent without assigning a human owner first
At Sneeze It, Radar runs our chief-of-staff function. Dash reads every ad account we manage and reports patterns daily. Dirk owns the sales pipeline. Tally pushes our KPIs to the scorecard four times a day. Arin manages the call center team.
Every one of those agents has a named human who owns the accountability for that seat. Not because the agent is a person, but because unowned seats drift. Silently. Confidently. For months.
SHRM's 2026 State of AI in HR found that 49% of organizations have an AI-use policy but only 25% call it clear. That gap is mostly this failure: policies that define what AI can do without defining who owns what it produces.
The fix is a seat assignment, not a policy document. Every deployed agent gets a named human owner before it runs a single task.
Failure two: writing agent metrics in runtime language
When we first deployed agents, their scorecards looked like infrastructure dashboards. Tasks completed per hour. Tokens consumed. Uptime.
These numbers tell you nothing about whether the seat is doing its job.
Dash's job is not to read ad accounts. It is to surface anomalies early enough that a human can act before they become client problems. The metric is alert accuracy and actionable timeliness. Tally's job is not to run a script on a schedule. It is to keep the KPI scorecard current so the Monday meeting has reliable numbers. The metric is staleness rate at meeting time, not script executions.
Korn Ferry found 42% of CHROs are prioritizing AI investment but only 5% feel fully prepared. Most of that gap is this: organizations invest in agents before defining what "working" looks like in terms the business can hold accountable. That definition is HR work.
Failure three: skipping governance because the agent "can't cause real harm"
Bersin: "For each dollar spent on machine learning technology, companies may need to spend nine dollars on intangible human capital." Part of that cost is governance.
HBR Analytic Services surveyed 603 senior leaders in late 2025. Only 6% fully trusted agents with core processes. Only 12% had governance controls in place. Eighty-six percent expected investment to rise. Investment scaling faster than governance is the failure the HBR/BCG paper documented.
Four components close it: scoped permissions (the agent accesses only what its job requires), a kill switch (the owner can halt output immediately), an audit log (every action recorded), and a review cadence (the owner checks on schedule, not just when something breaks).
At Sneeze It, our CRM is read-only for all agents except one write path requiring an explicit environment variable. Agents that post to Slack require human approval before anything goes out. When we retired Jeff, our former data integrity agent, we ran a formal hearing. Not because Jeff had feelings, but because we needed to document every capability the seat held and assign each to a named successor. The accountability never moved to Jeff. The hearing was a human process. Jeff was the subject of it, not a participant.
Failure four: treating agent onboarding and offboarding as optional
Agent onboarding has four required steps: define the seat's scope in writing, assign a named human owner, set scoped permissions that match that scope exactly, and establish the business-outcome metric. When all four are done, the agent is live. When any is skipped, the agent is not ready.
Agent offboarding is not turning off a subscription. The Jeff retirement took longer than it should have because we had not maintained a live scope document. We had to reconstruct what he was doing before we could assign it forward. The scope should have been written at deploy time. HR owns that checklist the same way it owns it for humans.
Failure five: letting the strategy-workforce gap close on its own
Korn Ferry found that 70% of senior leaders say their organization has an AI strategy, but only 39% of employees agree. That gap does not close passively.
When Pulse, our client retention agent, flags a churn risk, the account managers have to trust the signal and know their own accountability in the chain. If they do not, the agent flags and nobody acts. The agent did its job. The organization failed.
Forty-eight percent of employees globally told Korn Ferry they fear replacement by AI within three years. That fear makes people reluctant to trust agent output or share context with agents. HR does not close that gap with a seminar. It closes it by making the division of work visible: here is what the agent owns, here is what you own, here is where accountability flows up to a named human. Clarity is better than reassurance.
What the architecture looks like when it works
Janine owns billing. Bogdan owns operations. Radar owns chief-of-staff functions. Dash owns pattern detection. Crystal tracks project delivery. Nick drafts cold outreach. Pepper triages the inbox. Pulse monitors client health.
One seat, one owner, one measured metric. The agent gets scoped permissions and a clear brief. The human owner answers for the output. The scorecard runs the same way on Monday whether the row is an agent or a person.
The five failures above are what happens when organizations add agents before adding the governance and clarity that makes them accountable. HR closes that gap. Let agents carry the operational work so your people are free for the work that actually requires them.
See the live chart
The Sneeze It org chart is queryable from the OTP MCP, including which seats are agent-owned versus human-owned and which human carries accountability for each agent seat.
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 list every agent seat alongside its named human owner."
The accountability map is the whole model. If you can read it for one org, you can build it for yours.
Series: AI-Era CHRO. Part 37 of an in-progress series. Previous: HR does not disappear when half your workforce is agents. It changes shape entirely.