Josh Bersin put a number on something most executives sense but cannot prove.
For each dollar spent on machine learning technology, companies may need to spend nine dollars on intangible human capital.
Nine to one. That ratio reframes the entire conversation about AI investment. Most budget decks treat AI as a cost-reduction play. The technology spend goes in, headcount comes out, and the math works on paper. Bersin's number says the math is wrong before you start. You are not buying AI instead of human capital. You are buying AI and nine times as much human capital to make it work.
The companies that understand this are outbuilding everyone else. The companies that do not are spending the nine dollars anyway, just getting nothing back for it.
Here is what the nine dollars actually buys, and the five places it leaks out before it produces a return.
What the ratio is measuring
Bersin is not talking about training budgets or change management costs. He is talking about something harder to account for: the organizational redesign work, the judgment work, the governance work, and the cultural work that has to happen before an AI deployment produces durable value.
The technology is the easy part. You can procure a capable model in an afternoon. The hard part is restructuring the work around it. Who owns the seat the agent is filling. What outcome that seat is accountable for. What the human upstream and downstream of the agent now does differently. What happens when the agent produces a wrong output. Who decides the agent is not earning its seat and pulls it.
None of that is a technology problem. All of it is a human capital problem. And it compounds. Get it wrong on the first agent and you develop habits that poison the next ten.
The failure modes
There are five places where the nine dollars leaks before it produces anything.
The first is buying AI before you have designed the seat.
Most AI deployments start with the technology. A vendor demo, an impressive benchmark, a budget approval. The tool arrives before anyone has answered what work it is replacing, what outcome it owns, or who is accountable for that outcome when the tool gets it wrong.
This is the most common failure mode and the most expensive. You pay for the tool. You pay for the integration. You pay for the rollout. And then six months later someone notices the tool is running but nothing in the business changed, because the work was never redesigned around what the tool actually does.
At Sneeze It we run twelve active agent seats. Every one of them was designed as a seat first. Dash owns ad performance analysis. Tally owns KPI value reporting. Arin owns call center coaching. Crystal owns project status tracking. The agent came second. The seat came first. The human capital we spent designing the seat before deploying the agent is the reason any of them are producing value.
The second is skipping the named human owner.
A research team at HBR and BCG ran a large experiment on this in 2026. They found that when people anthropomorphize AI agents, treating them like teammates with autonomy rather than tools with owners, accountability drops, escalation increases, and review quality falls. The finding was counterintuitive to a lot of people who had been told to manage agents more like coworkers.
The synthesis, when you read the full literature, is narrower than either camp claims. MIT SMR found that 69% of experts agree agentic AI demands new management approaches. HBR put a new role on the map: the agent manager, someone who runs agents via dashboards and scorecards. But the HBR/BCG experiment is right that anthropomorphizing the agent itself, giving it social status that creates diffuse accountability, is a specific failure mode.
The resolution is not complicated. The agent gets a seat with a metric. The human gets named as the owner of that seat. The accountability lives with the human, not with the agent. "Onboarding" an agent means scoped permissions and a clear outcome metric, not a title and a performance review. "Retiring" an agent means a human decides the seat is not producing and pulls it.
We retired Jeff, our former data integrity agent, in April. The decision was mine. Jeff did not decide to retire. I held a hearing, weighed the evidence, and made the call. Jeff's capabilities were redistributed to other seats. The accountability never moved to the agent. It was always mine.
This is not anthropomorphizing. It is accountability architecture.
The third is measuring the agent and not the outcome.
When an agent gets its own technical dashboard, separated from the business scorecard, you get a common failure mode: the agent's numbers look fine while the business outcome the agent was supposed to move stays flat.
Only 6% of business leaders in an HBR Analytic Services survey of 603 executives said they fully trust agents with core processes. That trust level is rational, but the response to it is often wrong. The response is usually more monitoring of the agent. More dashboards. More runtime metrics. More latency tracking.
The right response is simpler. Put the agent's outcome metric on the same scorecard the rest of the business runs on. When the agent's number drops, the conversation is the same as when any seat's number drops: what changed, what is the fix, who is accountable for the recovery. Not a separate technical postmortem in a different room with different language.
Radar, our chief-of-staff agent, has rows on the same Monday scorecard as Bogdan, our COO, and Janine, our accounting lead. Dirk, our sales agent, has rows next to Nick, our cold-prospecting specialist. The rows are not labeled "agent" or "human." The discipline is identical.
When the agent's row drops, someone answers for it. That is the nine dollars producing a return.
The fourth is underinvesting in the humans who now do something different.
Bersin's ratio is not just about the cost of deploying agents. It is about the cost of what happens to human work when agents absorb the operational layer.
SHRM data from 1,908 HR professionals shows AI is 5.7 times more likely to shift job responsibilities than to displace jobs outright. That means the most common outcome of deploying an agent is not that a human leaves. It is that the human's job changes. The scheduling work goes to the agent. The screening work goes to the agent. The first-draft work goes to the agent. The human now has capacity that used to be consumed by toil.
What the human does with that capacity is the nine-dollar question.
If the organization does not redesign the human's role around what the agent freed up, one of two things happens. The human fills the freed capacity with lower-value toil. Or the human starts to feel like the job is shrinking. Korn Ferry surveyed 15,000 employees across 15 markets in 2025. Forty-eight percent fear their job will be replaced by AI within three years. The fear is not irrational. It is a reasonable read on an organization that has not told them what they are supposed to do now that the agent is doing what they used to do.
The human capital investment is in that redesign conversation. What is the human now responsible for that they could not do before because the toil consumed their time. The Deloitte 2025 Global Human Capital Trends report found that managers spend around 40% of their time on administrative work versus 13% on people development. That ratio inverts when agents absorb the administrative layer. The nine dollars of human capital is what makes the inversion productive instead of disorienting.
Let agents carry the operational work, so people are free for the work that matters. That sentence is the strategy. But it only pays off if someone does the human capital work of defining what "the work that matters" means for each person whose operational work the agent just took.
The fifth is confusing the technology rollout with the culture change.
Forty-two percent of CHROs in a Korn Ferry study say they are prioritizing AI investment for HR. Only 5% feel fully prepared. The gap between 42 and 5 is not a budget gap. It is a culture gap.
Korn Ferry also found that 70% of senior leaders say their organization has an AI strategy. Thirty-nine percent of employees agree. That 31-point gap is the culture failure. The leaders know the plan. The people doing the work do not trust it. And SHRM data shows that only 25% of organizations with AI governance policies call those policies clear. Which means even the organizations that wrote the rules are not confident the rules are understood.
You can deploy the technology without closing the culture gap. The deployment will run. It will produce runtime metrics. It will not produce durable organizational value, because the humans in the system do not know what the agents are supposed to do, who is accountable when they fail, or what the humans are supposed to do differently as a result.
The nine dollars of human capital is the investment that closes the gap between the org chart the technology assumes and the org chart that actually exists.
What the return looks like when you spend it right
Bersin's Superworker framework describes the progression: Assistance to Augmentation to Replacement to Autonomy. Most organizations are stuck somewhere between Assistance and Augmentation. They have agents helping humans. They do not yet have agents carrying full seat accountability.
The organizations that reach the Replacement and Autonomy stages are not the ones with the best technology. They are the ones that invested in the human capital work first. Designed the seats before deploying the agents. Named owners for every seat. Put the agent outcomes on the same scorecard as the human outcomes. Redesigned the human roles around what the agents freed up. Closed the culture gap between what leaders know and what employees trust.
That is not nine dollars wasted. That is nine dollars producing the return that the one dollar of technology made possible.
See the live chart
Every agent seat at Sneeze It has a named human owner and a measured outcome you can query from the OTP MCP. Ask which seats are agent-owned versus human-owned, and you get back a structured view of the accountability architecture, not just a list of tools.
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 chart and tell me which seats have named human owners and which are agent-held."
The response shows you what the nine-dollar investment looks like when it is working.