The question most operators ask when they start building an agent fleet is: what can I give to an agent?
It is the wrong starting question.
The right question is: what should never leave a human, and why? Get that answer first, and the agent assignment decisions that follow become much cleaner. Get it wrong, and you will spend months rebuilding trust with clients, teams, and yourself that an agent quietly eroded.
I have been running a hybrid team at Sneeze It for over a year. We have agents carrying the daily operational load across sales, analytics, inbox management, project tracking, call center management, KPI reporting, and cold outreach. The system works because of what we did NOT give to agents as much as what we did. The human seats on our chart are human by design, not by default.
This post is the decision tree I use when a new piece of operational work shows up and I have to decide where it lives.
The principle before the tree
Accenture has a phrase that has stuck with me: do not make inefficiency run efficiently. The point is that you fix the process before you add the agent. But there is a companion principle they do not spell out as clearly: do not give an agent work that is inherently human, not just operationally inconvenient.
Some work is repetitive and low-judgment and belongs on an agent. Some work looks repetitive but is actually made of judgment calls that require context, relationship history, and accountability that only a named human can hold. The difference between these two categories is not always obvious. That is why the decision tree exists.
McKinsey's framing on this is useful: managing in the age of AI means managing systems of people and agents together. The COO's job is not to maximize the percentage of work that goes to agents. It is to get the right work on the right seat. Sometimes the right seat is a human. Sometimes it is an agent. The COO has to know which is which before the assignment gets made.
The decision tree
Walk any piece of operational work through these four questions in order. The first question that produces a hard answer determines where the work lives.
Question 1: Does this work require a named human to be accountable for the outcome?
Some operational decisions carry weight that cannot be diffused onto a system. A client who is told their budget is being cut needs to hear that from a person who has a name, a title, and a phone number. A team member who is being coached on performance needs that conversation to come from someone who will be in the room next week and the week after. A vendor dispute that involves a contract needs a human signature on the resolution.
This is not about what the agent can technically do. An agent can draft the budget cut email, the coaching message, and the dispute letter. The agent is not the appropriate accountable party for any of those outcomes.
At Sneeze It, Bogdan is our COO. Pepper, our executive assistant agent, handles most of the inbox. But when a client relationship is at genuine risk, Bogdan or I have the conversation. Not because the agent cannot draft something appropriate, but because the client needs to know that a human who can be held accountable is taking responsibility for the outcome. If the answer to Question 1 is yes, the work stays human. Full stop.
Question 2: Does getting this wrong require a human to absorb the consequence?
This is a sharper version of Question 1. It is not just about accountability in the abstract. It is about consequences.
When Arin, our call center manager agent, flags that a caller's appointment rate has dropped below target, the coaching response goes through me before it goes anywhere. The agent surfaces the data and drafts the message. But if the coaching is wrong, a human caller's livelihood is affected. The error does not land on the agent's row on the scorecard. It lands on mine.
Wherever the consequence of an error has to be absorbed by a named human, the judgment that produces the action should come from that same named human. Agents can prepare that judgment. They cannot own it.
This rule eliminated a class of decisions I had initially planned to give to agents. Anything that could result in someone getting paid differently, a client relationship souring, or a vendor contract changing shape stays on a human seat. The agent prepares the briefing. The human makes the call.
Question 3: Does the work require reading something that is not in the data?
Agents are good at reading what is measurable. They are not good at reading what is not.
When Crystal, our project management agent, flags that a project milestone is late, it can tell me which milestone, by how many days, and which team member owns it. What Crystal cannot tell me is whether the reason the milestone is late is because the team member is going through something hard right now, or because the client changed the scope informally three weeks ago and never updated the brief, or because there is a team dynamic issue that has been building for a month.
That context is not in Accelo. It is in the room. It is in the tone of the Slack message. It is in the pattern across six months of Thursday conversations with that person. An agent cannot read it.
Any operational work that requires reading the room, reading a relationship, or reading a pattern that has not been formally captured belongs on a human seat. Not because the agent is incapable, but because the work itself is the act of reading, and agents are not present in the room where the reading happens.
Question 4: Does this work build organizational trust over time?
Some operational work is not just about the immediate output. It is about the relationship that the work maintains or builds as a byproduct.
Nick, our cold prospecting agent, drafts outreach. Nick does not send it. I review and send, or David sends manually. The choice to keep the send on a human is partly about trust in the output quality, but it is also about something else: the relationship between the sender and the recipient is real, and real relationships are built by real people. An agent-generated email that produces a reply starts a conversation. If that conversation is going to convert, a human has to carry it forward. The human who sends the first email has a stake in the reply.
Crystal flags project risk. Bogdan acts on it. The client who gets the call from Bogdan knows that a human with skin in the game looked at their situation and decided to pick up the phone. That knowledge is part of what keeps a long-term client retained. An automated escalation does not carry the same signal.
If the work builds trust as a byproduct of being done by a human, it belongs on a human seat. This category is smaller than people think, but it is real, and it is worth protecting deliberately.
What happens when you skip the tree
The cost of getting this wrong is not usually dramatic. It is slow.
We gave an early agent work that should have stayed human. The agent handled it competently by every measurable standard. The client whose account it touched eventually churned. In the exit conversation, the client said something I have not forgotten: "I just stopped feeling like anyone was paying attention."
The agent was paying attention in every way that was trackable. It was surfacing the right data, flagging the right anomalies, hitting every metric target. What it could not do was signal presence in the way that a human who is genuinely paying attention signals it. The work looked right. The relationship eroded anyway.
Deloitte found that only 21% of companies have a mature governance model for agentic AI (State of AI in the Enterprise 2026, n=3,235). The governance failure they are measuring is mostly about risk and compliance. But there is a softer version of the same failure that does not show up in a governance audit: the failure to protect the human work that should never have moved in the first place.
The seats that are human by design at Sneeze It
Walking our own chart through the decision tree gives a clear picture of what stayed human and why.
Bogdan holds the COO seat because client risk decisions, team performance calls, and vendor disputes all require a named human who absorbs the consequence. No version of the agent fleet changes that.
Janine holds the accounting seat because financial accuracy and vendor relationships require accountability that belongs to a named person with authority over the company's money.
I hold the strategic decisions because they require reading organizational patterns that are not in any data source and because the trust that accumulates around those decisions is trust in me as a founder, not in the system.
The call center callers stay human because the person on the other end of a lead call needs to feel like they are talking to a human being who cares whether they show up for the appointment. Arin coaches them. The callers carry the conversation.
Everything else on our chart is evaluated on the decision tree. Radar, Dash, Tally, Crystal, Dirk, Pulse, Pepper, Nick, and Arin all hold seats that passed through Questions 1 through 4 and came out on the agent side. They carry the operational work. The humans are free for the work that actually requires them.
The COO call
The decision tree is not a one-time exercise. Every time a new process shows up, every time a role evolves, every time an agent's scope gets proposed for expansion, the COO walks the new work through the four questions.
The reason this is the COO's call specifically is that the COO is the one who has to live with the answer. When agent work erodes a client relationship, the COO explains it to the CEO. When a human seat is overloaded because too much judgment work is piling up on one person, the COO redesigns the process. When an agent succeeds at the task and misses the purpose, the COO is the one who traces it back and fixes the framing.
The agent fleet runs well at Sneeze It because the human seats are protected deliberately. The agents carry the operational work. The humans are present for the work that the agents cannot hold.
That is the whole design. And it starts with the question most operators skip.
See the live chart
The Sneeze It org chart, including which seats are human by design and which are agent-held, is queryable from OTP MCP.
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 tell me which seats are held by humans and what type of work keeps them on a human seat."
The structure you get back is the decision tree applied to a real running company, not a hypothetical.
Series: AI COO. Post 10 of an in-progress series.