How Should an AI Agent and a Human PM Divide the Work?

The previous article explained why an AI agent needs a role, limited permissions, approvals, records, and stopping rules before it joins a project team. The next blockage is usually not the principle. It is the split. Once an agent can draft notes, create tasks, and flag risks, teams still need a clear answer to a simpler question: what is the human PM for?

The split should not be based on speed. It should be based on whether an action changes project state, is hard to reverse, or affects people outside the team. This article uses that test to separate work the agent should own, work the human PM must own, and work that should move between them. The line is not a habit of one tool or one team. It is already a global operating pattern in development, documents, and communication.

Why “Just Handle It” Fails

A broad instruction looks like a goal to an agent, but it is a weak operating rule. An agent can read meeting notes, task lists, old decisions, and outdated drafts with the same weight. If the human PM does not mark what counts as the current baseline, the agent will pick the most plausible next action. It will not replace the team’s priorities.

The usual result is a pile of drafts, a few quiet schedule changes, and no clear owner when something goes wrong. Without a split, automation does not reduce work. It increases the number of outputs someone has to check.

Work the AI Agent Should Own

Give the agent work that prepares judgment without immediately changing the project. Drafting, retrieval, gap-finding, and repeated checks belong here.

• Separate decisions, follow-ups, and open questions from meeting records.

• List tasks that have no owner or due date.

• Compare requirements, change requests, and draft deliverables, then mark missing items.

• Prepare a weekly draft of status, risks, and unresolved decisions.

• Find similar past decisions and related documents as evidence candidates.

At this stage the agent’s output should be material for review, not an executed change. Dates, owners, scope, and client messages should not change before a person reads the draft.

Work the Human PM Must Own

The human PM’s job is not to write more text than the agent. It is to set the current baseline, take decisions where interests conflict, and approve actions that are hard to reverse.

• Lock this week’s goal and the work the team will not do.

• Decide scope changes, schedule resets, and priority conflicts.

• Approve messages that will go to clients, partners, or internal stakeholders.

• Make calls that reach beyond the project, including cost, contracts, and acceptance criteria.

• Keep the current source-of-truth documents and the agent’s prohibited actions up to date.

• Decide whether to stop, hand off, or redo the work when the agent fails.

If the PM hands these decisions to the agent, the work may look faster while accountability disappears. The agent can propose. The final call stays with the human PM.

The Handoff Zone

Not every task belongs to only one side. In the middle, the agent should prepare the draft and the evidence, and the human PM should approve or send it back in one pass.

The agent prepares, the PM approves

• Proposed changes to owners or due dates

• Draft client updates

• A summary of how a scope change affects dates and deliverables

• A list of files and recipients for external sharing

A handoff should include the reason, the sources used, the impact, and whether the action can be undone. The split works only if the PM can approve or reject it quickly.

Seen together, the split looks like this.

A One-Line Test

When new work appears, start with these questions.

1. Would this action change dates, owners, scope, or an external message?

2. Would a mistake be hard to reverse?

3. Would it affect a client, a contract, or cost?

4. Can it be judged from the current baseline documents alone?

If the answer to 1, 2, or 3 is yes, do not execute without the human PM. If those are no and 4 is yes, the agent can draft and the PM can review a sample. If 4 is no, the agent should stop and ask the PM for the missing baseline.

A Weekly Split

A weekly rhythm can hold the split in place.

• Agent: Prepare meeting-note drafts, unassigned work, delay candidates, and the weekly report draft.

• Human PM: In the same cycle, lock the goal, the priorities, and what may be shared with the client.

• Handoff: The agent drafts schedule changes and client updates; the PM approves them.

• Stop rule: The agent does not change dates, owners, scope, or outbound messages without approval.

With that rhythm, the agent produces the same kind of material each week, and the PM makes the same kind of decisions. The split is not a one-time document. It is a repeating operating cycle.

This Is a Global Operating Pattern

AI drafts and people decide. That is not a feature of one or two products. Coding assistants propose a change and a person merges it. Issue trackers and docs tools can draft tickets, notes, and specs, but a person still sets priority and what may be shared. Chat and email assistants can write a reply; a person still checks it before a promise goes to a client.

The names differ — human in the loop, copilot, approval queue, draft review — but the line is the same. The agent prepares material. A person changes project state. GitHub, Jira, Linear, Notion, and Slack all need that boundary. Without it, an automatic draft starts to look like an automatic decision.

The test is the same for in-house teams, outsourced projects, and freelancer work. Different tools do not change the rule. If drafts and decisions are not split, the same failure appears. This is not a tip used in one corner of the industry. It is already a global way of running work.

Checklist Before You Freeze the Split

• Is there a line between work the agent may finish and work it may only draft?

• Is there a list of actions the human PM must approve?

• Is the current baseline document marked for the agent?

• Do handoffs include the reason, evidence, and impact?

• If something goes wrong, who stops the work and explains it to the client?

• Is there a regular point in the week when the agent and the PM meet?

A Good Split Makes the PM’s Decisions Clearer

Giving work to an AI agent is not the same as removing the human PM. It means the agent takes repeated drafting and cleanup, while the PM keeps goals, priorities, external judgment, and exception handling.

When the roles overlap, the agent imitates decisions and the PM repairs outputs after the fact. When the roles are split, the agent prepares better material and the PM makes fewer, more important calls.

After you introduce an agent, the next design problem is not a longer feature list. It is the line between a draft and a decision. That line is needed first, whichever tool the team uses.