"AI project manager" is a promise several tools now make. Motion's AI Project Manager generates projects and manages statuses, automatically updating them, assigning next steps and flagging delays. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025, and at least 15% of day-to-day work decisions to be made autonomously by agentic AI by 2028. So the useful question is not whether AI can manage a project, but which parts of the job it should take.
1. What It Does Well: the Coordination Busywork
Status updates assembled from the real board. Chasing overdue and blocked work. Triage of new requests into the right project and owner. First drafts of plans, reports and task descriptions. Turning meeting transcripts into action items. Re-planning when a date moves. These jobs share a shape: the inputs are in the system, the output is checkable, and a mistake is cheap to reverse. That is where AI earns its keep, and where most of a project lead's week goes.
2. What It Should Not Do: the Judgement Calls
Deciding priorities between two stakeholders. Approving spend or scope. Making a commitment to a client. Recording a decision on someone's behalf. These are the calls a project manager is accountable for, and they are also where an AI acting on unexpected or manipulated output does real damage: OWASP's Excessive Agency risk, which it traces to excessive functionality, permissions and autonomy. The fix it recommends is the same line drawn from the other side: minimum permissions, and a human approving high-impact actions before they are taken.
3. Why the Line Sits Where It Does
Anthropic's guidance on agents notes that autonomy brings higher costs and the potential for compounding errors, and recommends guardrails and checkpoints where the agent pauses for human feedback. A status report with a wrong number is caught at the next stand-up. A wrong email to a client, or a bulk reassignment across projects, compounds before anyone sees it. So the line is not about what AI is clever enough to do; it is about what is cheap to check and cheap to undo.
4. How Kavanah Draws It
AI employees in Kavanah prepare decisions and people make them: an employee can ask a named person to decide on a task's Approvals section, but no agent tool can approve or reject. The sprint planner packs the backlog at p50 and p90 confidence, and committing the sprint stays a human action. Meeting action items can arrive as triage candidates a person accepts or dismisses. Action policies hold categories such as outbound messages and deletions for approval, and every action lands in a ledger where many can be undone.
5. A Sensible Starting Setup
Give the AI the busywork first: a weekly status per project, a daily sweep of overdue work, meeting notes to tasks. Scope it to the projects it needs, hold anything external for approval, and read its log for two weeks. Then widen what the log shows was safe. That is the autonomy ladder our agentic project management guide describes, climbed one rung at a time.
AI project manager software takes the coordination; a person keeps the judgement. For what six tools' AI actually does, and a checklist to hold each to, see our AI project management software page.
