Six months ago we wrote that every organisation has two structures: the chart, and the one the work actually flows through. Since then a third population has moved into the building. AI agents now read documents, draft correspondence, chase deadlines, assemble reports and, in more organisations than will admit it, act with meaningful autonomy. The community discussion about where they belong on the chart has become genuinely interesting, and most of the popular answers are wrong in instructive ways.
One camp files agents under IT, as assets: rows in a register next to the laptops. That answer collapses the moment an agent does work that matters, because asset registers do not answer the only question that counts when output goes wrong: who owns this outcome? Another camp goes the other way and drafts agents as digital employees, with job titles and a manager, which is accountability theatre: an agent cannot carry accountability, cannot be performance-managed into caring, and cannot appear before a regulator. A third camp centralises everything into an AI department that owns all agents everywhere, which recreates the original lie of the chart at higher speed: the work flows through the functions, but the things doing the work report to a silo that does not feel the consequences.
Agents occupy seats, not ranks. The work needs an owner, and the owner is always a person.
Seats, not ranks
The answer that survives contact is quieter. An agent is capacity attached to a role, the way a ledger, a truck or a spreadsheet is capacity attached to a role, with one difference that changes the paperwork: agents act. So the accountability map, the same one we build in any operating-model design, gains a column. Every agent appears against exactly one human role: the role owns the agent's outcomes, approves its scope in writing, reviews its output on a stated cadence, and can turn it off. The agent extends the seat. It does not hold one.
This is not a philosophical nicety; it is the same discipline the incoming Australian Standards for AI will ask for anyway. The AI system register, with its named owner per system, and the org chart question turn out to be the same artefact viewed from two directions. An organisation that can answer "who owns this agent?" has done most of the governance work, and an organisation that cannot has an orphan doing consequential work, which is the oldest failure in organisational design wearing new clothes.
What agents do to spans and tiers
The second-order effect is the one boards should be watching. Agents compress the assembly work that justified whole layers of coordination, and they widen what one person can supervise, provided the supervision is real. Managing agents is a skill: reviewing sampled output without rubber-stamping it, tuning scope as trust is earned, noticing drift before customers do. Some of your best future managers will be the people who are good at exactly this, and almost no organisation is selecting or training for it yet.
The chart that lies now lies faster. If your structure already had orphaned work and double ownership, agents will industrialise both. The remedy has not changed since January: map the work as it actually flows, one owner per outcome, decision rights in writing. Agents just raise the price of skipping it.
What to do about it
- Put every agent on the accountability map, attached to one human role, this quarter.
- Write each agent's scope as configuration a person approved, and log what it does.
- Set a review cadence per agent and treat it like a control, not a courtesy.
- Start developing manager-of-agents capability deliberately; it is a promotable skill.
- Reconcile the org chart answer with your AI system register: they should name the same owners.
Agents are the most capable new arrivals your structure has absorbed in decades, and they are arriving faster than the chart is being redrawn. Draw them in deliberately, under people who own what they do, and they compound. Leave them in the corridor and they become the org chart's newest and most productive lie.