Human accountability when agents publish

The most dangerous sentence in an agency right now is: the AI posted it.

No it did not. A system you switched on posted it. A permission you did or did not write posted it. A brand you are paid to hold posted it. The model has no ABN, no client relationship, and no place in the dock. AI does not own the brand. A human does. If you cannot name the human, you do not have governed AI social publishing. You have a ghost with API keys.

This is step four of an online agentic closed loop. It is also the sentence clients are already asking, even when they use softer words. They are worried how you use AI. “We have a process” is not an answer. A name is.

Accountability is not a vibe

The OECD AI Principles (2019, updated 2024) put accountability on AI actors — organisations and people — not on the model. NIST’s AI RMF is voluntary, but it is not vague: trustworthiness is designed, not hoped. In Australia, the ACCC will not accept “the tool wrote it” as grounds for a claim.

Manual is not virtue. Auto is not sin.

Manual Review means a human sees, edits, or kills before the feed. Auto means the calendar fires inside the same brand rules and the same permission slip. The lever changes involvement. It does not change ownership.

People use Manual as a moral alibi (“we approve everything”) and then approve everything, which is theatre. People use Auto as a moral holiday (“it’s on autopilot”) and then are shocked by a claim. Both are failures of the same job: someone must be able to explain why this was allowed to leave.

A useful test: if the client rings at 7:12 a.m., can you show, without assembling screenshots:

• the brand truth it was generated from (Lions Law)
• the goal and the permission
• the so-what that justified ship (So What Law)
• who was accountable — named
• what would have refused it

If you cannot, you are not running agents. You are running hope. The measurement record is what that looks like when it is honest, including when something is stopped.

What principals must install

1. A named owner per brand — not a roster, a person. Deputies are fine. Anonymity is not.
2. The same rules in both modes. If Auto can say what Manual would kill, Auto is a second brand.
3. A refuse that leaves a record. Deleted-in-silence is how you teach the system nothing.
4. No “the model thought.” The model does not think in a way a client can fire. You do.

Qy Partners is the agency form of this proof: white-label delivery that still has a human accountable per client brand. If you cannot prove that, do not take the roster.

A junior with publishing access and no named owner is not “empowerment.” It is how accountability evaporates into a shift pattern. Put the name in the brand record. Put the deputy. Put the holiday cover. Then Auto is allowed to mean speed.

Fluency will always have an excuse. Authority has a name. Fluency is not authority. Restart, for the personal version of putting your name on work again: gailmacleod.ai. Soft door: app.theagencyiq.ai.

Frequently asked questions

Who is accountable when an AI agent publishes?

The named human who owns the brand in that system — not the model, not “the platform,” not whoever is on Slack.

Does Auto-publish remove accountability?

No. Auto is speed inside the same rules. Ownership does not move.

What should we show a client who asks?

Brand truth, permission, so-what, the name on the work, and what would have been refused — as a record, not a paragraph about process. So the operational question is small and unfashionable: whose name is on this post if it is wrong?Not whose login. Whose judgement. If the answer is “whoever is on Slack,” you have already transferred authority to whoever is least paying attention.