AI Chief of Staff: What Ours Decides and What It Hands Back

An AI chief of staff is an agent that reads everything happening in a business every day and decides one thing: which of it needs a human. Everything below that line it handles, logs, or ignores. Everything above it arrives on the owner’s desk as a card with the facts, a recommendation, and three ways to answer.

That is the version we run on our own group, and the line matters more than the reading. Grant Thornton’s 2026 AI Impact Survey, fielded 23 February to 18 March 2026 across 950 senior leaders, found only 5% of organisations let agents execute high-stakes decisions without human review. The other 95% need a rule for what counts as high stakes. Most do not have one written down.

This is the third post in our series on the governance layer behind our fleet. The previous post covered the task backbone, the one store every task lives in. This one covers the agent that reads that store, plus the books, the targets and the fleet, and decides what reaches a person.

First, an honest correction to our own title

“Chief of staff” in our stack is two agents, and only one of them earns the name today.

The first is a planner. It takes a stated goal plus feedback from the business, breaks it into a sequenced plan with owners and dates, and drafts the tasks. Nothing goes live until a human approves it with one click. It keeps a log of every decision made on its plans.

What the planner does not do is the half a real chief of staff spends most of the week on: checking that the work actually got done, chasing it when it did not, and learning from the result. Its own spec says most of that vision is unbuilt, and that the missing part is largely unspecced, which is further back than unbuilt. We would rather tell you that than sell you a brain we do not have.

The second agent is the executive one. It is the subject of the rest of this post.

What the chief-of-staff agent reads every day

Our chief-of-staff agent runs daily across five inputs: finance reports, strategy targets, the task backbone, people signals, and fleet health. It produces two things: a short written brief on the state of the business, and a queue of decision cards for anything that crosses the materiality bar. Both land in a digest the owner already reads.

There is no separate app to open. That was deliberate. A dashboard competes for attention; a daily email with a short queue does not. The weekly cycle computes the full brief, and a lighter daily pulse rides in the same digest. For what the numbers brief itself contains, see our post on automated business reporting. The difference here is the queue: reporting says what happened, the chief-of-staff agent says what you need to decide.

The materiality bar: four questions, not a threshold per topic

A matter becomes a decision card if it passes any one of four tests. It crosses a money-impact threshold. It cannot be undone. It touches more than one of our companies. Or it is a hire or fire call. Pass none, and the agent either acts on its own inside a small, logged free tier, or does nothing.

The free tier is intentionally boring. The agent may nudge, re-rank, schedule, draft, link records, and create tasks in a pending state that a human still has to approve. All of it is internal and reversible.

The gated list is where the value sits. Anything that moves money, any message leaving the building, any change to people, any edit to strategy, anything irreversible or outward-facing. The agent can recommend on all of those. It cannot do any of them.

Situation	Agent’s move	Human’s move
Task overdue, owner clear	Nudges through the messaging agent, logged	None
Plan needs re-sequencing	Drafts new order as pending tasks	One-click approve
Spend over the money threshold	Decision card with data and a recommendation	Accept, override with a reason, or ask for more
Cross-company matter	Decision card	Decides
Hire or fire	Decision card, never an action	Decides, always

Why it never talks to people directly

The chief-of-staff agent has no voice of its own. When it wants a person chased, it issues a directive, and a separate messaging agent on the task backbone delivers it. Exactly one system in our fleet speaks to humans.

This sounds like plumbing. It is the reason nobody on our team gets three bots asking the same question in three tones. It also means every outbound message passes the same risk gates regardless of which agent wanted it sent. The next post in this series goes deeper on that one-voice rule.

What a decision card looks like, and what happens when nobody answers

Each card carries a title, a domain tag (finance, people, strategy, fleet), a few lines of context, a snapshot of the data behind it, and a recommendation. The answers are accept, override with a written reason, or ask for more. The recommendation is the point: a card that only asks a question pushes the thinking back onto the person it was meant to save time for.

Unanswered cards do not die quietly. The agent chases through the messaging agent, and if the card still sits undecided it expires and re-escalates into the next brief. Overflow waits in rank order rather than flooding the queue.

Every decided card also gets rated by the owner as real or noise. Those ratings feed a trust meter, the agent’s hit rate and false-alarm rate, and that meter is what we look at before widening what the agent may do alone. The evaluation agent that scores the rest of the fleet scores this one’s briefs too. No agent in our stack grades its own homework.

Where it is still rough

Three things from its own run notes, as of September 2026.

One status job that tracks escalation rates threw a database constraint error when it tried to write a snapshot with a value the table did not allow. New action types arriving from a people-messaging channel went unclassified until someone fixed the code by hand. And the count of actions the agent dispositioned dropped sharply between two windows. The system flagged that itself, with the right question: did the business get calmer, or did the agent stop noticing? That question is still open.

One enforcement phase also stays in dry-run, writing nothing, until the owner makes one specific adoption decision. And the cash check is a point-in-time look, not a multi-week forecast. Both are stated limits, not hidden ones.

Why publish this? Because most organisations are further from control than they think. The Cloud Security Alliance’s March 2026 survey of 228 IT and security professionals, commissioned by Aembit, found only 33% can change an agent’s access policies in real time. Microsoft’s 2026 Work Trend Index, published 5 May 2026 from 20,000 knowledge workers in 10 countries, found only 26% say their leadership is aligned on AI. An agent that decides what reaches leadership is only useful if leadership agrees on the bar.

What this costs you to have

Our fleet runs on this layer. As of 30 June 2026 it held 41 registered agents with 9 live in production, and across the 13 days measured to 4 July 2026 those agents completed 5,450 runs with zero failures. The numbers and their dates are in our agent fleet case study. Every tomte, our word for one production agent with one defined job, reports into the same backbone the chief-of-staff agent reads.

For a client, the materiality bar is written with you in week one: your money threshold, your irreversible actions, your always-human list. Then it runs inside the Fractional AI Officer retainer at $2,900 a month, cancel monthly, with one new automation shipped each month and one written brief a week holding the decisions that are yours. You own the rules, the logs, and the code.

If you want the part of your week that is sorting signal from noise handed to something that shows its working, start async. Describe the business in writing and we reply within one business day. No calls.