What to look for in an AI chief of staff: a buyer's guide

The phrase AI chief of staff is now attached to a wide range of products. Some sort email. Some summarise meetings. Some are a chat window with a new name. For a managing director trying to choose, the label says very little.

A human chief of staff is defined by what they do, not by their title. They know the business, keep track of what has been agreed, notice what is drifting, prepare the leader for what is coming and make sure things get done. They also know what is confidential and when to ask before acting.

An AI chief of staff should be judged the same way. Because the best of these systems are agentic AI, meaning AI that plans and acts rather than only answering, the questions that matter are about judgement, memory and control as much as cleverness. Here are eight to ask before you choose.

1. Does it see the whole business, or one channel?

Inbox sorting is useful, and for a busy leader it can save real time. But it is the starting point, not the job. Context lives across email, meetings, documents and chats. A promise made on a call, a worry raised in a meeting and a change buried in a document never reach the inbox.

Ask what the system reads, and whether it joins those sources into one picture or treats each in isolation. An AI that watches only one window is a mail sorter, however good.

2. Does it remember, and can it show its source?

A chief of staff is valuable because they remember. Ask how the AI builds and keeps its memory of the business, and how accurate that memory is when tested on real questions about the company's records.

Just as important, ask whether every answer comes with its source. A leader making a decision needs to know where a fact came from, not only that the AI sounds confident.

3. Does it speak up first?

If you have to ask, you have already lost some of the value. A real AI chief of staff starts the day with a brief, prepares you before meetings, writes up what was agreed afterwards and raises the commitment that is about to slip. Ask what it does without being prompted, and how often.

4. Who sees what?

A company AI will hold information that not everyone should see: salaries, board papers, a manager's notes on a team. Ask how the system decides who sees what, and whether that follows each person's role.

Then ask how it was tested. Rules on paper are one thing. Deliberate attempts to get around them, on a realistic company setup, are the evidence that counts.

5. Where does it act alone, and where does it ask?

An agentic system can draft, sort, book, send and spend. The question is which of those it does on its own. Ask what waits for a person's approval, whether you can adjust that, and whether there is a record of every action with the reason for it.

A sensible answer keeps anything that goes out of the company, or costs money, behind a person's yes, and lets everything else be tuned to how you work.

6. Where does it run, and who else is there?

Ask whether your company's AI runs on a private server dedicated to your company, with its own accounts, or on a platform shared with other customers. The answer shapes how much you will be comfortable letting it see, which in turn shapes how useful it can be.

7. Who sets it up and keeps it running?

Many AI tools are sold as software and left for the customer to configure. In most companies, nobody has the time. Ask who installs it, who teaches it how the business works and who looks after it once it is live. Ask too whether it is tied to a single AI provider, or can move as the technology changes.

8. What is the evidence?

Finally, ask for results rather than adjectives. How was the system tested? On what kind of setup? What was measured, and what was the outcome? A provider that runs its own business on the product, and can show its test results, is telling you something no brochure can.

A short checklist

Taken together, a strong AI chief of staff should pass every line below.

• Joins up email, meetings, documents and chats.
• Remembers the business accurately and shows its sources.
• Speaks up first, every day.
• Shows each person only what their role allows.
• Asks before anything goes out or costs money, and logs every action.
• Runs on a private server dedicated to your company.
• Is installed, taught and run for you, and is not tied to one AI provider.
• Comes with evidence.

How Anabel answers these questions

Anabel is turnkey: we install her, teach her the business and run her, on a private server dedicated to your company that we provide and manage, with its own accounts and its own AI account, walled off from every other company. She builds one company memory from email, meetings, documents and chats, and each person sees only what their role allows. Anything that goes out or costs money waits for a person's approval, everything else is configurable, and every action is logged with its reason. She is never tied to one AI provider, and we run our own companies on her.

We tested her on a full company setup. More than 160 deliberate attempts to reach information a person should not see all failed. She answered 97% of questions about the company's records correctly, and 98% of her answers showed the right source. On GateMem, Anabel scored higher than every memory system it tested, across all of its area tests. Those were our own runs; the published results used other AI models.