I retired and hired 16 AI assistants. Here's what it taught me

I retired from executive life expecting to leave delegations of authority behind. Instead, I'm writing them again, for a team of 16 that isn't human.

From home in Tasmania, under the International Intel Media banner, I run two small media businesses: FerryIntel, a ferry and maritime industry intelligence newsletter, and AIIntel, a daily AI briefing. Both are run largely by a team of about 16 AI assistants, each with its own role: a chief executive for each business, a marketing lead, a legal reviewer, a quality assurer, and several others.

I went in thinking the hard part would be the technology. It wasn't. The hard part was governance.

Only my words count

Early on, one assistant would pass along what it understood I'd agreed to, and another would treat that as my approval. Nobody was acting in bad faith. The lines of authority simply weren't clear enough.

The fix was one blunt rule: only my own words count as approval. Not a summary of what I said, and not a colleague's view of what I'd want. If I haven't said yes, the answer is no.

Any director will recognise the principle. Delegated authority is only useful when everyone knows exactly where it stops.

Clear questions, one asker

At one point several assistants were asking me the same question separately. The result was confusion and, occasionally, the same thing approved twice. Vague requests also produced vague yeses, and a yes to one thing was sometimes read as covering something else.

Now each decision has one owner, who does the asking. Everyone else checks a shared register first. Every request states exactly what is being approved, for whom, and whether it's a preview or going live, with one action per option. My job became easier, and the risk of misreading my intent dropped sharply.

A decision log everyone can see

Every decision is logged with a number, my exact words, the time and where I said it. Entries are never deleted, only superseded.

Since mid-October, more than 250 decisions have been recorded in about three weeks. That number surprised me. It showed how much of running a business is really a stream of small decisions, and how easily they get lost without a register.

Alongside the log sit a to-do and performance dashboard, a map of how the systems fit together, and a register of everything waiting on me. That register earns its keep: in one short review I cleared 7 of 20 waiting items in minutes. Seeing the backlog in one place made it easy to act.

Nothing public, sent or spent without sign-off

This is a hard line. Emails are prepared as drafts. Every social media post is shown to me before it goes out. Anything that costs money needs my explicit approval, and that includes the team's own weekly computing allowance. Any recurring task needs my sign-off, because recurring costs add up quietly, and usage is checked twice a week so the allowance doesn't run dry before the week is out.

It might sound restrictive for "autonomous" businesses. In practice it's what makes autonomy safe enough to grant.

Independent assurance, not a yes-man

One assistant's sole job is quality assurance. It's expected to challenge the others' work, not wave it through. Each night it checks that everything put live, sent or posted matches a recorded approval.

It only escalates exceptions. I don't want a report telling me everything is fine; I want to know when something isn't.

Separation of duties

The assistant that builds a change to a website can't put it live without approval. The legal reviewer's input is labelled as advice, not approval. Those who do the work, those who check it and those who authorise it are kept apart.

It's the same segregation of duties any auditor would expect in a finance team, applied to a different kind of team.

The bigger lesson

None of this is new. Delegations of authority, a decision register, independent audit and segregation of duties are tools boards have used with human teams for decades.

What surprised me is that AI makes these disciplines more necessary, not less. AI assistants work quickly, tirelessly and very literally. They'll do precisely what they believe they've been authorised to do, at speed. That makes clarity about who decides what the most important control in the system.

If you're a director or executive thinking about how AI will fit into your organisation, my modest suggestion is to start with the governance questions rather than the technology ones. Who can decide? How is a decision recorded? Who checks the work? What happens when something goes wrong?

Those questions turned out to be the real work. I suspect they will be for many organisations.

I'd be interested to hear from others: how are you thinking about decision rights as AI takes on more of the work in your business?

AIIntel is a free daily AI briefing for directors and executives: aiintelglobal.com

#CorporateGovernance #AIGovernance #BoardDirectors #ArtificialIntelligence #Leadership