The Seven-Layer Skeptic: Why I Built My Own AI Chief of Staff Instead of Buying One Off the Shelf
Authored by: Zara Hunter
I was the most information-starved person in my own company, which is a strange thing to admit as a founder/CEO. My numbers lived in three different places and rarely agreed. Good ideas got buried under the ordinary noise of a working week. Every few months I’d realise, too late to do anything about it, that my time had wandered far from whatever I’d told myself mattered that quarter. Off-the-shelf AI assistants promised to fix this. None did, because none were built around how I actually work. So I built my own, one piece at a time, and only afterwards did anyone start calling it a Chief of Staff. I didn’t pick the name.
The Problem with Off-the-Shelf Assistants
Most AI tools are built to answer questions when you remember to ask. That’s the wrong job for what a CEO or decision maker needs. A real Chief of Staff doesn’t wait to be asked, they notice what needs attention first. A generic chatbot has no memory of your commitments, no sense of what you care about, no rhythm of its own. Buying a tool doesn’t fix that. Treating the system like a role with a job description does.
My Framework, Built in 6 Layers
- Sensing connects the accounts you already use, inbox, calendar, pipeline, so nothing needs checking manually.
- Memory Core is a handful of short, plain-language files covering clients, pricing, and recurring commitments. Keep it as several focused documents rather than one giant one. A system with less to dig through thinks faster and drifts less. This is where most people stop too early.
- Reasoning decides whether something can wait or needs a call now, whether a meeting clashes with what you’ve said matters. You’re not coding logic, you’re writing instructions the way you’d brief a new hire, then correcting them as you learn what it gets wrong.
- Action is where it does something: a routine reply drafted, a first-pass agenda, a deadline flagged early instead of the morning it’s due. Keep this narrow at first. One or two things done reliably beats ten done badly.
- Cadence means a short morning brief and an honest weekly summary, so the system becomes a habit, not one more app you forget to open.
- Interface is whatever’s easiest, chat, a voice note, a shared doc. The best one is whatever you’ll actually use before your first call of the day.
- Governance
None of this matters if you can’t trust what the system is doing, so governance isn’t optional. In practice that means three things. A clear line on what it can act on versus what only gets drafted for sign-off, anything with financial, legal, or people consequences waits. A record of what it did and why, so an outcome traces back to the instruction that caused it instead of a guess. And permissions that expire and get checked, rather than piling up forever. A system that gets more capable without getting more accountable is exactly how this goes wrong, and it’s the bit most “AI Chief of Staff” pitches never mention.
What This Actually Looks Like Day to Day
Instead of forcing every number into one dashboard, I gave Memory Core a single reconciled version pulled from all three systems, so I stopped second-guessing which figure was current before a decision. Cadence gives an honest weekly comparison, unprompted, of where my hours actually went against what I said mattered..
Where It Stops
It has no political instinct. It can’t sit in a room and clock that someone’s yes was really a no, or know which relationship needs a call instead of an email. That judgement comes from years of paying attention to people, not a memory file, and I’ve stopped waiting for it to arrive. It has no relationships of its own either. Everything it does well still depends on trust I’ve built with real people, the colleague who’ll tell me an idea is weak, the client who picks up the phone because we’ve earned each other’s time. It can protect those hours. It can’t make the relationships for me.
Start Here
Build Sensing and Memory Core first, they pay off fastest for the least effort. Add Reasoning once you trust what it notices. Keep Action narrow. The rest follows once those hold steady. What you end up with is a tireless assistant with a very good memory, not a second-in-command, and that’s the point. It doesn’t run your company. It buys back the hours and the clear head to run it yourself, and to be the person your people can still reach when it matters.
Author bio: Zara Hunter
Eduk8agentic
Agentic training for Decision makers from solo founders to team leaders