My wife asks me questions about our plants. Until recently, I couldn't answer them.
Which trees are getting enough sun. Whether the bed by my office window is failing because of shade or soil. I didn't know. And "I don't know" wasn't an acceptable answer at my house.
So I took it to Claude with two things: a goal — answer any question about sun and our plants with complete accuracy — and an inventory of what I had to work with. Our Real Property Report. Satellite imagery. Security cameras covering the yard.
What I didn't hand it was a plan. And that's where it got interesting: it started telling me what it needed. When I sent a photo of a struggling bed, it spotted that the neighbour's hedge was throwing shade I hadn't accounted for — then sent me out to the street to photograph the front of the house so it could calculate the hedge height against the building's dimensions. It flagged that sun patterns shift in September and turned the cameras I'd mentioned into an assignment: snapshots at specific times, on clear days, to refine the model.
It brought assets of its own, too — the solar geometry, the seasonal sun paths, the math. My side was the problem, the property, and the legwork. Its side was knowing what was missing and how to combine all of it. Neither of us builds this alone.
The result is what we call the SunMap: our property divided into seven zones — morning sun, afternoon sun, hot afternoon sun — with every plant and tree mapped against it.
A backyard project. But it runs on the lesson I've carried through three decades of adopting new technology: be ruthlessly clear about the problem you're solving and the value it creates. Here, the problem was my wife's questions. The value? Happy wife, happy life.
What the age of AI adds is the second half:
Don't tell AI what to do. Tell it what you want — and show it what you have. It will tell you what's missing.
Everything else I do with AI runs on that principle — at three levels. My personal life. My business operations. My client consultations.
The digital brain — and the fleet that runs on it
The foundation is what I call my vault: a digital brain built from plain text files, version-controlled, that holds a version of my thinking, my decisions, and the related context regarding everything I deem important. And I mean everything — the business, the investments, the house, health, my parents' care, the big decisions. Not a filing cabinet. A knowledge base of how I live, with the reasoning attached: what I choose, what I discard, and why.
Two rules shaped its design, worked out in the build with AI itself: low friction in, low friction out.
If capturing is work, you stop capturing. If retrieval is slow, the AI claim is really window dressing.
Retrieval is where it earns its keep. I sit in medical appointments with my father. I wrote about that journey in When Leadership Became Personal: Leading Without a Clean Handoff. When a doctor asks what was discussed three appointments ago, I have the answer in seconds. Doctors don't have time to wait, and neither does the decision. Same mechanism when my dad's health status shifts or a major household purchase is on the table: the vault surfaces timely, relevant information our brains have long since forgotten.
And it's model-agnostic. The SunMap was built with Claude; the same context now serves ChatGPT and Gemini. Models keep changing. My context is mine, and it compounds.
On top of the vault runs a small fleet of AI agents — on a machine in my office, running on schedules. They capture what comes in, keep my operating cadence, and watch each other for failures, handing work to one another through the vault. Let me be precise, because this space is full of overstatement: most of what they do still routes through me for approval. But the direction is one-way — with every use case we run, the agents earn a little more room to act without me. It works the way trust works with a new hire: small scope, reviewed output, expanding autonomy. Repeat. I'm not prompting a chatbot. I'm managing a small team that's steadily earning its independence.
And here's a part I didn't expect. I haven't really coded since university — I figured out early that programming wasn't my strong suit, so I built my career on the business side of tech. AI closed that gap. Building things used to require a programmer. Years ago at Yellow Pages, I had to hand an idea to a developer, wait for the prototype, then wait longer to learn whether it had business value. Now I prototype, prove the value, and bring in the developers to build it right.
One identity system, AI-extended
The website you're reading is hand-authored. No CMS, no plugins, no build system — plain files that will work unchanged in ten years. AI helped build it; I can maintain it with a text editor.
But the site is just the visible tip. The same brand system — the colours, the type, the templates — extends through everything my business produces: proposals, one-pagers, invoices, timesheets. All of it generated and kept consistent by AI, from one set of standards I own.
That used to require an agency. It's now one person with a system. That's the honest economics of this shift — not the tool subscriptions you cancel, but the entire categories of work you no longer need to buy.
Inside a client project: the automation hub
The third level is the live one. At RightMetric, a digital research firm, I'm helping build what we call the automation hub — the intelligence system that supports their operations. Four layers: a core knowledge base that holds what the company collectively knows, an input layer that keeps it current, an LLM layer that reasons over it, and an agent layer that delivers work against it.
The goal is bigger than automating tasks. It's giving each employee a digital twin — an agent counterpart that carries the institutional knowledge to deliver against the workflow. The humans carry the judgment, the client relationships, and the problems nobody has seen before. The business result so far: triple-digit year-over-year revenue growth without the need to hire headcount.
You don't have to take my word for the inside view. RightMetric's co-CEO Charlie Grinnell wrote about it himself — "I Gave AI Access to My Entire Business. Here's How I Sleep at Night" — a piece that grew out of many of these conversations.
If you've read my thinking on organizational intelligence, this is what it looks like in production: captured decisions and institutional knowledge systematized with agents operating against the knowledge the company actually owns.
What breaks and what I'd tell you to do Monday
None of this worked the first time. Agents fail silently — that's why the fleet includes a watchdog whose only job is catching the others. The SunMap still isn't finished; it's waiting on September sunlight patterns. My own instinct is to over-build. The rule I enforce on myself — and on clients — is what I call blast radius: match the build to what a mistake would cost. My father's medical records get built for accuracy. Financial data gets built for security. A backyard sun map gets built fast. Speed, accuracy, security — every build trades among the three, so decide what a mistake costs before you decide how to build. And when the blast radius is small, the lightest system that works wins.
If you're starting from zero, the playbook is short:
1. State outcomes, not instructions. Give the machine the goal, the constraints, and an inventory of what you have to work with. Let it surprise you with the path.
2. Build the knowledge base before the agents. The order matters: information first, input second, reasoning third, agents last. Agents without owned knowledge just improvise faster.
3. Own your context, rent your models. Keep your knowledge in a form you control. Every model should be replaceable; your accumulated judgment shouldn't be.
4. Measure output change, not activity. Prompts written and hours saved tell you nothing. Health wins, a flourishing garden, and revenue per headcount tell you whether the process is actually different.
The question I can't shake
When my parents moved into assisted living, we did the groundwork while everyone was still healthy: powers of attorney, personal directives for medical care. After my father's dementia took hold, that preparation was the difference between grief and grief plus chaos. When a doctor deems someone incapable, the paperwork you didn't do becomes the crisis you didn't need.
On a recent walk, my parents' financial advisor and I were talking about exactly that: the decisions you make while you're healthy, because eventually every family gets tested. Get your ducks in a row before you need them. Somewhere on that walk it struck me — a power of attorney is society's answer to a hard question: who acts for you when you can't? We hand that authority to the person who knows us best.
In twenty or thirty years, with decades of my decisions, values, and reasoning captured — why wouldn't AI act on my behalf, legally? It will know more about how I think than anyone else possibly could.
Can an AI be your agent for financial or health decisions? Most people's instinct is an immediate no. But the vault will hold a more intimate record of how I make decisions than any person could carry. Based on the value I'm already getting from captured decisions, I'm not sure the answer stays no. And either way, the infrastructure for that answer is being built now — one captured decision at a time.
The technology is available to everyone. The system you build around it is the part you own.
Start with what you want.
The machine will ask you for what it needs.