The Factory of Intelligence
A map for leaders trying to think clearly about Agentic AI: the foundations, the levels people operate at, and how to use the authority, team, and expertise you already have to move the dial.
Because of my work in AI, one of the questions I get asked most by leaders is how to actually think about this industry. It all seems incredibly abstract.
Consider this the attempt to put what is in my head in my own words to explain where I think this industry is going, and what it means for leaders operating within organizations across the world.
So this is written to you, the leader, seeking to understand and think about Agentic AI.
First the foundations, so this stuff stops feeling like magic. Then a map of the levels people operate at, so you can place any tool, any vendor and any team on it. Then how I look at the whole universe — how someone in your seat uses the authority, the team and the resources you already have to turn the dial on the things that matter, and how to cut through the noise. The practical "just start doing this" bits are at the end.
The Foundations: What an "Agent" Actually Is
The most useful thing you can hold onto is that there are three different things people lazily lump together as "AI". Keep them separate and most of the confusion out there disappears:
The model is the brain — Claude, GPT and so on. Raw intelligence. On its own it's a brain in a jar: brilliant, but with no hands, no memory and no context about your world.
A harness is the thing the brain lives in — the product you type into. Claude the app is a harness. It's general-purpose and probabilistic by design: wonderful for thinking, drafting and exploring, but loose. It'll happily be 90% right, which is exactly wrong for a journal entry or a SOX control. It's a vessel for that brain: something to interact with the world through.
An agent is a specific combination of a model, a set of tools, and the right context, pointed at a job. That combination can live inside a consumer harness (Claude) or inside a custom one built for the task (more likely the latter for any real agentic work).
The distinction that matters to you: consumer tools are probabilistic — great for personal leverage, dangerous and spiky anywhere the answer has to be right every time. A custom agent lets you pin down a combination of deterministic workflows and tools — generation at the steps where judgment helps, hard rules at the steps that must not drift. That's the difference between an impressive demo and something you'd let touch the ledger. That also is the very handy property that allows you to make these things autonomous (more on that later).
There are 'levels' to this game:
If there's one lens I'd want you to walk away with, it's this: almost everyone talking about AI is standing on a different rung of the same ladder and doesn't know it: and they generalise from their rung to the whole of what AI "is". Once you can see the rungs, you can place any tool, any vendor, any breathless article, and your own team exactly where they actually sit, and you stop mistaking the bottom of the ladder for the ceiling.
The AI baked into your phone, your social apps, the box on a search page. This is most people's entire exposure to AI, and it's the worst possible sample to generalise from. These products largely aren't sold to you; you are the product. The economics underneath are ugly — providers haemorrhaging money on inference so people can generate memes and selfies — and the output is consumer-grade and unrefined. Judging AI from here is like judging aviation from a paper plane.
People paying for a Claude or ChatGPT plan, or handed a Copilot licence at work, and using it earnestly. They draft emails faster, stop hand-writing Excel formulas, tidy up documents. Useful - but they're working the same way they always did, just mildly quicker. This is where most corporate "AI investment" lives, and where most of it quietly loses money: throw generic tools at people and hope productivity appears, and you get very high yield-loss on the spend: you can't create something from nothing, you can't create context where none exists, and you can't use tools you don't have. Worse, it anchors everyone's expectations — they touch this and conclude this is all there is.
People tailoring those consumer tools to specific workflows (think most users of Claude Cowork or GPT 'Agents'): a good system prompt, a couple of connected tools. This gets sold as "agents", but it mostly isn't agentic in any real sense; it's more instructions poured into a fundamentally probabilistic system. Not bad — just not going to get the bulk of real work done, and it won't change how the underlying system behaves. This rung is also where the big industry lie lives: that scale alone will solve everything (the promise of the scaling maximalists). The trouble is you're scaling the same substrate. Same hammer, swung harder, at every job. The key: you're still in the loop, feeding context in real time. A lot of people who say "they get AI" really are just playing at this level. They are just spending a lot of their time talking to it rather than building systems to direct and orchestrate it.
Here it gets more interesting. People using developer-grade tools — Claude Code and the like — with real customisation: many MCP tool calls, their own servers, personal knowledge graphs, semantic/vector context. This is the first rung genuinely shaped to how AI works rather than bolted on top. They start automating tasks, running things after hours, standing up a few medium-length jobs that run while they do something else. This is the step right before the leverage jump because these people have taken the time to organise the context and tools AI would need to "do the work they want to do" outside of botsitting and hand-feeding it every step.
This is the leverage jump — a phase-change, not a step. Everything below rung five is you working with a powerful tool: your productivity scales linearly, because it's still a multiplier on your time and attention. At rung five you start delegating to the tool, and productivity goes non-linear, because it's no longer bounded by your hours. If I am honest, this is where I spend most of my time. In practice I don't hand-crank a custom system for every workflow — I command a fleet of extremely well-specced agents inside existing harnesses, because for most work (planning, building, coding) a well-tailored consumer harness beats a bespoke build.
On a normal day I'm running four or five parallel workstreams — several orchestrators at once, each coordinating three to five agents beneath it, some running long into the night — with what amounts to a chief of staff running across all of them. The shift in the job is subtle but total: I've stopped doing the work and primarily specify and supervise it. My AI agents are able to operate autonomously because they don't need me to tell them what context they would need and what tools to use.
The top rung — still only being scratched — is true agentic delegation: long-running workers living in custom harnesses, always on, each an expert tuned to a real job, interfacing with real systems rather than guessing probabilistically. The difference from rung five is the difference between delegating an activity and delegating an entire job. You can hand a chatbot a task with a couple of tools wired into Excel and supervise it closely; a true virtual worker you can hand the whole function and walk away. That's not a script sitting in a desktop app — it's a product in its own right, a genuine colleague.
Here's the part that matters for how you place your bets. Rung seven is simply the factory for rung six, and what makes rung six hard is that assembling these workers is genuinely difficult and the know-how is scarce (for now)— and that scarcity is exactly what commoditises next. While the world is still stuck learning rungs one to three, the real race, the one almost nobody's running, is building the factory that produces rung-six workers on demand: a self-extending, self-improving system that builds and replicates agents itself. The shovels are already in the ground on that — it's the whole point of what I'm focusing every day of my time on designing and building towards.
So the heuristic I'd hand you is this: for any tool, any pitch, any idea your team brings you — ask which rung it's actually on, and which rung the job actually needs. Most enterprise AI disappointment is a rung mismatch: buying rung two and expecting rung six, or paying to hand-craft a rung-six worker for something a rung-two subscription already does.
The truth? The painful back-office grind — the filings, the reconciliations, the compliance work — doesn't need better prompting. It needs rung-five-and-six thinking: real workers doing whole jobs that you direct rather than operate.
How I would think about it as a leader.
The ladder tells you where things sit; here's what to do with it. Three shifts, and if you internalise them you'll read every vendor, every pitch and every internal idea correctly.
The intelligence is already here and gets cheaper and better every few months, so don't anchor your strategy — or your team's habits — to one model or brand. Claude is the best general tool today; when that flips, it shouldn't cost you a thing. The brain is becoming a utility, like electricity — you don't build your business around which power station you're plugged into.
Don't buy, or build, point solutions (this is where millions are already being wasted on "custom AI builds" with the big consultancies). This is the one most people get wrong. As the ladder shows, even the hard part — assembling real agentic workers — is on its way to cheap and repeatable.
Personally, I love to look to similar analogues in history. My take: Carnegie didn't invent steel and Ford didn't invent the car; they built the machinery to produce them at scale, and that's where the era's fortunes were made. In AI terms: the hand-built "agent for one narrow problem" gets replicated or wiped out by the next model release. The durable things are a production system that can build and reconfigure agents cheaply and repeatedly, and your own domain expertise encoded so it's reusable. Hold that thought, because it's the crux of the portability of those agents to the specific and unique experience you hold as a leader.
Picture a virtual version of your own function — a treasurer, an engineer, a risk expert, a SOX expert working alongside your people, compressing time and lifting quality — this isn't a productivity hack. It's the actual shape of the next few years — those level 5 and 6 workers.
The right mental model isn't "tools my people use." It's "colleagues my people direct."
You stop adding headcount to keep up and start building a digital workforce that your existing, talented people orchestrate. Working smarter rather than simply hiring more is the whole game. The new way of working is effective delegation that runs 24/7.
What to do about all of this?
For you personally: put in an hour a day with a tool like Claude and build the muscle. It compounds, and it's real leverage. Get a personal-assistant agent running your prep, your inbox triage, your first drafts. This is the "learning to fish" part, and it's worth it.
But — and I'll be direct, because it's the most valuable thing I can tell you — don't mistake becoming a builder for the goal. Your leverage isn't in learning to wire agents together yourself. It's in the three things you already hold: the authority to say "we're doing this," a team of genuinely capable people, and years — often decades — of knowing how your domain actually works. The scarce input is that domain judgment, articulated clearly enough that a production system can build something even better around it. Point it at the right problems and you'll move the dial far harder than any amount of personal tinkering.
To put it another way: you can use the knowledge above, with the resources at your disposal, to build the ecosystem you inhabit and orchestrate — without doing the heavy lifting to build it yourself. Use your unique position to find the right partners to work with and people to support in your experimentation in this world. It's classic comparative advantage at its' best.
To help separate the signal from the noise as you navigate it: four questions you can put to any vendor or any internal idea:
If it's a clever prompt over a probabilistic model doing something that has to be right, walk away.
A thing configured once and frozen is a liability. A thing you can reshape just by describing how your needs have changed is an asset.
Does it remember and improve over time and across your domains, or is it a one-off that dies when the contract ends?
This is most common with AI add-ons to existing software products. Expensive, locked in, and human-driven.
A worry I hear a lot: if you invest years building this capability and then move on, do you lose it? It depends entirely on how it's built. The confidential data stays with the company that owns it — as it must. But the logic and the runtime, the actual capability, can be portable, and stood up wherever you go on day one. You don't build a career's worth of capability and leave it behind; you build a transportable digital team. That only holds, though, if it's built on a real production system rather than hand-assembled by whoever's cheapest — which loops straight back to the two commoditisation points above.
The honest headline: the window where being early is a genuine advantage is open now, and it won't stay open. The intelligence is already sitting there. Most of the market is running the wrong way: buying narrow point solutions that the next model release makes obsolete or lock them in to rigid ways of working. The people who move now (on the right foundations) build a lead that's very hard to catch later inside their businesses, and personally.
That's the map. Turning it into a real digital workforce for your function — a system that builds and reconfigures those specialist agents and keeps the whole thing transportable — is exactly the problem I've spent the last few years dreaming and working on.
If you've read this far, you're already thinking about it the right way. The intelligence is here and the window is open; the only real question left is whether you start now.
If you'd like to think through what this looks like for you, I'm always happy to talk.