The bottom of the ladder
On a training call last week, a woman told me she was at the bottom.
We'd spent most of an hour on her situation. Her company runs on Microsoft, her IT department keeps things sensibly locked down, and she wanted something modest: a way to have the tender notification emails piling into her inbox screened and summarised, so she could stop reading them one by one. Then I showed her the system I run for myself, the one that reads my tenders while I sleep, and she went quiet. "I'm on the very, very bottom," she said. "I would never get there."
She was wrong, and not about where she stood. She was wrong about the picture. She thought she was looking at a ladder. She was looking at a map.
There are three ways of working with AI, and one question tells them apart. Not how clever the AI is, and not what it costs. Just this: for the piece of work in front of you, are you using the AI, directing it, or running it? You use a product. You direct a tool. You run a service. That's the whole map.
Here's the same idea in your hands, literally. When you use a product, your hands are on the work and the AI advises through a window. When you direct a tool, its hands are on the work and your eyes are on it. When you run a service, its hands are on the work and nobody is watching at all.
You use a product
Most of us met AI as a product: a chat window in someone else's app, under someone else's rules. You type, it answers, and you carry the useful bits back into your real work. Every decision about how the thing operates was made before you arrived.
That's not a criticism. Products are where nearly everyone should start, and where plenty of work should stay. The woman on my call was using AI as a product, and using it well - summarising documents, pulling wage rates out of months of correspondence, asking questions of her own inbox. Real value, delivered through a chat window, on a system she'll never see and doesn't need to think about.
You direct a tool
A tool works under your direction, on your files and systems, while you watch. Coding agents like Claude Code are the clearest example today: you point one at a job, it does the work in front of you, and you steer.
The line between using and directing is simple: whose hands are on the actual work? With a product, yours are - the AI is advising, you're doing. With a tool, that flips. It does the work; you supervise. Something shifts the first time you feel that. You've stopped being a visitor in someone else's app and started operating machinery of your own, though the engine inside is still theirs.
You run a service
A service is a system that works without you.
Here's mine. Early every morning, before I've looked at a screen, a program on my server wakes up and reads through my incoming email, looking for messages from the public tenders website. It pulls out the link, archives the email so it never reaches my inbox, and visits the tender page itself. Then it scores what it finds. Training, digital, AI: plus fifty. Tarmac, housing, playground fencing: minus fifty. Sligo, my home county: plus ten. Anything scoring over seventy earns an email to me: here's a tender you might actually want. Everything else disappears. The system has read a great many tenders about tarmac so that I never have to.
The line between directing and running is different. It isn't whose hands are on the work - it's whether you're in the room. And here's the part worth slowing down on: a tool you start and then walk away from is not a service. It's a tool with the supervision stripped off, which is a risky thing to leave running. A service is built for your absence from the start. The checking is designed in - logs, spot samples, a standing rule that says when in doubt, stop and fetch the human. The boundaries are set before anyone leaves the room. Rules decide when it wakes and what it may touch: a timer, an arriving email, a schedule. The AI supplies the judgement inside those boundaries - and the boundaries are enforced by permissions, not promises.
Same engine, three relationships
Notice what separates the three ways. It isn't intelligence, and it isn't how advanced the AI is. The same AI can be all three. Claude in the chat app is a product. Claude Code under your direction is a tool. Claude Code set to run on a schedule is a service. Nothing about the model changed. What changed is your relationship to it, from using, to directing, to running.
That's also why this is a lens for thinking rather than a set of boxes to file things in once and for all. The same software can sit in different places on the map depending on how you're working with it that day. When something doesn't fit neatly, the honest answer is usually that you're doing two things at once - and the map has still told you which two.
A map, not a ranking
Here's the part I most wanted the woman on my call to hear. Any three-part list begs to be read as a ladder, with chat users at the bottom and automation wizards at the top. Resist that reading.
Services suit a narrow kind of work: high-volume, well-bounded, the kind you'd trust to run overnight. Most work doesn't fit that description and never will. Plenty of valuable AI use belongs at product and tool permanently, and a business that runs everything through a chat window can still be using AI brilliantly.
She wasn't at the bottom of anything. She was standing at product, looking at one specific piece of work - screening tender emails - and asking whether it was worth moving to running it. That's the right question, asked from the right place.
Crossing a boundary
So what does the map buy you? Two things: knowing where you stand, and knowing what you take on when you move. Each step from using to directing to running hands you decisions the provider used to make on your behalf. What is the system allowed to do - draft for your review, or act for you (Produce vs Act)? What may it start without you? And where does it run - whose ground does the machinery sit on?
That last question turns out to matter more than any of the others, especially if you handle data you're not free to send anywhere. It gets its own piece: Where Does It Run? For now, one question is enough. Look at any AI in your business and ask: are we using this, directing it, or running it - and did we choose that on purpose?
The one-page card version, built so you can hand it to your own AI, is here: The Three Ways of Working.
Hit reply and tell me where your work sits on the map. I read everything.
-- Alastair