Automate last

Every founder I talk to is circling the same question: how do we use AI to transform the business? And the internet is happy to answer. Agents this week, AI OS the next, loops after that. The terms change fast enough that you can never feel caught up, which is the point. Content that makes you feel behind gets shared. Content that tells you to slow down doesn't.

When founders ask me where to start, I give them an answer that has nothing to do with AI. It's Musk's five-step algorithm, the one he beat into Tesla and SpaceX after automation nearly killed the Model 3 ramp. He over-automated the factory, then spent months pulling robots back out. "Humans are underrated," he said afterward. The algorithm:

1. Question every requirement.
2. Delete the part or process.
3. Simplify.
4. Accelerate.
5. Automate.

He attaches a corollary worth framing on a wall: the most common error of a smart engineer is to optimise a thing that should not exist.

The list looks banal until you notice that an entire industry is currently selling step five, and only step five. A company gets excited, brings in an agency on a three-month project to put AI into the sales motion, and points it at a complex legacy flow nobody has questioned in years. The demo is impressive. Six months later, ask what actually changed.

Why does starting at step five fail? Think of a business as a program. Not metaphorically. A company is a set of procedures for turning inputs into money, and almost none of it is written down. It runs distributed across the heads of the people who work there. The pricing logic is three people's intuition. The quality bar is whatever the most experienced person in the room refuses to let through. An AI agent needs a spec: rules, context, definitions of done. Drop an agent into a business and you are asking it to execute source code that was never written. The vendor solves this by transcribing whatever the team says the process is, the what and the how, recorded faithfully, with all the accumulated nonsense intact. That's documentation without deletion. The company's scar tissue, automated. The mess doesn't go away. It compiles.

And here's the uncomfortable truth about that unwritten program: most of it was never designed. It accreted. The processes in most businesses are not the work of someone who deeply understood the problem and chose the cleanest answer. They're whatever got the job done at the time, made by people who didn't know what they didn't know, then inherited by people who never asked why. Doing it well takes a rare combination: people who really know their craft, who dig all the way down to the why, and who then have the discipline to pick the simplest solution instead of the familiar one. Most businesses never had that combination in the room when their processes formed. In software we'd call this unreviewed legacy code. And business legacy code is worse than the software kind, because a business has no git. Nothing can be safely reverted, so nothing ever gets deleted, and the program only grows.

Which is why the real work of automation is deletion, and why deletion almost never happens. The only honest test of whether a process is load-bearing is to remove it and see what breaks. Musk's heuristic: if you aren't occasionally adding things back, you aren't deleting enough. So you delete until something breaks. When it breaks, good. You've found a real requirement, the first one you can actually trust. Fix it fast and keep going. It's build-measure-learn, pointed at subtraction.

Notice what this requires: you have to be there when it breaks. Deletion is only safe for someone who sticks around. This is why no vendor on a six-week contract has ever deleted anything. The rational move, when you're leaving, is to add. Additions demo well now and break later, on someone else's watch. Deletions break now, on yours. The incentive gradient of the entire agency model points away from the one activity that creates the most value.

The Japanese have been making this argument about food for centuries. The philosophy of the cuisine is subtraction: strip away the sauce, the seasoning, everything external, until what remains is the true flavour of the ingredient, pushed to an extreme. It looks like the simplest food in the world. It is the hardest to make. Simplicity is the highest level of complexity, not the absence of work but the residue of it. A simple business is the same thing. Not a primitive one. One where someone did the brutal work of removing everything that didn't need to exist.

Simplicity is also what unlocks everything the noise promises. A simple process is one you can write down. Once it's written down, AI can read it, answer questions about it, and eventually run it. Simple, then documented, then automated. That's the whole path, and the hype skips the first two steps because they're slow, invisible, and don't demo.

Could AI do the questioning and deleting itself? Eventually, maybe. Not soon. The inputs it needs don't exist in any dataset: the requirements live in heads, the process was never written down, and the load-bearing nuance is things like who actually does the work versus who is supposed to. Even if a model handed you a perfect strategy, nothing would change until people worked differently on Monday. A good strategy executed this week beats a perfect one that is still a document next week. Execution is presence, and presence is the one thing you can't download.

The strongest evidence comes from the AI labs themselves. The hottest job in AI right now is the forward-deployed engineer, Palantir's old playbook, postings up 800% in a year. In May, OpenAI launched an entire deployment company with billions behind it. Anthropic announced a joint venture days later. Read that again: the people with the best models on earth are spending billions to put humans inside businesses. They are telling you, with their capital, where the bottleneck is. It isn't the model. It's the part where someone understands the business.

So if you're serious about transformation, the shape of the work follows. Someone embedded, in-house or outside, it doesn't matter, as long as they stay long enough to do the first four steps properly before plugging in the fifth. And sometimes the honest answer is that the legacy flow isn't worth renovating. Then you build the clean version beside the business, a new system, even a new brand, and migrate onto it slowly while the old one keeps the lights on. It looks slower on paper. It's far more likely to still be working in five years.

None of this is an argument against AI. I build with it every day. Work that used to take a quarter now takes weeks. But that's exactly why the order matters more now, not less. When a step gets cheap, the value doesn't disappear. It moves upstream, to the steps that are still expensive. Automation is becoming free. Knowing what to automate is not.

Two woodcutters. One chops all day without a break. The other keeps stopping to sharpen his axe, and goes home with more wood. The first four steps are the sharpening. The agent is just the chopping.

Automate last. Not because automation doesn't matter, but because it's the only step that gets easier every month. The other four are still on you.

---

Weekly notes from conversations with my business partner, where we share what we're seeing on the ground across clients and the market - not the hype, the real shifts.