Why Your Principles Don't Survive Writing Down

I wrote the best brief of my career this spring. It took about two hours and it was addressed to a machine. Crucially, not one line of it had to survive a disagreement.

That last part took me longer to notice than it should have.

The brief was ordinary enough in its parts. Context: here is the situation, here is what has been tried, here is why the obvious answer is wrong. Principles: this is what good looks like here, here is what we never do and where the line to hold is even when it hurts. Then the operating instructions. When to stop and check in before going further.

I read it back and thought: I have never written this for a person.

Not in twenty-five years. Not for teams I trusted and rated and would have gone to the wall for.

The comfortable explanation arrived straight away. It goes like this: I had been carrying all of it in my head, my people deserved it written down, and the machine had finally forced me to do the work I should have done years ago. Pay the debt, give the team the brief.

That reading is flattering, it is half true, and I now think it is the less interesting half.

The reason it was easy

I could not have written that brief for my team. Not did not. Could not.

Put those same principles in front of six people who know the work and the meeting takes a day and a half. Someone says the third principle contradicts something we agreed in March. Someone else wants to know what happens when the line I say we hold costs us a launch. A third points out, correctly, that the thing I described as what we never do is something we did twice last year for good reasons.

I would have defended some of it and to be fair, I would have lost some of it. What came out the other end would have been harder to write down, yet much better.

The machine took the whole thing. It thanked me for the clarity.

So the completeness of that brief is not evidence that I had the good version all along. It is evidence that I was writing for a reader with no standing to object.

The tell

There is a line in my own brief I skipped past a dozen times before I saw what it was doing.

I told it where to push back on me.

I had to ask.

Nobody asks a good team to push back. Disagreement is the default condition of working with capable people, and most of a leader's energy goes on holding it in some useful shape rather than summoning it in the first place. With a machine it is a feature you remember to switch on, in the sections where you happen to suspect you might be wrong.

Which is exactly where you are not wrong. You already know where you are shaky. That is not where the damage lives.

The damage lives in the places you were so certain about that it never occurred to you to invite an argument.

What judgement is made of

Here is the thing I think the current conversation keeps stepping over.

Judgement is not the set of conclusions you hold. It is the record of what survived being contested.

That distinction sounds academic until you try to move judgement from one place to another, which is what every company is now attempting to do at speed. You can move conclusions easily. Why? Because they write down beautifully. What does not travel is the thing that made them trustworthy. Every time they were tested by someone who wanted a different answer and lost is what made them trustworthy.

Strip that out and you still have the sentences. They look the same on the page.

You can see the difference in how the two behave under pressure. A conclusion that has been argued with knows its own edges. The people holding it can tell you where it stops applying, what it costs, which cases nearly broke it, and what they would need to see to abandon it. That is not extra detail, it IS the judgement. A conclusion that has never been argued with has none of that, and it does not know it is missing anything.

Every experienced leader can name someone in their company who holds strong views they have never had to defend. They know the hard-won reasons behind the conclusions everyone has to apply. We are now building that person into the infrastructure and calling it institutional knowledge.

Three decimal places

Last month I wrote about a rule I worked under for most of my career. Connection tolerances held to three decimal places, taught to everyone who touched the product, old enough to outlast everyone applying it, and strong enough that good ideas lost arguments to it.

I described it then as a rule with teeth and left it there. Here is the part I did not say.

That rule did not have authority because it was written down. Plenty of things were written down. It had authority because it kept being argued with. Every year somebody arrived with a genuinely good reason to bend it: a cost saving, a new material, a design that would have been beautiful. The rule won those arguments on the merits, repeatedly, in front of people who badly wanted the answer to go the other way.

Decades of winning arguments is what a rule with teeth is actually made of. The three decimal places were only the notation.

Now load that same rule into a system on day one. Every decision downstream complies with it immediately. No cost saving ever gets to test it. Nobody has to defend it against a beautiful design. It reads identically and it is a completely different thing: a conclusion wearing the authority of something that earned it.

What the industry is building

This spring the World Federation of Advertisers (WFA) and BCG put a joint report in front of a room of global CMOs. Everyone is using AI yet very few are getting compounding value, and around seventy percent of the effort separating the leaders from the rest is people and change rather than technology. Their proposed fix is what they call an enterprise context layer. Encode the company's knowledge, principles and guardrails into the systems themselves, so that every AI-assisted piece of work draws on the institution's judgement by default.

Codify the thinking. Wire it in. Make judgement a property of the infrastructure instead of the people.

I understand the appeal, and I would rather a company did this than nothing, because writing your principles down well enough for a machine to use them is real work and most companies have never done it.

But look at what you are encoding.

You are not encoding the institution's judgement. You are encoding the version of it that was current on the day somebody sat down to write it out, produced under exactly the conditions I have just described: no room in the process for the six people who would have taken it apart. The context layer inherits the confidence of the original without inheriting the arguments that earned it.

Then it does something worse. It makes the thing unarguable going forward. A principle held by people can be tested on Tuesday by somebody with a good reason. A principle wired into the system that produces the work is not encountered as a claim at all. It arrives as the shape of the output. There is nothing there to disagree with.

The usual objection to codifying judgement is that it cannot notice when the world has moved. That objection is fair and it is too kind. The deeper problem is that it was never as sound as it looked, and now nothing will ever find out.

Why any sensible person would do it anyway

I want to be fair to the move, because I have made it myself and I will probably make it again.

Argument is expensive. It takes the time of your most capable people, which is the most costly thing you have. It is uneven, so the same question gets a different quality of challenge depending on who is in the room that week. It is slow at exactly the moments when speed feels like survival. And it produces nothing you can show a board. There is no artefact at the end of a good disagreement, only a decision that is quietly better than it would have been and no way to prove that to anyone who was not there.

Encoded judgement has none of those problems. It is fast, even, it applies at three in the morning, and it audits beautifully. You can show it to a regulator. You can hand it to a new starter on day one and they will produce compliant work by the afternoon.

So the choice companies are making is not stupid, and telling them it is will not get you anywhere. They are choosing something legible and cheap over something illegible and expensive, which is a choice organisations make correctly most of the time.

It is the wrong trade here for one reason. In almost every other case where you swap a slow human process for a fast encoded one, the thing you encoded stays true while you use it. A tolerance is a tolerance. A payment rule is a payment rule. Judgement is the one input that degrades precisely because you stopped arguing about it, and it degrades invisibly, and the system built on it goes on producing confident output the whole time.

You will not get a warning. That is the property that makes this different, and it is why the trade that looks prudent on the day looks reckless in the third year.

Every one of them agrees with me

I have spent the last few weeks building a set of agents to take over parts of my own working week. It has been the most interesting thing I have done in a while, and it has also been a slow lesson in the argument above.

Every one of them agrees with me.

Of course they do. I wrote them. They run on principles I set, in a context I supplied, with a tone I chose. When one of them produces something wrong, it produces it confidently and in my own register, which turns out to be an effective disguise. Twice now I have caught something on the second read that I would have caught instantly if a colleague had said it out loud in a meeting, because when a person says something slightly off, you hear it.

I am one person with a handful of agents. Scale that to a company where every team has a fleet, all of them drawing on the same encoded principles, and you do not have more capacity to think. You have one opinion, held many times, at speed.

Building the argument back in

So what do you do instead, if writing it all down is not the answer?

You build the contest rather than the codex.

The clearest example I know sits in Oslo. Norway's sovereign wealth fund has pushed AI into its work as hard as any institution I have watched. Alongside the automation they built something most companies would not think to build: a simulator in which their investment professionals make calls, see the consequences, and get structured feedback on the quality of their reasoning. A machine for being wrong in front of evidence, on purpose, before it costs anything.

Read it against everything above and notice what it actually is. It is not a system that holds the institution's judgement. It is a system that keeps putting the institution's judgement back into the argument it came from.

That is the design choice, and it is available to any company willing to make it. Ask which of your principles has been genuinely tested in the last two years, by someone with standing and a real reason to want a different answer. Ask what happened. If you cannot name the occasion, you do not have a principle. You have a habit that has not been challenged yet, and you are about to wire it into everything.

The practical version is smaller than it sounds. Before anything goes into the system that produces your work, make it survive a room. A room where someone is asked to take it apart and is thanked for succeeding.

What I still do not know

I have been careful in this piece to argue from things I have seen. Here is the part I have not resolved.

The systems I built will outlast me. I made sure of that, and it is the professional fact I am proudest of. They run in places I have never visited, operated by people I have never met, and they have not needed me for years.

What I have never fully tested is the judgement underneath them. That was never written down anywhere, for the reason this whole piece has been circling: it did not need to be, because it was alive in argument every day, and I was in the room for the argument.

I am not in the room forever. Nobody is. And there is only one way to find out whether what you built can keep winning the arguments without you, which is to stop being there and see.

It is not a document. It was never going to be a document.

Next month, a more personal piece. There is something I have been carrying through every one of these essays, and it is time to say it plainly.

Sources

Evidence

• WFA + BCG, AI Community — Global Marketer Week 2026 (Stockholm, April 2026): 100% of marketers using AI, 16% at advanced stages; ~70% of the effort separating leaders is people and change; the "Enterprise Context Layer" proposal engaged in the body. Link
• Norges Bank Investment Management, How we use AI in practice (AI Summit, 2026): the Investment Simulator as deliberately engineered judgement-building alongside aggressive automation. Link

Building on

• Michael J. Mauboussin, The Wisdom of Crowds (Morgan Stanley, Consilient Observer, 2026) — the conditions under which a group aggregates information rather than amplifying one error: diversity and independence. Remove either and the crowd stops working. The formal statement of what the agent-fleet passage argues informally. Link
• Jennifer Garvey Berger, "Disagree to Develop" (2026) — the move from disagreement as something a decision has to survive to disagreement as the thing that builds the people and the idea at once. Link
• Paul Willmott, Considered Machines (Substack, 2026) — the organisational rather than tooling read of AI that this piece extends. Link