Not quite right

I find myself saying "not quite right" far more often than "looks good to me" when I'm collaborating with AI.

LGTM (looks good to me) has become a catch cry for a lot of people using AI. It's what you type when something is close enough, and going round again on the subtle bits feels tedious. Usually when you're short on time, and back-to-back meetings leave no room to really do the work. A lot of AI-assisted work now runs on LGTM. NQR, not quite right, is slower. It's also where the thinking happens.

AI slop often turns up as derivative and formulaic writing that fills our inboxes and documents. It's easy to complain about. But that kind of slop is only a symptom. Better writing can't fix anything if we're skipping the real job, which is critical thinking. That means doing the messy work of thinking it through with others — including AI — and doing it properly. After that hard work, a short document that gets everyone aligned is trivial.

“Better writing can't fix anything if we're skipping the real job, which is critical thinking.

The fix isn't a procedure or a policy on AI use. It's culture. It's how a leadership team thinks together before anyone writes anything down.

Writing used to prove the thinking

I worked inside Amazon's writing culture, and I liked it. Meetings started in silence. Everyone read the six-page memo for the first half hour, pen in hand, and when the talking started, everyone started from the same place. Jeff Bezos put the logic simply: "There is no way to write a six-page, narratively structured memo and not have clear thinking."

The document was proof that the author had done the thinking before asking for an hour of everybody's time.

Steve Huynh spent almost twenty years at Amazon. He quotes Bezos's line, then adds: "That last sentence is no longer true. With AI, you can trivially produce six pages, no critical thinking required."¹ One senior engineer he knows now tags documents "AI;DR". AI, didn't read. If a model wrote it, a model can summarise it.

I'm seeing a version of this in every company I work with, written and spoken. Sometimes the presenter hedges with an apology: "this is what my AI came up with in the 15 minutes I had before the meeting." Other times I see furrowed brows, as the room sits through an awkward retelling of unfamiliar ideas that aren't really the presenter's own. Or frustrated stakeholders, annoyed that the simplest questions have no answers.

Skipped thinking doesn't disappear. It moves to the reader. In most companies the reader is the person who has to decide, so the work piles up on the same few people. They're the ones wading through the slop, trying to work out whether there's a real idea in there, and where the risk is.

“Skipped thinking doesn't disappear. It moves to the reader.

Do the thinking

In my 2025 piece I said there's nowhere to hide in a document. That's still true, but think about why. A good document needed the thinking to happen first, and again in every revision that follows. At Amazon, and anywhere with a written culture, the writing forced the work.

So the point isn't the document. It's to do the thinking. Talk to customers. Get out of the building. Explore the situation, and understand the problem and the people who have it. AI has tools that make that work better and faster. It doesn't replace it.

That work takes time, and it's where "just give it to AI" does the damage. It assumes the understanding is as free and instant as the document at the end. It isn't. I've caught myself thinking I must be doing it wrong when I'm 15 revisions deep, and changing what I think (again) as I look at the opportunity through yet another lens. The time goes on the conversations and the second looks. The sentences are quick.

Getting back in the room is how you do it together. A contest of ideas, in real time, with people who see the problem differently and hold different evidence. Think together, progress together. AI can be one of those thinkers. Some of the most useful AI work I do is when it asks me hard questions instead of handing me answers.

“Some of the most useful AI work I do is when it asks me hard questions instead of handing me answers.

Then go and test the idea. That doesn't mean building it. Whether you can build something is the risk that kills ideas least often. Whether anyone wants it, and whether it's worth doing, are much bigger killers. Find the riskiest assumptions and test them, cheaply, before anyone writes a business case, or builds anything. Even a prototype.

Thinking together, without workshop theatre

In 2025 I wrote that workshops favour "charismatic extroverts, HiPPOs (Highest Paid Person's Opinion), and quick thinkers", and argued for a written culture instead. I've changed my mind, at least in part.

The problem wasn't the room. It was the theatre: sticky notes, high energy, and an outcome nobody could remember a week later. Amazon's study hall was a room too. People read first, then argued. The writing got everyone to the starting line, and the contest in the room did the rest.

“The problem wasn't the room. It was the theatre.

What makes a room a contest instead of theatre is culture.

Huynh tells of a review meeting where a senior product manager was obviously reading his own six-pager for the first time. Everyone in the room could see it. Nobody said a word, because he was the most senior person there.

A writing policy wouldn't have saved that meeting. A different culture could have. In the teams I've learned most in, the facts and the rationale settled the argument, regardless of who liked or disliked the idea. Amazon writes this into its leadership principles². Leaders "seek diverse perspectives and work to disconfirm their beliefs", and they're "obligated to respectfully challenge decisions when they disagree, even when doing so is uncomfortable or exhausting."

At ThoughtWorks, "strong opinions, weakly held" was everywhere. The phrase is Paul Saffo's³, but we lived it. You argue hard for what you think, and you change your mind when the evidence says so. Just like I'm now changing my mind about what I wrote with strong conviction in 2025.

That's how trust gets built. Colleagues need to see proof of thought, wherever it happened. If someone wants to do their thinking in a document, fine. If they'd rather do it in the room, just as fine. Trust goes when the thinking is skipped, or handed to AI with nobody checking the work. When evidence wins the argument, people stop guessing what the boss wants and start bringing what they've found.

“When evidence wins the argument, people stop guessing what the boss wants and start bringing what they've found.

Making room for the thinking

None of this needs a workshop. These are habits a leader builds into how they run things.

• Ask what's been tested, not for the proposal. "What did we try, and what did we learn?" becomes the price of entry to a decision.
• Fund tests, not projects. Every small bet comes with the question it's meant to answer.
• Thank the person who kills their own idea with evidence. Do it in front of others. It tells everyone what you value.
• Speak last. If your view enters the room first, the contest is over before it starts.
• Sort decisions by whether you can undo them. Most can be made quickly and fixed later. Save the room — and your own attention — for the ones you can't take back.

Why not let the AI decide?

One objection is that the models reason now, so why not let them do the thinking? Because business isn't only logic. It's creative and it requires good judgement in conditions of uncertainty. Besides, a leader can't hand off the accountability for a decision, especially when AI did the reasoning behind it.

“A leader can't hand off the accountability for a decision, especially when AI did the reasoning behind it.

And yet, in the right conditions, you should let it decide. Decisions you can undo are the obvious place to start. Bigger ones can follow, once you've done the hard work of building trust into your judgement infrastructure: who checks what, what an agent is allowed to decide on its own, and how you'd know when it gets something wrong. That's the work we call Agentic Operations. You earn your way there by doing the work.

Write it down afterwards

Writing still matters. It just comes later. Once the thinking is done, or at least properly started, write down what you understood, so the people who weren't in the room can act on it. It'll come out short, because the thinking has already been done.

Blaise Pascal knew this in 1656, apologising in one of his letters: "I have made this letter longer only because I have not had the time to make it shorter."⁴ A short, clear document takes time, spent on thinking. So make time for it.

“A short, clear document takes time, spent on thinking. So make time for it.

It's also why I've started putting a colophon on what I share. It tells the reader how a piece was made, how much time and thinking went into it, who's accountable for it, and how much attention it deserves. Usually it says I've worked on this enough for you to spend time reading it. Sometimes it says the contents are machine-generated and limited thinking time has been invested, so treat it accordingly. Both have their place.

The craft that matters now is thinking. Writing is how you hand it on.

Notes

¹ Steve Huynh, "On the death of Amazon's writing culture", A Life Engineered

² Amazon Leadership Principles: "Are Right, A Lot" and "Have Backbone; Disagree and Commit"

³ Paul Saffo, "Betting on 'Strong Opinions Weakly Held'", EDUCAUSE Review interview, 1998

⁴ Blaise Pascal, Lettres Provinciales, letter 16 (1656). On the line's history, and why it's often credited to Twain