In April I wrote about Conway’s Law and the pitch pause. The short version: organizations build things that mirror how they communicate. Before the pitch pause, half our company thought in AI-native terms and half didn’t, and that gap was going to show up in everything we made. So we spent two weeks raising the level of the conversation. Everyone got hands-on time with the tools, and we came out the other side able to talk about what was possible in the same language.
That worked. It’s fall now, and the tools are everywhere. We were acquired in June, so I work for ChartSpan now. That’s a side note here, but it’s why you’ll see me say ChartSpan where the earlier posts said Validic. The real change is that agentic AI isn’t something one part of the company does. Sales, support, product, finance, leadership. Everyone scans, reads, summarizes, drafts, and creates with AI as the default.
I think that’s a net positive and I want to say so clearly before I spend the rest of this post poking at it. People are doing things they couldn’t do a year ago. I’m not going back.
But I’ve been sitting with the Conway’s Law idea again, and I don’t love where it leads.
We talk through the artifacts now
In February, the problem was that we needed to elevate our conversations so we could understand the tools well enough to talk about what to build.
In November, a growing share of our conversations happen through what the tools produce.
Here’s the pattern. Someone sends me a document. A proposal, a call summary, a plan. It’s clean and well organized and has a nice little table. It’s also pretty obvious it came out of an AI tool with a light pass of review. It’s not slop. The content is usually reasonable. But it’s shaped for a generic reader, not for me. Long where it should be short. Thorough on the things I don’t need and thin on the one thing I do.
So I paste it into Claude. “Summarize this.” “What’s the gist of what they’re getting at?” And I read the summary instead of the doc. Which means a model wrote most of it, and a model read most of it, and the two humans mostly handled delivery. If the person on the other end does the same thing with whatever I send back, and I’d bet they often do, the loop closes.
A model wrote it. A model read it. We carried the messages.
It’s not laziness
It would be easy to call this lazy. It mostly isn’t.
The real reason is volume. These tools let you produce far more than you could before, and your capacity to carefully review it didn’t grow at all. So you make a tradeoff that sounds reasonable in the moment. Skim it, looks right, send. If you gave everything the review it deserved, you’d lose most of the speed that made the tool worth using.
I wrote about a version of this in February, in the post about work looping instead of flowing. The trap there was becoming a human approve button on your own work. What I’m seeing now is that trap at the scale of an organization. Not one person rubber-stamping their own output. A chain of people rubber-stamping each other’s.
The cost of producing something dropped to almost nothing. The cost of actually reading it didn’t move.
So what does Conway say?
Here’s where I keep landing.
If the things we build mirror how we communicate, and more of how we communicate runs through models, then what we build starts to mirror the model.
And models have a shape. They produce the most likely next thing. That’s what makes them so useful, and it’s also what makes them pull toward the middle. The familiar framing. The standard structure. The answer that is reasonable and broadly correct and has been said many times before.
So I think two things happen at once.
The floor goes up. Fewer bad documents. Fewer sloppy proposals. Fewer obvious mistakes. The worst version of our work gets better, and that’s real. It’s part of why I call this a net positive.
The ceiling comes down. The strange idea, the strong opinion, the “I think we’re wrong about this and here’s why” gets a little smoother every time it passes through a model. Not deleted. Just rounded off. Summarized into something more agreeable than it started.
That applies to software, but it applies even more to everything around software. The decisions. The strategy docs. The conversations that actually move a company forward. Those are the things that are most likely to be passed back and forth as AI-generated artifacts now, and they’re the things where the ceiling matters most.
Here’s the part that bugs me. The pitch pause gave us a shared vocabulary, and that was the whole point. Now we’re getting another shared vocabulary, by accident, and it isn’t ours. It’s the model’s. If you’ve built a few websites with Claude, you already know what I mean. You can spot other sites built with Claude. There’s a look to them. The same layout instincts, the same spacing, the same visual rhythm. None of it is bad. It’s just recognizable. The same thing is happening to our documents. Same structure. Same headers. Same three bullet points. Everyone starts to sound alike. That’s not shared understanding. It’s just convergence. This paragraph, written by Claude, has exactly the feel I’m talking about.
Now zoom out
Run the same logic across an industry.
Most companies are routing more and more of their communication through the same handful of models. If Conway holds, those companies start to build similar things. Not because anyone copied anyone. Because the strategy doc at your company and the strategy doc at your competitor were both shaped by the same model, pulling toward the same middle.
Differentiation doesn’t get destroyed in a single meeting. It erodes, one reasonable-sounding document at a time.
This isn’t just a hunch. A couple of researchers, Anil Doshi and Oliver Hauser, ran an experiment on exactly this back in 2024. They had people write short stories, some on their own and some with story ideas from GPT-4. Then they had other people rate the stories.
The AI-assisted stories were rated more creative, better written, and more enjoyable. The biggest jump was for the writers who were the least creative on their own. Floor up.
But the AI-assisted stories were also more similar to each other. Everyone got individually better, and the group as a whole produced a narrower range of ideas. The researchers called it a social dilemma. Each person is better off using the tool. Everyone together ends up with less that’s actually new. Ceiling down.
It’s one study, about short stories, with a model that’s old news now. I’m not going to pretend it proves anything about how a company makes decisions in 2026. But it’s the clearest picture I’ve found of what I’m worried about. Every individual doc gets better. The collection of all our docs gets more alike. And the things at the edges, the odd observation from the support lead, the dissenting opinion in the planning meeting, the customer quote that doesn’t fit the story we’ve been telling ourselves, are the first things to get smoothed away.
What we actually lose
The first thing is the thinking. Half the time I don’t know what I actually think about something until I sit down and try to write it. That’s usually where I find out my idea has a hole in it, or that I care about it more than I realized. If Claude writes the first draft and I just nod along, I can skip that part entirely and never notice I skipped it.
The second thing is harder to see. When someone writes a doc themselves, it has their fingerprints on it. You can tell which part they’re unsure about because they hedge. You can tell which part they really care about because they spend three paragraphs on it. That stuff is often the most useful information in the whole document, and it’s exactly what gets stripped out when it’s generated, and again when the reader summarizes it.
Then there’s the new stuff. Models are great at taking an idea and making it bigger and cleaner. But the idea itself usually came from somewhere. A customer call. Something someone noticed in the support queue. A thought that comes to you during a lunch break. If what goes into the model is just another model’s output, nothing new is coming in. We’re rewording the same thing back and forth and calling it collaboration.
And honestly, some of it is just the relationship. I trust the people I’ve argued with and worked things out with. I’m not going to build that with your agent.
The meeting is next
Here’s my prediction. It might be a little hyperbolic, but I don’t think it’s outlandish.
In the next six to twelve months, someone is going to productize sending an agent to a meeting in your place. Not a notetaker. A proxy. An agent with your documents, your history, your positions, that joins the Zoom or Google Meet and participates as you.
And then the obvious thing happens. If you’re sending your agent, why should I show up? I’ll send mine. Now the meeting is four agents and one confused human, and the document problem has moved into the room. Everyone gets the summary afterward. Most of them will have their AI summarize it.
Meetings are one of the last places where unfiltered human input still reliably enters the system. That’s the part I’d least want to automate away.
There’s also a quieter version of this already happening. When agents are iterating on a document or a problem, there’s a pull to step back and wait for the good version to show up. Sometimes that’s fine. But sometimes I’m on the sidelines of my own work, watching two models trade revisions on something that a ten-minute conversation between two people would have settled. The models will happily keep going forever. They never get tired. That’s not a feature when the real blocker is that two humans haven’t talked.
Where I’ve landed so far
I don’t want to leave this as all problem and no answer, so here’s where I actually am with it. It’s not much.
The one thing I’ve genuinely changed is that when I’m asking someone for a decision, I write the top of the doc myself. Just a few lines. What I need from you, what I’ve already decided, what I’m still not sure about. Claude can write the rest, but those lines are mine, because they’re the part nobody else could have written.
On the reading side, I’ve been experimenting. When a doc shows up that clearly came out of a model, I’ll ask Claude what the author was actually trying to say, not just what the document says. And when I ask it to summarize something, I tell it not to drop the details. Keep the hedges. Keep the numbers as the person wrote them. Keep the one-off customer comment that doesn’t fit the pattern. Those are small prompts, but they push back on the smoothing I described above, at least on my end.
There are a couple of things I’d like to start doing and haven’t yet. Before sending something, asking myself whether I could explain it in two minutes without opening it. If I can’t, I didn’t review it. I forwarded it. And telling people what I actually reviewed. Something like “I went through the first two sections carefully, the rest is Claude’s first pass.” Ten seconds, and the reader knows where to spend their attention.
But even if I did all of that, they’re habits. Habits don’t change a communication structure. They’d just make me a slightly better link in the same chain.
The bigger thing I keep coming back to is this. If everyone has the same tools, the document isn’t really the valuable part anymore. Any of us can generate one. What none of us can generate is what the other person put into theirs. Right now all of that is sitting in private AI sessions that nobody else can see. That’s the second half of this, and it’s a post of its own: What If a Document Worked Like a Repo?
In April I wrote that you can’t design for a world you don’t share. We do share it now. The question is whether we’re still the ones doing the talking.