[Chapter 6 from Deliberate Alignment by Walter Reid, written in spring 2026]
AI does not fix unclear thinking. It amplifies it.
Somewhere in your organization, right now, thirty conversations are happening in parallel.
Each person, with their AI tool open in a separate tab, is asking a version of the same question the organization is supposedly aligned on. Each is receiving an answer shaped by the specific way they framed the question, the context they brought to it, the assumptions they did not know they were making. Each answer is locally coherent. Each one makes sense given the conversation that produced it.
Collectively, they are pointing in slightly different directions.
This has always been true. Organizations have always been collections of parallel, partially overlapping mental models of what the organization is doing and why. The watercooler conversation that produced a different understanding than the all-hands. The email thread that resolved something the official process had not gotten around to resolving. The tribal knowledge that lived in the heads of the people who had been there longest and evaporated when they left.
What is new is the speed at which each conversation now produces artifacts.
Before, unclear thinking moved slowly. A misaligned understanding of what the product was for would take weeks to materialize as a feature nobody wanted. There were checkpoints — reviews, handoffs, the moment where someone would say, wait, that is not what we meant. The friction was expensive in time and morale, but it was also a correction mechanism. Slowness caught errors that speed would have shipped.
Now the artifact follows the conversation almost immediately. The unclear thinking is instantiated — made real, built, sometimes shipped — before anyone has had a chance to ask whether it was right.
In the Zansei project, this dynamic became visible on a specific Thursday. One person had spent the week building a research report — seventeen pages analyzing the site’s copy, with the second half promised for the following week. Another person had been running ads at a thousand dollars a week. Zero people had completed the funnel in the previous seven days. A third person had been exploring a new feature direction that the founder had mentioned in passing on Monday. Each of these efforts was internally coherent. Each person could explain what they were doing and why. Together, they were spending money to drive traffic to copy that was being researched rather than fixed, while the product’s scope expanded in a direction nobody had agreed on. The backlog was no longer a prioritization tool. It was an anxiety archive accumulating in three different directions simultaneously.
That is the thousand conversations problem. And it is not a coordination problem.
It is a meaning problem wearing coordination’s clothes.
Why the Distinction Matters
Coordination problems have known solutions. You establish a process, a single source of truth, a set of handoffs with clear ownership. You standardize the input and the output converges. Organizational design has been solving coordination problems for a century.
Meaning problems are different. They are not about who does what or when. They are about whether the people doing the work share a sufficient understanding of why — of what the work is actually for, what success would feel like from the outside, what they would sacrifice and what they would not.
Meaning cannot be standardized. You can issue a directive that creates the appearance of resolution while the actual divergence continues underneath, now invisible because everyone has learned the right vocabulary.
The reason AI amplifies this specific problem is that the tools are exceptionally good at producing outputs that feel coherent. A person with an unclear understanding of the product vision, working with an AI assistant, will produce work that is internally consistent, well-structured, and legible — and pointed in the wrong direction. The professional surface has never been easier to achieve. The gap between professional surface and genuine alignment has never been harder to detect.
The Zansei researcher’s seventeen-page report was well-structured. It had sections and findings and recommendations. It was also answering a question nobody had agreed on, produced by someone who had no way to know the question was wrong because the question was never made explicit. The report was a coordination success — the researcher coordinated with the tools beautifully — and a meaning failure. The meaning of what the research was supposed to accomplish had never been established clearly enough to guide it.
Another team member was building something genuinely impressive in parallel — a video generation tool using Runway, Veo3, and a custom renderer. Marketing briefs that turned into thirty-second videos. High production quality. The kind of output that would have required a studio and a team five years ago. But he was working in a silo, disconnected from clients and product direction, solving a problem no one had validated. He had heard people mention wanting video content and started building. His lived experience on real production sets made him feel powerful with the AI tools, and the power was real — but it was not connected to the room.
The last ten percent is ninety percent of the effort, and while the work felt productive at first, the problem was that his solution was looking for a problem. He was not a developer or a product person or close to the client. He was building things he had heard people needed but with no validation, because there was no structure connecting his capability to the team’s direction. The thousand conversations problem embodied in one person: locally brilliant, collectively disconnected.
The Temptation of Standardization
The instinctive organizational response is standardization. One prompt template. One AI tool. One system of record that everyone queries. One voice that speaks for the organization’s intent so that everyone’s parallel conversations start from the same place.
This is not wrong. It is incomplete. And the part it misses is the part that matters.
The parallel conversations are not a problem to be eliminated. They are a resource to be synthesized. Each person who approaches a question from a different angle — the one who treats nuance as a feature, the analytical mind, the contrarian who sees what consensus is missing — each of these people is producing signal. Signal that a standard template would suppress before it could be captured.
Standardize the input and you get consistent output. You also get sterile output. An organization that has optimized its parallel conversations for convergence has also optimized them for the elimination of the unexpected. And in a world where judgment is the scarce resource, the unexpected is where the value lives.
The solution to the thousand conversations is not to make them all the same conversation. It is to build the infrastructure that lets them remain different while still being synthesizable.
What That Infrastructure Looks Like
The OpenAI harness engineering team — introduced in Chapter One — encoded their direction in a repository document that any agent — human or AI — could read before beginning work. Not documentation for its own sake. A living encoding of direction, specific enough that every agent working from it would produce consistent results without continuous re-alignment. Their early approach — one large instruction file containing everything the agent needed to know — failed. When everything is marked important, nothing is. They replaced it with a structured directory: a short entry point that served as a map, with pointers to deeper sources of truth. Design decisions documented not just as outcomes but as reasoning. Architecture described not just as what exists but as the principle behind what exists, so that the next decision made by a different person in a different context would be consistent without being identical. That is progressive disclosure applied to organizational intent — the same principle that applies to human teams.
The insight underneath that example applies to human teams with at least as much force as it applies to AI agents: anything that is not encoded explicitly enough to be read by someone who was not in the room effectively does not exist.
This is not metaphorical. During the RouteForge development described in Chapter Nine, I built a new client portal — designed it, coded it, integrated it. When I was finished, Claude Opus informed me that a client portal already existed in the codebase. It was just not documented and not connected to anything. Something that literally existed in the codebase did not functionally exist because no one had encoded it legibly enough for anyone — including the AI — to know it was there. The organizational legibility argument applies to code as much as it applies to teams. What is not encoded does not exist, even when it does.
The Slack thread that aligned the team on a decision but was never written down — it does not exist for the person who joined three months later. The whiteboard session that resolved the ambiguity in the product brief — it does not exist for the team member who was traveling that week. The institutional knowledge that lives in the head of the person who has been there longest — it does not exist for anyone the moment that person walks out the door.
Organizations have always known this. The practice of knowledge management is decades old. But the urgency was manageable when execution was slow, because the gap between losing the knowledge and feeling the consequence was wide enough to catch and correct.
When execution is fast, that gap closes. The undocumented decision becomes a shipped feature before anyone realized the decision had been made implicitly. The divergent understanding becomes a launched campaign before the synthesis that would have caught it has had a chance to happen.
What this requires in practice starts with presence. I have been in Click to Pay meetings with twenty people in person, and sometimes there is a legitimate dialog. But I don’t often hear from people on the conference line — not because they don’t have ideas, but because they aren’t present. Both literally and figuratively. They would need to shout on the phone just to make their presence felt, and if they had that type of will to be heard on the phone, they would have been in the room anyway. Presence is self-selecting. Remote participation in alignment work is observation with the appearance of inclusion. The infrastructure that makes the thousand conversations synthesizable starts with the people who are doing the synthesizing actually being in the room where the synthesis happens.
Beyond presence, the infrastructure requires legibility — not comprehensiveness. The instinct when asked to make something explicit is to make everything explicit — the complete knowledge base, the exhaustive style guide, the documentation that covers every edge case. That instinct fails for the same reason the OpenAI team’s single instruction file failed. When everything is marked important, nothing is. The file became a graveyard of rules that may or may not still reflect reality, and nobody could tell which was which.
The solution is structural: a short entry point that serves as a map, with pointers to deeper sources of truth. That is organizational legibility. Not transparency — everyone can see everything. Not documentation — there is a record of what was decided. Legibility: the organization’s intent encoded in a form that survives the people who hold it.
The Person Who Reads the Conversations
There is a person this chapter has been describing without naming.
In the previous era, they existed in the translation layers between roles. The senior developer who understood what the business actually needed. The product manager who held the tension between what users said they wanted and what the data showed they did. The account director who maintained a relationship sophisticated enough to know what the client had not yet said.
These people were synthesizing — reading the parallel conversations, finding the signal in the divergence, surfacing the question that everyone was dancing around without asking, and converting the resulting clarity into a direction the rest of the organization could move in.
As the translation layers dissolve — as the designer and the developer and the account executive converge toward a shared capability — this work does not disappear. It becomes more important, more visible, and more load-bearing. Because the parallel conversations are happening faster, the artifacts are materializing sooner, and the cost of letting the divergence compound is higher.
The skill is not technical. It is closer to curatorial, or editorial — something that has not quite found its vocabulary because it has only recently become urgent enough to require one. It requires the ability to hold contradictory inputs without prematurely resolving them. To distinguish between productive tension — divergence that is generating signal — and destructive divergence that is consuming energy without producing meaning. To know when the conversations have generated enough to act on and when they are still processing something important that premature synthesis would foreclose.
This is not a new kind of person. They have always been in your organization. What is new is that their work, which was previously invisible because it happened in the gaps between the legible processes, is now the work that everything else depends on.
In film production, the person who holds this function has a title: second assistant director. Not the creative lead. Not the one whose name goes on the poster. The one who holds the entire production’s logistics in her head — who knows that tomorrow’s call sheet conflicts with a location change the director mentioned in passing, who reads across every department’s needs simultaneously and surfaces the collision before it becomes a cost. The function is synthesis. The title undersells it. It always has.
The version of this person in your organization may be the one who is most skeptical of the AI tools — not because they don’t understand them, but because they understand what gets flattened. The people best suited for this work are not always the ones who embrace the tools that accelerate it.
Find them. Give them the room.
Or, if you are that person: you already know. The question is whether the organization around you knows yet.
Before Anyone Opens a Laptop
This is where the chapter connects to Deliberate Alignment as a practice rather than a philosophy.
The thousand conversations problem cannot be solved in the meeting where the work gets reviewed. It cannot be solved in the retrospective where the misalignment becomes visible. It cannot be solved in the integration moment where the separately-built artifacts collide and the distance between them becomes apparent.
It can only be solved before it starts. Before the conversations diverge past the point of easy synthesis. Before the artifacts begin to accumulate in directions that will cost weeks to unwind.
And here is the part that is uncomfortable to say plainly: most organizations do not currently do that work. They approximate it. They hold the meeting that is supposed to produce alignment and leave with the feeling of having achieved it without checking whether what everyone took out of the room was actually the same thing. They produce the strategy document and assume that reading it produces the understanding it was meant to convey. They rely on the accumulated shared context of people who have worked together long enough to develop a kind of shorthand — and discover the limits of that shorthand the moment the team changes, or the pace accelerates, or the AI agents begin executing on the shorthand as if it were explicit when it never was.
Coordination problems have known solutions. Meaning problems might not. That is not a comfortable thing to write in a methodology book. But the honest version of what this chapter is saying is that the thousand conversations are already happening, the artifacts are already accumulating, and the infrastructure that would make them synthesizable rather than incoherent is, in most organizations, not built yet and not being built.
The conversations are not waiting for the organization to be ready.
And the synthesis problem does not stop at the organizational level. The person running multiple AI conversations in parallel — each producing coherent artifacts, each pointed in a slightly different direction — is the thousand conversations problem at the scale of a single mind.
The question is whether what they produce, when they converge, will be the thing the organization meant.
Coordination problems have known solutions. Meaning problems require something the field has not yet fully described. The conversations are not waiting for the description to arrive.
← Previous: Chapter 5: The Client in the Room · Contents · Next: Chapter 7: Relief or Panic — How to Know Which You Are →
Copyright Walter Reid, 2026. All rights reserved.
