Behind the BuildWhy I Stopped Having AI Conversations and Started Holding AI Meetings

Why I Stopped Having AI Conversations and Started Holding AI Meetings

Why I Stopped Having AI Conversations and Started Holding AI Meetings

Why I Stopped Having AI Conversations and Started Holding AI Meetings

For about two years I used AI the way most founders do. I had conversations with it.

I would open a chat, give it the context, work the problem with it, get an output, sometimes excellent, and act on it. The conversations were productive. The thinking was sharper than I would have done alone. The outputs were directionally right more often than they were wrong.

But the decisions did not survive the next morning's coffee.

I would re-read what I had decided the night before. The output was still there. The logic was still there. The reasoning was sound. And yet, in the daylight, the decision felt thin. Not wrong. Thin. As if I had been part of a conversation that ended too soon, with too few perspectives, and had committed too quickly to a path that one voice had agreed with too easily.

I am writing this for founders who already use AI for thinking and are starting to notice the same pattern. The conversations are good. The decisions, somehow, are not.

What was actually wrong with the AI conversations

I sat with the pattern for several weeks before I understood it.

The AI conversations were exactly that: conversations. One voice. One point of view. One probabilistic centre, talking to me. When I challenged the output, the voice updated its position. When I pushed back, it conceded. When I asked the same question framed differently, it gave a subtly different answer that I could not always reconcile with the first one.

What I was missing was not better outputs. The outputs were already good. What I was missing was a room. A room with more than one voice in it. A room where one voice pushed back on another, where a disagreement was held on the record, where the synthesis at the end was a synthesis of actual contradictions and not a confident summary of a single line of reasoning.

The decisions felt thin in the morning because they had been made in a single voice's room. The lighting in that room was good. The acoustics were not.

The move from chats to meetings

So I rebuilt the surface as actual meetings.

A meeting is not a longer conversation. A meeting has a structure conversations do not have. It convenes a specific set of people, each with a specific lens. It runs a sequence — agenda, discussion, synthesis, decision, mark outcome, implement. The voices are named. The disagreements are visible. The decision is recorded with the disagreement attached, not buried.

That is what an AI meeting needed to be too. Not a long chat with a thoughtful single voice. A convened group of named voices, each reasoning from a stable standing brief, talking through a sequence, producing a synthesis that survives because it has been earned against opposition.

The first room I built was for the Strategic Council. Eight named voices — the Council Director, the Operator, the Growth Advisor, the Margin Advisor, the Red Team, the Capacity Advisor, Peer Review, the Chairman. Each agent has a standing brief in the Brain that gives it stable identity and lens. The brief is cached so the cost of bringing it to the meeting does not balloon. The agents speak in their own voices across sequential turns. Attachments are read in as text and held as cached context. The Chairman synthesises at the end. An eight-template dispatch engine turns the decision into the actions it implies.

That sounds like a lot of machinery for what used to be a chat. It is. The machinery is the point.

The day I knew it had worked

Five weeks ago, I needed to decide whether to pull forward the Behind the Build product-catalogue posts into the daily content stream or hold them as their own track.

In the AI-conversation era, I would have opened a single chat, given it the context, worked the problem, and decided in twenty minutes. The decision would have felt clear. By Monday morning it would have felt thin.

I convened the meeting instead. The Margin Advisor came in with the unit-economics read on whether the product-catalogue posts move conversion enough to justify the daily-stream displacement. The Growth Advisor came in with the funnel-leak analysis on which product pages were under-fed by traffic. The Operator came in with the workflow impact on the seven-specialist content pipeline. The Capacity Advisor flagged what would have to stop on the EFO side to make room. The Red Team flagged a failure mode I would not have surfaced alone — what happens to the daily-stream voice consistency if product-catalogue posts get visibly different tonally. The Chairman synthesised across the disagreement.

The synthesis I walked out with said: pull forward two per week, never adjacent, with the topic-bank structure that names the product and its CTA destination in the h3. Hold the rest as their own track. Reconsider in eight weeks against measured product-page traffic.

That decision survived the next morning's coffee. It is the architecture you are reading inside right now.

The conversation version of the same problem would have produced a similar answer, faster. The meeting version produced a decision I had not had to re-litigate three times since. That difference is the product.

What the room actually contains, and what it costs

The AI Meeting Rooms product brings any group of agents from the registry into one room. Each speaks in its own voice across sequential turns, reasoning from a Brain-cached standing brief. Attachments read in as text, held as cached context. The Chairman synthesises. The eight-template dispatch engine turns the synthesis into ready-to-apply downstream actions.

The cost per meeting is roughly eighty-five percent lower than the same panel would cost without prompt caching and tier-based model selection. That is not a marketing line; the savings are mechanical. Standing briefs cache. Lower-tier models handle the procedural moves; higher-tier ones handle the synthesis. The total bill per decision lands inside what a founder running this for their own week can absorb.

Customer one was me, in production since 29 May. The Council you have been reading about all week — the Margin Advisor's seat, the Growth Advisor's seat — runs in this room. The room itself is the engine.

Who else should be holding meetings instead of conversations

The product is built for founders who are already comfortable using AI for individual thinking and are starting to notice the same thinness in their decisions that I was noticing in mine.

If you have ever closed a chat session feeling that the answer was good and the decision was somehow not quite settled, this is the room you are missing. If you have ever re-opened the same chat the next day and found yourself going back over ground you thought you had finished, you have been missing the synthesis layer. If you have ever wished there was someone in the conversation pushing back on the AI's confident centre with a different lens, you have been wanting the second voice in the room.

The room is what you have been wanting.

TL;DR

I used AI for two years and noticed my decisions did not survive the next morning's coffee. The conversations were good; the decisions were thin. They were made in a room with one voice in it. I rebuilt the surface as actual meetings — named agents, sequential turns, cached standing briefs, Chairman synthesis, an eight-template dispatch engine. The room costs roughly fifteen percent of what an uncached panel would cost. Customer one was me, production-live since 29 May. The same room is what powers the Strategic Council you have been reading about all week. The product exists because the AI-conversation surface stops scaling at the point where you stop having ideas and start needing decisions. The room is what comes after.


If you have stopped trusting your AI conversations to be decisions and want to see what an actual meeting looks like, here is the room. → /products/ai-meeting-rooms

Learning Materials

Key Vocabulary

foundernoun · B2

A person who establishes a company or organisation, especially a startup.

I used AI the way most founders do.

thinadj · C1

Lacking depth, substance, or convincing reasoning; weak in argument or evidence.

In the daylight, the decision felt thin. Not wrong. Thin.

concedeverb · C1

To admit reluctantly that something is true, or to yield a point in an argument.

When I pushed back, it conceded.

reconcileverb · C1

To make two ideas, accounts, or statements compatible or consistent with each other.

It gave a subtly different answer that I could not always reconcile with the first one.

synthesisnoun · C1

The combination of different ideas, perspectives, or elements into a single coherent whole.

The Chairman synthesises at the end.

conveneverb · C1

To formally call together a meeting, group, or assembly.

A meeting convenes a specific set of people, each with a specific lens.

lensnoun · C1

A particular perspective or framework through which someone analyses or interprets a situation.

Each agent has a standing brief in the Brain that gives it stable identity and lens.

standing briefphrase · C1

A permanent, reusable set of instructions and context that defines a role's identity, scope, and viewpoint.

Each agent has a standing brief in the Brain that gives it stable identity and lens.

cacheverb · C1

In computing, to store data temporarily so it can be retrieved more quickly on subsequent uses.

The brief is cached so the cost of bringing it to the meeting does not balloon.

displacementnoun · C1

The act of moving something from its proper or usual position to make room for something else.

Whether the product-catalogue posts move conversion enough to justify the daily-stream displacement.

failure modephrase · C1

A specific way in which a system, process, or plan can break down or produce a bad outcome.

The Red Team flagged a failure mode I would not have surfaced alone.

re-litigateverb · C2

To argue or debate a decision or issue again that was supposedly already settled.

The meeting version produced a decision I had not had to re-litigate three times since.

unit economicsphrase · C1

The direct revenues and costs associated with a single unit of a business model, used to assess viability.

The Margin Advisor came in with the unit-economics read.

absorbverb · B2

To take on a cost, impact, or burden without passing it on or being damaged by it.

The total bill per decision lands inside what a founder running this for their own week can absorb.

scaleverb · B2

To grow or expand efficiently while maintaining or improving performance.

The AI-conversation surface stops scaling at the point where you stop having ideas and start needing decisions.

Grammar Notes

Past perfect for sequencing reflection on earlier decisions ('what I had decided', 'they had been made')

The past perfect (had + past participle) is used to mark an action or state that happened before another past reference point. The post uses it to look back on yesterday's decision from the perspective of this morning's coffee, and on the closed-room nature of those decisions from the perspective of the present analysis. This double-layered past is central to the post's argument.

I would re-read what I had decided the night before. The decisions felt thin in the morning because they had been made in a single voice's room.

Common mistake: Learners often default to the simple past in both clauses ('I re-read what I decided the night before'), which loses the temporal contrast between the morning re-reading and the evening decision.

Habitual 'would' for repeated past behaviour ('I would open a chat', 'the voice updated its position')

'Would + base form' describes a typical, repeated action in the past — a habit or characteristic behaviour. It is distinct from 'used to', which signals a clear contrast with the present. 'Would' suits sustained patterns of behaviour you want to describe vividly. The opening paragraphs of the post use it to set up the two-year pattern.

I would open a chat, give it the context, work the problem with it, get an output, sometimes excellent, and act on it.

Common mistake: Using 'would' for states ('I would have a chat tool') instead of habitual actions, or mixing it with 'used to' inside the same paragraph for the same idea, which sounds clumsy.

Parallel structures with anaphora for rhetorical weight ('If you have ever closed a chat... If you have ever re-opened... If you have ever wished...')

Anaphora is the repetition of the same word or phrase at the beginning of successive clauses or sentences. Combined with parallel grammar — here, the present perfect 'have ever + past participle' — it builds rhythm and authority, and lets the writer aim three precise scenarios at the same reader without losing flow.

If you have ever closed a chat session feeling that the answer was good and the decision was somehow not quite settled, this is the room you are missing. If you have ever re-opened the same chat the next day... If you have ever wished there was someone in the conversation pushing back...

Common mistake: Breaking the parallel by changing tense or clause shape midway ('If you have ever closed a chat session... When you re-open the same chat... If you wish someone would push back'), which dissipates the rhetorical effect.

Nominalisation for compressed professional shorthand ('the daily-stream displacement', 'funnel-leak analysis', 'workflow impact')

Nominalisation turns verbs and adjectives into nouns ('to displace' → 'displacement'; 'to leak' → 'leak'; 'to impact' → 'impact'). It compresses a full clause into a noun phrase that can carry a hyphenated modifier in front of it. Senior professional English uses nominalisation heavily because it lets you pack causal or analytical content into a tight subject or object.

The Growth Advisor came in with the funnel-leak analysis on which product pages were under-fed by traffic. The Operator came in with the workflow impact on the seven-specialist content pipeline.

Common mistake: Over-nominalising ('the operationalisation of the implementation strategy'), which sounds bureaucratic. The skill is to use nominalisation only where it removes a clause rather than dressing up a simple verb.

Comprehension Questions

  1. 1.According to the writer, what was the recurring problem with the decisions he reached through AI conversations?
  2. 2.What are the structural differences the writer identifies between a conversation and a meeting?
  3. 3.Why does the writer claim that the high cost of running an AI meeting is not actually a barrier?
  4. 4.What does the worked example about pulling forward the Behind the Build product-catalogue posts demonstrate about the value of meetings over conversations?
  5. 5.If a founder reading this post recognises themselves in the 'closed the chat feeling the decision was not quite settled' pattern, what does the writer suggest they do, and why is that suggestion specific rather than generic?

Run your own diagnostic

Use the same Strategic Council I run my own decisions through. The assessment preview is free. The specific central human intelligence it is based on is verified in person during the call.

Start the free diagnostic →