AI and the Shape of Your Company

Why the same tool feels like magic at 3 people and a mess at 50

Jesse Flores · Compound

The puzzle

A solo consultant runs circles around their old agency with AI.

A 200-person company spends millions on it and gets… dashboards nobody opens.

Same tool. Why?

Open on the contradiction everyone in the room has felt. Don’t mention Compound yet. Promise: by the end you’ll know why, and you’ll see it coming in your own clients. ~1 min.

The one idea

AI is only as good as the context it can reach.

The documentation, the data, the institutional knowledge that lets it act on more than what’s in the prompt.

This is the lens for the entire talk. Everything that follows is this one idea at four different company sizes. Say it slowly. ~1 min.

Stage 1   You + AI

When it’s just you, you are the system.

• You’re the system of record and the system of knowledge.
• Everything AI needs is in your head — and you hand it over in a sentence.
• Zero context gap. That’s why it feels like magic.

You + AI genuinely outproduce a team of five from a few years ago. That leverage is real.

This is most of the room. Validate it honestly — their leverage is real, don’t undersell it. The reason it works: no context gap, because they ARE the context. ~1.5 min.

Stage 1   Win at your stage

Close the context gap with one tool.

• Projects — give Claude a persistent home per client/workflow. Stop re-explaining yourself.
• Custom instructions / styles — encode how you work, once.
• Connectors — let it reach your real tools (Drive, calendar, docs), not just what you paste.

Every one of these is the same move: shrink the gap between what AI knows and what you know.

The “use this Monday” payoff for THIS audience. CONFIRM/UPDATE these to the exact Claude features you want to demo live — current product naming. Frame each as closing the context gap. ~1.5 min.

Stage 2   The small team

The first crack: context lives in more than one head.

• What you know, your VA and your contractor don’t.
• The fix is still small and cheap: shared Projects, shared instructions, a shared home for the team’s context.
• Now AI works off the team’s knowledge, not one person’s.

At this size you can still solve the context problem by hand. That’s about to stop being true.

Claude Teams slot — shared projects, shared knowledge, seats. Confirm specifics. The key turn at the end: the by-hand fix has a ceiling. ~2.5 min.

Stage 1 Stage 2 Stage 3   The growth wall

At 25–100 people, the knowledge scatters.

It fragments across three places at once:

• People — what’s in someone’s head after 31 years.
• Systems of record — the CRM, the ERP.
• Systems of knowledge — docs, spreadsheets, the 147-row pricing file on one desktop.

Nobody holds the whole picture anymore. AI walks in like a new hire looking for a seat — and there’s no seat, because the work was never explicitly designed.

This is the turn of the whole talk. Make the three buckets concrete. The “new hire with no seat” line is the bridge to why their AI spend fails. ~1.5 min.

Complexity · people + knowledge

Add people, and the connections explode.

● people   ● shared systems   • personal silos (inbox, files, one head)

The complexity slide. Click to advance the three stages: small (legible) → growing → 25–100 (the hairball). Cream = people, red = systems of knowledge. The payoff line: every red node is context AI has to reach, and at scale no one even holds the map. Ground it verbally: communication paths grow as n(n-1)/2 (Brooks); the who-knows-what directory (transactive memory) fragments as you grow. ~2 min.

Complexity outruns headcount

Linear people. Quadratic complexity.

People	Connections
3	3
7	21
25	300
50	1,225
100	4,950

Drives the “complexity outruns headcount” point with hard numbers. Cream dashed = people (linear), red = connections n(n-1)/2 (quadratic). Add people one at a time, the work of keeping everyone in sync grows far faster. Capped at 50 on the chart; the table carries it to 100 (4,950). ~1 min. (Optional spoken: Reed’s Law puts possible sub-contexts at 2^n — exponential.)

Why the AI spend fails

It should work. That’s the trap.

• When you’re small, build-first is the right move. You’re the context, so you hand AI a sentence and it works.
• That’s the instinct everyone has earned — including everyone in this room.
• At 25–100 people, they run the same move on an org where the context is scattered across people and systems. The seat AI needs was never built.

The failure isn’t a wrong tool. It’s a wrong order — inherited from the smaller company they used to be.

The key idea of the talk: the build-first instinct isn’t dumb, it’s a reasonable extrapolation. It’s exactly how AI works at Stage 1/2, so of course they expect it to work at scale. The category error: treating a fragmented org like an individual who holds all the context. Spoken closer: “If you’ve watched a client buy AI and get nothing back — this is the reason.” ~1.5 min.

Why Compound exists

Closing the gap is a discipline, not a tool.

Give a 50-person company the thing a solopreneur has for free: AI working against context it can actually reach.

The method in one line: diagnose before you build. Name the real constraint, find where the knowledge lives — then build. Run it as a Sprint. Repeat on a rhythm.

Don’t teach the whole six-stage Sequence. Just establish there’s a real, repeatable method — operating design — and that it’s what lets a bigger company recover the magic. ~1 min.

The Compound Sequence

Six stages, in order.

01

Signal

Name the real constraint, and price it.

→

02

Source

Map what the org knows, and where it lives.

→

03

Design

Decide who's accountable: people and agents.

→

04

Build

Build the system against a locked design.

→

05

Deliver

Install it into how the work gets done.

→

06

Compound

Capture the gain; the next one's cheaper.

Signal + Source = Diagnose — the part big companies skip. The other four execute and compound. The order is the discipline.

Walk it left to right; the order IS the method. Why each one solves the large-company (fragmentation) problem: Signal — the loudest symptom isn’t the constraint; name and price it so you don’t automate the wrong thing. Source — find the scattered knowledge (the red nodes); you can’t give AI context you haven’t located. Design — build the seat that never existed: who’s accountable across people and agents. Build — fast precisely because the diagnosis is done; this is the step everyone wrongly starts at. Deliver — at scale many roles touch one workflow; this is where the change holds instead of getting abandoned. Compound — capture it so the next sprint is cheaper and the company compounds. ~2.5 min.

After the Sequence

The shape resolves.

● people   ◆ agents   ● knowledge

Each person now manages a team of agents, and the agents hold the knowledge. A human touches a handful of agents — not a hundred scattered systems. The tangle is gone, and the people don't have to.

The resolution slide — the payoff for the whole arc. Three tiers: people manage agent teams, agents hold the connections to knowledge. The scattered person-to-knowledge silos from the earlier chart get subordinated to the agent layer, so a human touches a few agents instead of a hundred systems, and the human-to-human tangle is gone. Honesty guardrails: (1) people don’t disappear — they stop being routers and move up to managing agents and doing judgment work; (2) agents, not humans, carry the knowledge touchpoints, which is what removes the n² mesh; (3) the agent team is hierarchical, supervised by a person, operating inside guardrails humans set. ~1.5 min.

The next level up

Who’s ready to compound?

You’re mostly Stage 1 and 2 — keep riding it. But you work with Stage 3 companies every day.

25–100 people. Growing. Knowledge stuck in people’s heads. AI spend not paying off.

When you meet one — that’s a Compound conversation. Send them my way.

The referral, framed as a position on a curve, not a sales ask: “your clients are one stage up from you.” Drill the repeatable sentence — that’s what gets passed along when you’re not in the room. End on the book + how to reach you. ~1 min.

Thank you

Co-Intelligent Co-Operation

Download the book.

Free. Scan the code, or visit the link.

Jesse Flores · Compound

go.compoundorg.com/cointelligent-cooperation

Hold for questions. The QR and the short link both go to the free book download (go.compoundorg.com/cointelligent-cooperation). ~30 sec.