The Four-Layer Map — Drawing It Live

📝 Context

The whiteboard diagram you actually draw in a kickoff — the signature, dual-labeled version of the foundations on-ramp, drawn for a room that contains both engineers and executives. The foundations page explains the model; this page is for drawing it live: the diagram dual-labeled (technical term + plain-English gloss), a "who sees what" lens, and the exact words to say at each layer. Screenshot it before a kickoff, or redraw it from memory on their whiteboard — the redraw is the flex.

🗺️ The Diagram, Dual-Labeled

flowchart TB
  L4["L4 · Architecture and governance — what we are allowed to build"]
  L1["L1 · Apps and agents — the thing the user touches"]
  L2["L2 · LLMOps and infra — the engine room: fast and affordable"]
  L3["L3 · MLOps and data — the discipline for what we train ourselves"]
  L4 --> L1 --> L2 --> L3

Placeholder diagram. This mermaid is a working stand-in. The signature, dual-labeled version (engineer term and the exec one-liner, in the flat-vector house style) is a polished image slated to replace it — see IMAGERY-PLAN.md, visual four-layer-map. Each layer's engineer-vs-exec phrasing lives in the table below until then.

Each layer carries two labels: what an engineer calls it, and what you say to an exec. Holding both in one frame is the translation move.

👁️ Audience Lens — One System, Three Views

The same architecture looks different to each person in the room. Make that explicit and nobody feels talked past.

Layer	Engineer cares about	Exec cares about	Customer cares about
L1 apps & agents	retrieval quality, prompt design, eval scores	does it solve the workflow	"does it answer my question correctly"
L2 infra	latency, throughput, GPU cost	cost per query at scale	"is it fast"
L3 MLOps & data	pipelines, drift, reproducibility	do we own our data advantage	(invisible — and that's fine)
L4 governance	guardrails, audit trails	legal/regulatory risk	"is my data safe"

When the exec and the engineer start talking past each other, point at the layer in question and name whose concern is on the table: "We're on L2 — that's the cost question, which is yours, Dana; the retrieval quality John's describing is L1. Both matter, let's take them in order." You just made the meeting productive.

🧩 Worked Scenario: Drawing It in a Kickoff

A prospect wants an "AI assistant over our policy docs." You draw the four boxes and walk up the stack:

• L1 · draw first — "This is a retrieval app — it reads your policies before answering." The visible product.
• L2 · the cost box — "Here's what decides the running cost — which model, hosted or self-run." Park a number to confirm.
• L3 · cross it out — "You're not training a model, so we skip this whole layer." Crossing it out builds trust — you're not upselling.
• L4 · circle it — "This is the one to settle early — where the policy data lives and who can see it."

Say it like this

"There are four layers to any AI system. The one you're describing lives up here — the app. Below it is the infrastructure that decides what it costs to run, and wrapping all of it is governance: what's allowed and how we prove it's safe. I'm going to cross out the data-science layer entirely, because you're not training anything — and that saves you a lot of money and time."

🚨 Failure Path

The number-one SE mistake in a mixed room is answering at the wrong layer. The exec asks an L4 question — "is our data training their model?" — and the engineer answers with L1 detail — "we use hybrid retrieval with reranking." Both true; only one is responsive. The exec hears evasion; the deal cools.

• Symptom — the exec goes quiet or repeats the question; you answered a different layer than they asked.
• Root cause — engineer reflex: answer with the most technically interesting detail, not the one asked for.
• Recovery — "That's an L4 governance question — let me answer that directly, then John can go deeper on the how." Name the layer, then respond to it.

⚠️ Gotchas

• Quoting the cost box as authoritative on a first call — cost/latency are illustrative until verified against the customer's real volume.
• Skipping the redraw and just sharing a slide — the live redraw is what earns trust; practice it.
• Leaving L4 for later — the data question can reshape the whole architecture; circle it early.

🔗 Links

• The Four-Layer Map (foundations) — the explainer this visual is drawn from
• The Real Cost of a RAG System — the L2 cost box, worked
• AI Vocabulary for Customer Conversations — the L4 governance terms
• IMAGERY-PLAN.md — the polished image slated to replace the placeholder