ATOM • the four-layer model

The middle is where it is won or lost.

The tools above are rented; the systems below are fixed. Between them sit the two layers the enterprise actually owns — where the work happens, and the rules, memory, and review steps that make it useful.

Layer 1 · rented

AI tool layer

Powerful, but rented — buying more of it doesn’t create adoption.

Chat & assistants
ChatGPTClaudeGemini
Embedded copilots
M365 CopilotWorkspaceCRM copilots
Coding & agents
CursorCodexCoding agents
Connectors & platforms
MCPConnectorsWorkflow platforms
Admin & controls
ConsoleAnalyticsAuditRuntime

The customer-owned operating core

Two distinct layers, one owner — the work, and the rules it runs within.

Layers 2 & 3 — owned
Today — people stitching tools, context & systems together by hand Turned into reviewed, repeatable work
Layer 2 · the work

Work execution layer

Where people and agents do the work.

Who works People AI-assisted people Autonomous agents
Build & run Skills Workflows Custom agents

→ meetings, research, deals, documents, decisions, handoffs

↑ Layer 2 runs the skills & agents as live work · Layer 3 defines the contracts, context & controls they run within ↓
Layer 3 · the system

Rules, memory, and review (the harness)

The structure around the work: governance, security, policy, and memory.

Context & policy Context contracts Source rules Security & policy Permissions
Skills, control & proof Skill/agent contracts Action boundaries Memory promotion Receipts & gates Economics
Measured here — this is where the work becomes a readout that turns a pilot into a renewal
Cost
seats, context, rework
Context
approved, missing, stale
Control
policy, actions, gates
Memory
reusable vs local
Performance
speed, quality, outcome
Portability
survives a tool swap
Layer 3 → Layer 4 · approved reach

Approved connectivity

The rules decide what connects — and what it may touch.

Connectors APIs MCP servers Approved plugins
Layer 4 · the ground truth

Third-party systems, sources & data

The company’s own systems and data — the third-party products it runs and the records they hold.

Systems of record
CRMERPHRISFinanceMarketing ops
Data & analytics platforms
SnowflakeDatabricksWarehouses & lakesBI / semantic layer
Communication & calls
EmailChatCalendarCall transcripts
Docs, content & knowledge
Docs & driveWikisFiles & evidence stores
Engineering systems
Code & reposCI/CDIssue trackersObservability
Service & product
Support / ticketingCS toolsProduct telemetry

Flagship visual · calibration v2 — four distinct layers, the owned middle as the centerpiece

Pipeline Rebel