AI Maturity in Engineering Organisations
A Framework from Autocomplete to Self-Evolving Factory
How to Read This Framework
This framework maps the adoption of AI in software engineering across twelve levels. Each level represents a distinct shift in one or more of five dimensions:
Who drives the work The human or the agent
Who initiates the work A person, a system event, or the AI itself
Scope of AI awareness Token-level, file-level, codebase, or organisational
Integration surface Editor, terminal, messaging platform, CI/CD, or autonomous pipeline
Knowledge context None, generic, codebase-specific, or institution-specific
The levels cluster into three eras. The boundary between Individual AI and Institutional AI is where the unit of productivity shifts from the person to the organisation. This is the most important line in the framework.
The Three Eras
| Era | Levels | Defining Characteristic |
|---|---|---|
| Era I: Individual AI | L1 — L8 | A developer uses AI to amplify their own output. Autocomplete, chat assistants, agentic coding tools. The human drives everything. |
| Era II: Transitional | L9 | The organisation builds shared AI infrastructure. Work remains human-driven, but the tooling is institutional. The bridge between two worlds. |
| Era III: Institutional AI | L10 — L12 | Agents operate as an organisational system. Humans govern, direct, and review. The factory runs. This is where the future lies. |
Summary
Institutional AI
L12Self-Evolving FactoryThe factory builds itself. Humans hold strategic authority. No company operates here yet.
L11Agent FactorySystem-initiated agents. Cron-driven quality sweeps, auto-remediation, coordinated multi-agent pipelines. e.g. StrongDM, AlphaEvolve
L10Human-Triggered Inst. AgentsA human sends a message or creates a ticket; agents plan, implement, test, and submit PRs within institutional infrastructure. e.g. Stripe Minions, Ramp Inspect, Uber, Coinbase
Institutional AI Line
L9Org-Tuned ToolingShared AI infrastructure: RAG, custom MCP servers, org-specific prompt libraries, CI/CD integration. Tooling is institutional, workflow still individual. e.g. Factory AI, NemoClaw, CodeRabbit, Snyk AI
Individual AI
L8Agent MultiplexingMultiple uncoordinated agents in parallel. Output metrics explode, organisational coherence collapses. The trap level: +98% PRs, flat delivery. e.g. parallel Cursor sessions, multiple Devin instances
L7Custom Trigger InterfacesDeveloper builds personal automation: Slack bots, custom scripts calling foundation models. No shared context or coordination. e.g. OpenClaw, Agentic Coding Flywheel
L6Agent-First DevelopmentThe agent is the primary author. Developer specifies intent and evaluates output. The IDE becomes a review surface. e.g. Claude Code, Devin, Codex, Replit
L5AI Co-pilot, AutonomousDeveloper stops reviewing every diff. Steers at intent level rather than approving individual changes. e.g. Cursor Agent, Windsurf Cascade, Copilot Agent mode
L4AI in IDE, SupervisedAI inside the editor with project context. Proposes multi-file edits, but every change requires explicit approval. e.g. Cursor, Windsurf, Trae, Kiro, Cline
L3AI Chat AssistantSeparate chat interface. No project context. Every interaction starts from zero. Integration layer is copy-paste. e.g. Claude.ai, ChatGPT, Gemini
L2AutocompleteAI predicts the next tokens inline. No dialogue, no multi-file awareness. Accelerates typing only. e.g. GitHub Copilot, Tabnine, Supermaven
External Frameworks
Trust Levels
→Individual developer trust progression with AI tools
Maps to: L1-L8 (individual trajectory)
Five Levels: Spicy Autocomplete to the Dark Factory
→Developer autonomy levels modelled on NHTSA driving automation levels
Maps to: L1-L12 (broad mapping)
Software Factory
→No human code, no human review — team practice at one company
Maps to: L11
DORA
→Organisational delivery metrics across the industry
Maps to: Measures outcomes at all levels