How mature is your organisation's AI practice, really?
A free scoring system for tracking AI progress — ten ladders, eight levels each, built from advisory work since 2022. Enable JavaScript for the full interactive rubric.
The ladders
- Mode of thinking — How people actually use AI day to day.
- Prompting — How well the organisation instructs AI.
- Human–AI trust / QA — How you know AI work is safe to ship.
- Heartbeat agent invocation — Whether AI keeps working without you.
- Model routing — Which AI model does each piece of work.
- Spend absorption — How much AI spend the organisation can digest.
- Sovereignty / model control — Who controls your models, data, and continuity.
- Parallelization — How many AI jobs run at the same time.
- AI impact measurement — Whether AI is actually paying off.
- Governance of AI rollouts — How safely AI reaches production.
Sample of the matrix
Showing plainLanguage for levels L1, L2, L3, L5, L7.
| Ladder | L1 | L2 | L3 | L5 | L7 |
|---|---|---|---|---|---|
| Mode of thinking |
AI is a one-off answer box. |
A person directs AI to produce a requested artefact. |
AI helps someone challenge and reframe a decision. |
Teams share reusable AI workspaces, assistants, and harnesses. |
Humans supply intent and accountability while agents work in specialist silos. |
| Prompting |
Ask AI without examples or a required format. |
Show examples and specify the answer format. |
Deliberately structure conversations around the task. |
Reflect on prompts and give subagents distinct briefs. |
Build reusable knowledge and systems that improve prompts. |
| Human–AI trust / QA |
Humans author the work and AI only inserts fragments. |
AI drafts the work, but human inspection controls quality. |
An agent retries against its own executable definition of done. |
Independent checks combine so no producer certifies itself. |
Independent systems revise assurance within human-set constitutional limits. |
| Heartbeat agent invocation |
AI use is chat or one-shot only. |
Work exists only while someone directly runs it. |
Agents run remotely or on simple timers. |
A paid always-on agent remembers earlier runs. |
A self-feeding factory works unattended within human-set constraints. |
| Model routing |
The system chooses the model without making it a decision. |
People manually choose a model for each kind of task. |
Context determines fixed model roles for a job. |
A shared gateway centralises selection, failover, and early cost control. |
Traffic data improves routing policy across the estate. |
| Spend absorption |
AI is free-only and treated solely as a cost. |
Each person has one tightly capped low-cost account. |
Selected teams can spend modestly, usually by exception. |
Productive power users routinely run expensive parallel agents. |
AI capacity is procured as major organisational infrastructure. |
| Sovereignty / model control |
No meaningful model choice is expressed. |
Quality is prioritised while data and legal controls are ignored. |
Basic controls exist but teams apply them inconsistently. |
Automated controls govern model choice, data, and local failover. |
The organisation controls the hardware and jurisdiction of inference. |
| Parallelization |
One private chat handles work serially. |
One person juggles several chats but integrates them personally. |
Reusable lanes retain context outside a single chat. |
People delegate background work and return for the result. |
Organisational queues feed an always-on agent fleet. |
| AI impact measurement |
The organisation measures access but not use or impact. |
One tool’s activity is mistaken for organisation-wide AI use. |
Teams measure local before-and-after task changes. |
A specific AI setup is causally linked to a business outcome. |
Measurement automatically improves the whole AI production system. |
| Governance of AI rollouts |
AI use is invisible until an incident exposes it. |
Policy exists, but there is no usable governed route. |
The organisation detects some AI use after the fact. |
Fast-built systems have nowhere safe to enter the estate. |
Governance is an externally enforced platform capability. |