Methodology

The AIMI 5-Level Maturity Model — public by design, grounded in published research and real enterprise implementations.

What the external evidence shows

AIMI follows the published research on where AI value is actually won and lost. Each finding maps to a dimension we measure.

~5% of enterprise GenAI pilots reach measurable P&L impact

The other ~95% stall — and the cause is integration and learning, not talent or infrastructure. AIMI measures exactly that gap.

MIT NANDA, “The GenAI Divide,” 2025

Only ~10% of AI value is the algorithm; ~70% is people & process

The model is the easy part. AIMI weights People & Skills and Process Integration accordingly.

BCG, 10-20-70

Redesigning workflows is the #1 driver of bottom-line gen-AI impact

Tracking well-defined KPIs is the top practice — exactly AIMI’s Process Integration and Value & Measurement.

McKinsey, The State of AI

Regulation & risk is now a top barrier to scaling AI (~38%)

Governance is the gate, not an afterthought — which is why it caps the AIMI level.

Deloitte, State of GenAI (2,773 leaders)

AI adoption ~78% while inference cost fell ~280× in two years

Cost and access are no longer the constraint. Organizational maturity is — the question AIMI answers.

Stanford HAI, 2025 AI Index

Figures are as reported by the cited organizations; where editions or definitions differ we cite a conservative published figure. Full sources accompany every emailed report.

The 5 levels

L1 — Manual / Ad-hoc

No systematic AI; at most individual, ungoverned tool use.

L2 — Assisted

Point AI tools assist humans in isolated tasks; pilots; no workflow integration; ROI untracked.

L3 — Integrated

AI embedded in core workflows; task-specific agents; ROI tracked; basic governance.

L4 — Orchestrated

Collaborating agents across the function; human-in-the-loop; continuous improvement; risk controls.

L5 — Transformed

The function is re-designed around AI; end-to-end agents under governance; cross-app, cross-function.

The 6 dimensions

Data & Knowledge

Is the operational data and knowledge structured, accessible and AI-ready?

Tooling & Agents

What AI capability is actually deployed — assistants, agents, orchestration?

People & Skills

Team AI-literacy, attitude and change readiness.

Process Integration

Is AI bolted-on or embedded in the core workflow, with an improvement loop?

Governance & Risk

Oversight, EU AI-Act posture, data protection, human-in-the-loop, ethics. Caps the overall level.

Value & Measurement

Is ROI / outcome tracked and acted on?

Governance caps the level. A function cannot sit above what its oversight can carry — you cannot be more "Transformed" than you can govern.

Research foundation

AIMI is not an opinion quiz. Four foundations keep it defensible enough to put in front of a board:

1. An established instrument

Staged capability-maturity models are a decades-old research instrument (the CMMI lineage, later digital- and AI-maturity frameworks). AIMI applies that discipline to a single operational function rather than a whole company — the level where transformation is actually won or lost.

2. An evidence base of real implementations

Every roadmap ships with named, published deployments and their reported outcomes — Klarna, IKEA "Billie", Anthropic's adoption of Intercom Fin, finance, and field-service operators. Each report carries the source link. The ROI and payback ranges are anchored to these public cases, not invented.

3. A regulatory anchor

The Governance dimension is mapped to the EU AI Act (Regulation (EU) 2024/1689). The main tranche of obligations — transparency, high-risk conformity, human oversight — applies from 2 August 2026. Maturity that outruns oversight is flagged as exposure, not progress.

4. Honest benchmarking

Peer benchmarks are released per segment only at N≥30 respondents. Below that we show a clearly-labelled, model-based reference — never a fabricated peer number. The benchmark becomes empirical as the sample grows; the methodology that produced it stays on this page.

Selected evidence

A sample of the published implementations behind the function roadmaps:

  • Klarna — AI assistant handled 2.3M chats in month one (~700 agents' workload), two-thirds of chats; later rebalanced toward humans for complex cases — a governance lesson, not a failure.
  • IKEA — "Billie" resolved ~47% of inquiries while 8,500 call-centre agents were reskilled into design advisors, linked to €1.3B in new sales.
  • Anthropic — adopted Intercom's Fin agent (build-vs-buy): a 50.8% resolution rate (96% Fin involvement) in month one, under a hybrid model.
  • Gartner — finance AI adoption reached ~59% of leaders in 2025 (from 37% in 2023); the gap is execution, not intent.
  • Field service (Geotab 2025) — 88% of field-service organizations report better uptime/cost/CX and 75% better first-time-fix from AI.

Full source links are included in every emailed report. Figures are reported by the cited organisations and public coverage; AIMI does not independently audit them.

How the score is computed

Scoring is deliberately legible — a board member can re-derive it. Each of the 6 dimensions is scored 1–5 from behavioural questions (concrete operational ladders, not agree/disagree). The overall level is the rounded mean of the six dimensions — then the Governance gate applies: Level 5 requires top Governance (5/5), Level 4 requires Governance ≥3, Level 3 requires Governance ≥2. Governance can only pull the level down, never up: a function that has deployed AI it cannot oversee is not more mature — it is more exposed. Benchmarks are shown per segment (function × industry × geography × company size) only at N≥30 (also a minimum-cell / k-anonymity floor); below that, a clearly-labelled model-based reference, never a fabricated peer number. AI-generated content is labelled per the EU AI Act (obligations from 2 August 2026).

Known limits — stated, not hidden

Credibility means naming the limits, not just the strengths:

Self-report

A single respondent self-scoring their own function carries optimism and framing bias. We run consistency checks and exclude low-effort patterns (identical answers across all dimensions) from the aggregate; multi-respondent input is on the roadmap.

Convenience sample

Early respondents skew toward the AI-engaged — not the whole economy. That is exactly why the benchmark stays model-based and clearly labelled until the sample is representative.

Uniform weights (v1.0)

The six dimensions are weighted equally until the data justifies differential weights empirically.

Independent validation pending

The methodology is published here to be challenged; external academic review is a goal, not yet a claim.

Methodology v1.0 — published in full. Any change to the scoring bumps the version and is dated.