The metric everyone gets wrong
Many rollouts still start with seat count: how many people have a Copilot or ChatGPT licence. That tells you who's paying, not who can use it. I've watched a company roll out a tool org-wide and get real leverage from a handful of teams, because licence count and capability aren't the same thing, and nobody was measuring the second one.
I built the framework below while rolling out AI across a scale-up, after the seat-count metric stopped being useful. It's a five-level fluency ladder crossed with the major functions of a company, so "good" is defined per function rather than as one vague, org-wide bar.
The ladder
Zapier published a four-level model for hiring and onboarding: Unacceptable, Capable, Adoptive, Transformative. It's a solid starting point, and I extended it two ways: an explicit floor, and a fifth rung between "uses AI well" and "sets AI strategy".
- Resistant. Avoids AI tools or treats them as a threat to how work already gets done.
- Aware. Uses AI occasionally for low-stakes personal tasks, but has no repeatable workflow or quality bar.
- Applied. Runs AI inside a real workflow with a verification step and clear data boundaries, and can point to the gain.
- Advanced. Builds reusable workflows with review, evals and monitoring built in; owns the failure modes, not just the prompts.
- Strategic. Sets direction: risk appetite, data governance and where the function's operating model changes, and owns the tradeoffs.
The floor matters more than it looks. Most fluency frameworks start at "beginner", which quietly implies that using AI at all is optional. Naming Resistant as its own level makes the bar explicit: it isn't a starting point on the ladder, it's a gap the rest of the org has to notice and close.
Why one column isn't enough
A single fluency rubric fails the moment you apply it across functions, because "Applied" looks completely different in engineering than it does in finance. An engineer at Applied is running an agentic coding tool on scoped tasks and reviewing every diff before merge. A finance lead at Applied is using AI to draft variance commentary and reconciling it against source systems before it reaches a decision. Same level, same rigour, entirely different work.
That's the actual insight behind the framework: the unit of assessment is the cell, not the column. You don't ask "is Engineering fluent?". You ask what Applied looks like for Engineering specifically, write it down, and use that as the bar.
Resistant
Avoids AI tools or treats them as a threat to the way work already gets done.
Aware
Uses AI occasionally for low-stakes personal tasks, but has no repeatable workflow or quality bar.
Applied
Runs AI inside a real workflow with a verification step and clear data boundaries, and can point to the gain.
Advanced
Builds reusable workflows with review, evals and monitoring built in; owns the failure modes, not just the prompts.
Strategic
Sets direction: risk appetite, data governance and where the function's operating model changes, and owns the tradeoffs.
| Function | Resistant | Aware | Applied | Advanced | Strategic |
|---|---|---|---|---|---|
| Engineering | Won't use an AI coding assistant; reviews AI-authored PRs from teammates with blanket suspicion rather than normal scrutiny. | Uses an AI coding chat for explanation or debugging now and then, but doesn't trust it on anything touching more than one file. | Runs an agentic coding tool daily for scoped tasks, tests and refactors, reviews every diff before merge, and never lets it merge unsupervised. | Maintains the repo context, task templates and CI checks that make agent output consistently mergeable across the team, and catches the subtle bugs an agent introduces before they land. | Sets policy on sandboxing, secrets, production access and dependency updates for agents, and tracks the quality and throughput numbers that justify it. |
| Product & design | Insists every spec and mock is made by hand, on principle, even when a first draft would save a day. | Uses AI for rough flows or copy variants but can't turn the output into a testable product decision. | Uses AI to draft PRDs and synthesise interview transcripts with evidence traced back to source, then edits hard before anyone else sees it. | Builds working prototypes in an AI app builder to test interaction and copy assumptions with real users before committing engineering capacity. | Redesigns discovery around AI-assisted synthesis and measures the effect on decision quality and experiment velocity, not just speed. |
| GTM (sales & marketing) | Refuses AI-drafted outreach on the grounds that it feels impersonal, without weighing the actual risk to trust or deliverability. | Has used AI to write a subject line or two but still drafts full sequences manually, with no read on what it's actually improving. | Runs AI-assisted account research and first-draft outreach as a standing pipeline step, with CRM hygiene and human personalisation before anything sends. | Builds and maintains an AI lead-qualification agent wired into the CRM, monitoring false-positive rate and handoff quality, not just prompt tuning. | Redesigns GTM around signal-based automation and human-led high-intent moments, measured against CAC and payback, not headcount saved. |
| Finance & legal | Rejects even access-controlled, already-scoped AI use cases on blanket principle, without weighing the risk that was actually assessed. | Has asked an AI tool to summarise a document once but doesn't use it for anything that reaches a real decision. | Uses AI to draft first-pass variance commentary and flag anomalies, reconciled against source systems and materiality thresholds before it's trusted. | Runs a reviewed AI workflow for contract redlining or month-end close checks that the team now relies on, with an audit trail attached. | Owns AI spend and the approved-use policy, including privilege, retention, model-risk review and where sign-off is non-negotiable. |
| People & talent | Treats all AI-screened resumes and AI-drafted job ads as inherently unfair, without weighing the bias-audit and notice obligations either way. | Has tried AI for a job description once but still writes most people documents from scratch, with no sense of where it's actually safe to lean on it. | Uses AI to draft job ads and survey summaries, but keeps candidate scoring and other high-stakes assessment on a separate, more cautious track. | Builds role-specific AI work-sample assessments and manager calibration workflows that other functions adopt. | Owns workforce planning, role redesign and assessment validity as AI reshapes them, and sets the AI-fluency bar used in hiring itself. |
| Operations & support | Blocks approved AI-assisted triage or retrieval tools on principle, even after the risk has been scoped and signed off elsewhere. | Has used an AI chatbot to draft a reply once but still writes most tickets from scratch, with no fixed sense of when it helps. | Uses AI to draft replies and summarise recurring issues as a standing step, always edited by a person before anything goes out. | Runs an AI triage system with confidence thresholds, QA sampling and escalation rules, and tracks deflection against customer trust, not just speed. | Redesigns the support model around what AI can resolve end to end, with rollback paths and service tiers, freeing the team for judgement calls. |
Engineering
Resistant
Won't use an AI coding assistant; reviews AI-authored PRs from teammates with blanket suspicion rather than normal scrutiny.
Aware
Uses an AI coding chat for explanation or debugging now and then, but doesn't trust it on anything touching more than one file.
Applied
Runs an agentic coding tool daily for scoped tasks, tests and refactors, reviews every diff before merge, and never lets it merge unsupervised.
Advanced
Maintains the repo context, task templates and CI checks that make agent output consistently mergeable across the team, and catches the subtle bugs an agent introduces before they land.
Strategic
Sets policy on sandboxing, secrets, production access and dependency updates for agents, and tracks the quality and throughput numbers that justify it.
Product & design
Resistant
Insists every spec and mock is made by hand, on principle, even when a first draft would save a day.
Aware
Uses AI for rough flows or copy variants but can't turn the output into a testable product decision.
Applied
Uses AI to draft PRDs and synthesise interview transcripts with evidence traced back to source, then edits hard before anyone else sees it.
Advanced
Builds working prototypes in an AI app builder to test interaction and copy assumptions with real users before committing engineering capacity.
Strategic
Redesigns discovery around AI-assisted synthesis and measures the effect on decision quality and experiment velocity, not just speed.
GTM (sales & marketing)
Resistant
Refuses AI-drafted outreach on the grounds that it feels impersonal, without weighing the actual risk to trust or deliverability.
Aware
Has used AI to write a subject line or two but still drafts full sequences manually, with no read on what it's actually improving.
Applied
Runs AI-assisted account research and first-draft outreach as a standing pipeline step, with CRM hygiene and human personalisation before anything sends.
Advanced
Builds and maintains an AI lead-qualification agent wired into the CRM, monitoring false-positive rate and handoff quality, not just prompt tuning.
Strategic
Redesigns GTM around signal-based automation and human-led high-intent moments, measured against CAC and payback, not headcount saved.
Finance & legal
Resistant
Rejects even access-controlled, already-scoped AI use cases on blanket principle, without weighing the risk that was actually assessed.
Aware
Has asked an AI tool to summarise a document once but doesn't use it for anything that reaches a real decision.
Applied
Uses AI to draft first-pass variance commentary and flag anomalies, reconciled against source systems and materiality thresholds before it's trusted.
Advanced
Runs a reviewed AI workflow for contract redlining or month-end close checks that the team now relies on, with an audit trail attached.
Strategic
Owns AI spend and the approved-use policy, including privilege, retention, model-risk review and where sign-off is non-negotiable.
People & talent
Resistant
Treats all AI-screened resumes and AI-drafted job ads as inherently unfair, without weighing the bias-audit and notice obligations either way.
Aware
Has tried AI for a job description once but still writes most people documents from scratch, with no sense of where it's actually safe to lean on it.
Applied
Uses AI to draft job ads and survey summaries, but keeps candidate scoring and other high-stakes assessment on a separate, more cautious track.
Advanced
Builds role-specific AI work-sample assessments and manager calibration workflows that other functions adopt.
Strategic
Owns workforce planning, role redesign and assessment validity as AI reshapes them, and sets the AI-fluency bar used in hiring itself.
Operations & support
Resistant
Blocks approved AI-assisted triage or retrieval tools on principle, even after the risk has been scoped and signed off elsewhere.
Aware
Has used an AI chatbot to draft a reply once but still writes most tickets from scratch, with no fixed sense of when it helps.
Applied
Uses AI to draft replies and summarise recurring issues as a standing step, always edited by a person before anything goes out.
Advanced
Runs an AI triage system with confidence thresholds, QA sampling and escalation rules, and tracks deflection against customer trust, not just speed.
Strategic
Redesigns the support model around what AI can resolve end to end, with rollback paths and service tiers, freeing the team for judgement calls.
Read any row across and the ladder holds: each level asks for more accountability for what the AI actually produces than the one before it, whether that's a reusable workflow at Advanced or a governance call at Strategic. Read any column down and the function stays specific: what counts as Strategic in People & talent has nothing to do with what counts as Strategic in Operations.
Running it
The framework only earns its keep once it's operational, not just published. The shape that's worked for me:
Baseline first, self-assessment second. Have each person place themselves on the matrix before any training happens. The gap between self-assessment and manager assessment is itself useful data: it tells you where confidence and capability have drifted apart.
Calibrate across managers, not just within teams. Two managers' idea of "Applied" will diverge unless someone forces a shared read of the matrix. A short calibration session, comparing real examples against the cell definitions, closes most of that gap.
Aim enablement at the Aware-to-Applied gap. That's usually where the bulk of the organisation sits, and it's the gap training budget actually moves. People stuck at Aware don't need more tool access, they need one worked example specific to their function and a verification step to go with it.
Use it in hiring, the way Zapier does. A role-specific Applied-level bar in a job description does more than a generic "AI-savvy" line, because candidates can see exactly what's being asked of them.
What I'd change
The matrix needs a refresh cadence. Cell examples that felt Advanced eighteen months ago are Applied now, because the tools moved. Treat it as a living document, not a document.
Self-assessment inflates, especially at the Applied/Advanced boundary. Pair it with one real work sample per person rather than trusting the self-rating alone.
And the six functions here are a generic starting shape, not a fixed list. The most obvious gap is Security and IT: as AI agents get real access to systems and data, that function needs its own row, not a mention buried inside Engineering. The value of the framework is the method: name the levels, name the functions that matter to your company, and write concrete cell-by-cell examples of what good actually looks like. Copy the structure, not the content.