Anthropic's 4D AI Fluency model — what it gets right, and where it stops

Anthropic’s 4D AI Fluency model — Delegation, Description, Discernment, Diligence — captures the cognitive shift behind effective AI use better than most published frameworks. It sequences the behaviours a fluent user actually performs: deciding what to send to the model, framing the request clearly, judging the output against a quality bar, and verifying the consequential parts before they ship.

The model maps onto how we already sequence training. We use four cognitive metaphors, drawn from Nielsen’s Four Metaphors for working with AI:⁹ AI as an Intern (direct, review, correct), AI as a Coworker (parallel collaboration), AI as a Teacher (just-in-time learning), and AI as a Coach (critical-thinking partner). Anthropic’s 4D layer fits inside this sequence. Delegation and Description sit inside Intern and Coworker behaviours. Discernment is the Teacher stance. Diligence is the Coach posture.

So what’s the criticism. The 4D model is a cognitive-fluency framework. It describes what a fluent individual user does. It does not describe what happens when a fluent individual user sits inside an organization that hasn’t done the adoption work.

That distinction matters because most enterprise rollouts fail at the adoption layer, not the fluency layer. We see this on every engagement. The team trains well. Individual users get to Coworker or Teacher level inside three weeks. Then they go back to their teams and the workflow doesn’t change. The systems don’t change. The manager doesn’t ask a different question in the one-to-one. The fluent individual produces better drafts in private and revises the workflow back to its pre-AI shape in public.

What addresses that gap is not a different fluency framework. It’s a change-management layer alongside the fluency one. ADKAR tracks the individual readiness conditions every user needs before fluency converts into observable behaviour. Kotter’s 8-Step sequences the organizational moves that make the new behaviour stick. Bridges Transition Model names the psychological transition underneath. None of those layers are inside the 4D model — they’re not meant to be — and a rollout that trains for fluency without addressing them produces fluent users whose teams still don’t run on AI.

Two layers, both required. Cognitive-fluency training (Anthropic 4D, Nielsen four metaphors, Bloom’s, 70-20-10) addresses the individual capability. Change-and-adoption frameworks (ADKAR, Kotter, Bridges, Diffusion) address whether the capability becomes the team’s standard practice. We pair them on every engagement because the rollouts that don’t pair them are the rollouts that stall.

The 4D model is a strong piece of the puzzle. It’s not the whole puzzle.