The 4D Framework — Anthropic's own playbook for human-AI work (and why most AU firms don't use it by name)

In April 2025 Anthropic, Rick Dakan (Rollins College), and Joseph Feller published a free, open-licensed course called AI Fluency: Framework & Foundations. Over fifteen short video lessons, it teaches what is probably the single most useful framework in enterprise AI right now — and almost no AU consultancy uses it by name.

This post is the short version. If you're considering AI investment in 2026 and read only one piece this week, read this one.

The framework in one diagram

Every human-AI interaction sits on two axes.

Axis 1 — the mode you're engaging in (three options):

• Automation — AI executes a specific task on your precise instructions. Scripted, repeatable, no surprises.
• Augmentation — you and AI work turn-by-turn as collaborators. Claude Code pairing, chat-assisted research, drafting-plus-review.
• Agency — you shape AI's policies and knowledge so it works independently on your behalf. Configured agents, Managed Agents, Routines.

Axis 2 — four competencies you apply across all three modes (the 4Ds):

• Delegation — deciding what work to do with AI vs yourself
• Description — communicating effectively with AI systems
• Discernment — evaluating AI outputs and behaviours critically
• Diligence — ensuring responsible AI collaboration

That's the whole model. Three modes. Four competencies. A 3×4 matrix, 12 cells. Every real AI decision you face — "should we automate report generation?", "how do we trust this agent?", "what do we disclose under Privacy Act 2026?" — fits somewhere on it.

Why the two-axis model matters

Most AI failures in 2025 came from mode-mismatch — trying to operate in Agency mode before mastering Delegation. Businesses bought managed-agent retainers for tasks they hadn't even specified clearly. They delegated nothing correctly, so the agent did things they hadn't actually asked it to do, and then they blamed the agent.

The 4D framework's quiet insight: the competencies are the same regardless of mode. Whether you're automating one task, collaborating with Claude on a draft, or running an autonomous agent on a policy, you still need to Delegate well, Describe clearly, Discern outputs rigorously, and hold Diligence. The mode changes; the discipline doesn't.

Most AI consultancies pitch Agency mode because Agency mode has the most revenue per client. The framework cuts through that: you should start in Automation, prove the 4Ds on one task, and only ladder up to Augmentation and Agency as trust is earned. Which is exactly how we run our own consulting practice.

The four Ds, expanded

1. Delegation — deciding what to do with AI

Has three components:

• Problem Awareness — clearly understanding your goal before involving AI
• Platform Awareness — knowing what different AI systems can do and can't do today
• Task Delegation — strategically dividing work between you and AI

The most common failure here is Platform Awareness. Teams read blog posts about what Claude "will" do and base decisions on the roadmap rather than the current capability. Opus 4.7 has a 1M-token context window but a 128K output cap. You can hand it a 400-page document and ask questions; you cannot ask it to produce a 400-page response.

2. Description — communicating how with AI

Three components:

• Product Description — clearly defining what you want the AI to create
• Process Description — guiding how the AI approaches your request
• Performance Description — defining how you want the AI to behave during collaboration

The Description competency is where most prompt-engineering advice lives. The framework's addition: it's not enough to describe the product (what you want made). You also have to describe the process (how to go about it) and the performance (how to be during the collaboration — tone, persona, pushback style).

3. Discernment — evaluating what you got

Three components:

• Product Discernment — evaluating output quality (accuracy, appropriateness, coherence, relevance)
• Process Discernment — assessing how the AI arrived at its output (logical errors, attention gaps, tool misuse)
• Performance Discernment — evaluating how the AI behaved during the interaction

Discernment is the flip side of Description. You describe, the AI responds, you discern, you re-describe. This loop is explicitly named the "Description-Discernment loop" in the course — and naming it is more useful than you'd think. Once you've named it, it's easier to diagnose when something goes wrong. A persistently bad output usually means a Description problem. A weirdly bad process usually means a tool or context problem.

4. Diligence — taking responsibility for the collaboration

Three components:

• Creation Diligence — thoughtfulness about which AI systems you choose and how you engage with them
• Transparency Diligence — being honest about AI's role in your work with everyone who needs to know
• Deployment Diligence — taking ownership for AI-assisted outputs you ship

Under AU's Privacy Act 2026 reforms (ADM disclosures commence 10 December 2026), Diligence stops being optional best-practice and starts being a legal requirement. If your business uses AI for any decision materially affecting a person — hiring, insurance, credit, health, housing, government, professional judgment — you need to disclose it, document the DPIA, and own the outcome.

The 3×4 matrix — every cell matters

When you map any concrete AI decision onto this matrix, the weak cells stick out. If your pitch to a boss is "we're going to deploy an autonomous agent for compliance reporting," the first questions the matrix asks you are:

• Agency-Delegation: what policies govern it?
• Agency-Description: have you defined its constitution?
• Agency-Discernment: how will you monitor its behaviour?
• Agency-Diligence: who owns the output when it fails?

If you can't answer any of those, the answer isn't "deploy in Agency mode." The answer is "drop back to Automation, do one bounded task, learn where the edges are, then ladder up."

How we use this at Adaptation AI

Before a client commits a dollar, we map their problem onto the 4D × 3-mode matrix. If any cell is weak, the engagement gets smaller, not bigger. Our 48-hour Paid Assessment (Rung 4 of our product ladder) is a deliberate Automation-mode exercise — one real task, precise instructions, measurable output. That's it. If the 4Ds hold on one task, we ladder up to Rung 5 (Automation Workflow Build) and then Rung 6 (Augmentation Pod) and only later Rung 7 (Agency Managed Retainer).

Clients who want to skip ahead — who want the managed-agent retainer before proving Automation — get told no. That rule exists because we've seen it fail the other way, too many times.

Further reading

• The course itself (free, ~90 minutes): anthropic.skilljar.com/ai-fluency-framework-foundations
• Dakan and Feller's canonical framework site: https://aifluency.ai
• Anthropic's public lesson library: https://www.anthropic.com/learn
• The companion AI Fluency variants for educators, students, and nonprofits also published by Anthropic

What to do now

If you've read this far and one of these is true:

• You're running an AI pilot that's stalling
• You're considering an AI investment but don't know where to start
• You're wondering whether your use case is in scope for Privacy Act 2026 reforms

Do one of:

1. Read the framework in depth — the course is free, 90 minutes, worth the time
2. Use it as a decision filter — on any AI investment, ask the four Ds for each mode
3. Book a 48-hour Paid Assessment — adaptation.ai/assessment — if you want to see what the framework looks like applied to one of your real tasks

— Willie Prosek, Founder, Adaptation AI

---

Framework content © 2025 Rick Dakan, Joseph Feller, and Anthropic. Released under CC BY-NC-SA 4.0. This blog post cites the framework by name and under fair-use quotation. Full course material remains the authors' copyright.