4D AI Fluency Framework: How to Govern AI Agents Safely
Use four clear decisions (what an agent may do alone, how its limits are written, how its work is checked and who owns it) to deploy AI agents with confidenc…
Topic: AI/SI Adoption
Why It Matters
The first AI agent in a team usually arrives quietly. Someone connects an assistant to the shared inbox, or an agent starts drafting supplier replies, reconciling invoices or updating the CRM overnight. It saves hours. Then one day it sends something it should not have, or changes a record nobody can trace, and the first question in the room is: who approved this?
That question is the gap the 4D AI Fluency Framework helps close. It gives managers four plain-language lenses for deciding what an agent may do on its own, how its boundaries are written down, how its work is checked and who answers for it. Applied as a governance model, it turns ad hoc experimentation into a deliberate operating arrangement, without the weight of a full enterprise AI policy.
It matters now because agents are moving from suggesting to acting. When AI only drafted text, a person was always the final step. When it can send, book, approve or update, the controls have to be designed in advance. Managers who can show that their agents are governed earn the trust to deploy more of them.
A few questions can tell you whether this is worth your time:
Is an AI agent in your area taking actions (sending, updating, approving) rather than only drafting for a person to review?
If an agent made a costly error tomorrow, could you name, in one sentence, the person accountable for it?
Are team members setting up their own agents faster than anyone is writing down what those agents are allowed to do?
Origin & Underlying Logic
The 4D AI Fluency Framework was developed in 2025 by Professor Rick Dakan of Ringling College of Art and Design and Professor Joseph Feller of University College Cork, in partnership with Anthropic.
It defines AI fluency as working with AI in ways that are effective, efficient, ethical and safe, built on four interconnected competencies: Delegation, Description, Discernment and Diligence.
It was designed first as a set of individual skills for anyone using AI. It also distinguishes three modes of interaction: automation (AI carries out a defined task), augmentation (AI acts as a thinking partner) and agency (a person configures AI to act independently on future tasks).
This guide focuses on the third mode and treats the four competencies as a governance model for a team or function. That is an extension of the original intent, not a replacement for it.
The underlying logic is simple: an agent failure can usually be traced back to one of four decisions made badly. The wrong work was handed over, the boundaries were described poorly, the output was not checked properly, or nobody owned the result.
The framework sits alongside established thinking such as the RACI matrix for accountability, the four eyes principle in financial controls and formal AI risk management standards. Its advantage is that it is light enough for a team manager to use without a compliance department.
Structure & Application
The four components work as a cycle rather than a checklist. Each one answers a different governance question, and a weakness in one puts pressure on the others.
Delegation: what may the agent do alone? This is the decision about which tasks to hand over and at what level of autonomy. It exists because not all work carries the same risk. Drafting a meeting summary and issuing a customer refund both look like admin, but the cost of an error is very different. Delegation sets the autonomy level for each task: suggest only, act with approval, or act and report.
Description: how are the boundaries written down? An agent can only stay inside limits it has been given. Description covers the output expected (product), the approach the agent should follow (process) and how it should behave, including when to stop and escalate (performance). It exists because unwritten expectations do not survive contact with an automated system.
Discernment: how is the work reviewed? This is the review layer: checking the quality of outputs, how the agent reached them and whether it stayed within its role. It exists because a fluent, confident output is not the same as a correct one, and agent errors tend to repeat at scale before anyone notices.
Diligence: who owns each agent? Every agent needs a named human owner who answers for what it does, a record of which tools and data it uses, and openness with the people affected by its work. It exists because accountability cannot be delegated to software.
The interactions matter more than the parts. Generous Delegation demands sharper Description and heavier Discernment. Weak Description shows up later as review workload, because reviewers are left to catch what was never specified. Diligence holds the cycle together: an owner is the person who notices when the other three have drifted.
In practice, the tool usually takes the form of an agent register: a simple table with one row per agent and a column for each D. Some teams add an autonomy ladder showing the permitted level for each task. The 4D AI Fluency Framework works best at team level first, where the manager knows the work well enough to judge risk, and scales to organisational level once several teams share the same register format.
Quick-Start
Before reading further, list every AI agent or automated assistant currently acting in your area, and write one name next to each as its owner. If any row stays blank, you have found your first governance gap. The harder question, which the deep dive answers, is how much autonomy each of those agents should actually have.
FAQ
Q: What is the 4D AI Fluency Framework?
A: The 4D AI Fluency Framework is a model developed by Rick Dakan, Joseph Feller and Anthropic that describes four competencies for working with AI: Delegation, Description, Discernment and Diligence. It was created to help people use AI effectively, efficiently, ethically and safely. Applied to AI agents, it becomes a lightweight governance model covering autonomy, boundaries, review and ownership.
Q: Is the framework only for individual AI users?
A: It was originally designed as a set of individual skills, and that remains its core purpose. Its four questions translate directly to team-level governance, though, especially for agents that act independently. A manager can use the same vocabulary to agree with their team what each agent may do, how it is instructed, how it is checked and who owns it.
Q: How does it differ from a formal AI risk management standard?
A: Formal standards are built for organisation-wide risk programmes, with defined processes, documentation and audit trails. The 4D framework is lighter and more practical: it gives a team manager four questions they can answer this week. The two complement each other well, and a team's 4D register can feed an organisation's formal risk process rather than replace it.
Q: Does this framework work for small teams with only one or two agents?
A: Yes, and small teams can benefit most because they set good habits before agent use spreads. With one or two agents, the whole framework fits on a single page. The value lies less in paperwork and more in forcing a clear decision about autonomy and ownership before an error makes that decision for you.
Q: How do I know if this framework is right for my situation?
A: It is a strong fit if an AI agent in your area takes actions rather than only drafting, and if accountability for its work is unclear. It is less useful if your team uses AI purely as a writing or research assistant with a person reviewing every output; the individual-skills version is likely enough there. Setting the right autonomy level is covered in the deep dive.