AI Fluency: The Four Competencies That Decide Whether AI Works for Your Company — or Against It
Published on 03 September 2026 · Last updated on 07 September 2026
A sentence I hear in almost every Discovery: "Everyone's already using it somehow." That's exactly where the problem starts. Somehow is not a competency — it's chance, with your own colleagues as the test subjects.
What's actually missing in most companies is rarely the will to use AI sensibly. It's a shared vocabulary for what working well with AI even means. That's exactly what the AI Fluency model by Joe Feller and Rick Dakan, developed in partnership with Anthropic, provides — four competencies I now use in every engagement on the topic of enablement: Delegation, Description, Discernment, Diligence. What you're allowed to delegate. How you instruct. How you check results. And who's ultimately accountable.
Four terms, one system: only when all four work together does AI use become a competency — instead of a loss of control everyone saw coming. Below, each tile links straight to the matching section.
The four competencies at a glance
Delegation: Which Tasks Can AI Take On in Your Company — and Which Can't It?
AI can take on what can be clearly bounded, whose mistakes you can afford, and for which a human is still ultimately accountable at the end. Everything else stays with people for now. In the AI Fluency model by Joe Feller and Rick Dakan, developed in partnership with Anthropic, this distinction is called "Delegation" — the first of four competencies that decide whether AI works for a company or creates chaos.
Delegation isn't a gut call — it's a leadership task. In practice, a simple three-way split works well: uncritical tasks (internal draft texts, for example), where a mistake has no real consequences; conditionally suitable tasks (working with customer data, for example), which need a review loop; and tasks that stay with humans as a matter of principle — automated decisions without oversight, for instance. That exact classification is one of the four building blocks of my AI Governance Discovery: the risk map.
What's missing in practice is rarely the will to delegate — it's a clear rule for who is allowed to delegate, and on what basis. According to McKinsey, two out of three companies name exactly that, security and risk questions, as the biggest obstacle to scaling AI — not the technology.
Description: Why the Instruction Decides the Quality of the AI Result
An AI result is rarely worse than the instruction it was given. Vague instructions produce vague results — and the responsibility for that still lies with the human who approves them. In Feller and Dakan's AI Fluency model, this skill is called "Description": describing clearly enough what you want, so AI can deliver what you actually need.
For a company without its own legal or IT department, that means, concretely: Description isn't about prompting tricks — it's about expectations set down in writing: on tone, on sources, on what must never be invented. Those exact expectations belong in the AI ground rules I draft with leadership teams in the AI Governance Discovery: a compact policy draft that fixes responsibilities and standards, instead of leaving them to individual employees' judgment.
Description and Discernment — the next competency — belong together: clear instructions produce better results, and checking those results is how you learn whether the instruction was actually good. Together, the two form the basis for working productively with AI day to day.
Discernment: How to Tell Whether an AI Result Is Usable — Before It Reaches the Customer
An AI result only becomes usable once someone with real expertise — and real time pressure — has actually checked it, not once it sounds convincing. Feller and Dakan's AI Fluency model, developed with Anthropic, calls this ability "Discernment": recognizing what's accurate, what's missing, and what's too smooth to be true.
For someone with a journalism background, this isn't a new concept — it's daily practice: nothing gets published that hasn't been checked. That exact reflex is what many teams lack when working with AI — not out of carelessness, but because no one has defined who checks, what they check for, and where the line runs between "good enough for an internal draft" and "ready to go to the client."
Without that line, what happens is what companies with weak governance report time and again: AI results get approved too quickly because no one is officially responsible for holding them back. That exact responsibility — who decides, who escalates — is part of the AI ground-rules draft the AI Governance Discovery delivers.
Diligence: Who's Liable When AI Makes a Mistake — and How to Prepare
When AI makes a mistake, the AI isn't liable — the company is, usually its leadership. That's exactly why "Diligence," the fourth competency in Feller and Dakan's AI Fluency model, developed with Anthropic, isn't optional polish — it's the competency that underwrites all the others: staying accountable for what you've delegated.
Since February 2026, the EU AI Act has made this formally explicit too: under Article 4, responsibility for AI use cannot be delegated — not to IT, not to a tool, and certainly not to the AI itself. What that means for your company specifically is a question for your legal counsel; I don't replace that. What I deliver is the organizational foundation legal counsel can actually build on: named responsibilities, documented decision paths, a review process that can show, if it ever comes to that, that due care was taken.
In the end, Diligence means this: if it ever comes to that, you can show who decided what, and when. That's exactly what the prioritized action list at the end of the AI Governance Discovery delivers.