What Is AI Fluency? It Shows Up Between the Prompt and the Finished Work
AI literacy describes what someone understands. AI fluency becomes visible in the choices they make while turning that understanding into useful, verified and accountable work.
Two Prune Research··4 min read
AI fluency is the demonstrated ability to turn AI knowledge into sound work: frame the task, find the evidence, direct AI, check the result, exercise judgment and deliver something another person can use.
Imagine two managers answering the same question: should a supplier contract be renewed?
Both know that generative AI can invent facts. Both can write a detailed prompt. Both receive a polished recommendation. One manager forwards it. The other notices that two performance reports cover different periods, checks the contract against the source documents, removes a confidential term from an unapproved tool and tells the decision-maker what remains uncertain.
Their difference is not AI vocabulary. It is what happens between the prompt and the finished work.
AI literacy is the foundation
AI literacy deserves a fair definition. The OECD and European Commission describe it as knowledge, skills and attitudes that help people understand AI, critically evaluate outputs and use it ethically and creatively.[1] Skills England similarly treats foundation skills as a combination of technical, non-technical, responsible and ethical capabilities for using simple AI tools at work.[2]
A literate user should understand that models can be wrong, know what information is sensitive, recognise where human responsibility remains and have enough practical knowledge to interact with a tool. Those capabilities matter. The mistake is assuming that knowledge alone shows how someone will handle a consequential task.
AI fluency is performance in context
At Two Prune, AI fluency means demonstrated performance in AI-enabled work. It is contextual: someone may be fluent in one class of task and less prepared in another. It is observable: conclusions should be tied to what the participant reviewed, asked, changed, checked and delivered. And it is bounded: evidence from one simulation is not a permanent label or proof of performance everywhere.
We organise that performance into six connected pillars:
Context Framing: understanding the audience, constraints, decision and definition of success.
Evidence Navigation: finding, weighing and using the information that matters.
AI Orchestration: decomposing the work, directing AI, iterating and deciding what should remain human-led.
Two Prune is beginning a research programme—and inviting academic collaboration—to develop task-bounded AI-fluency assessment into a responsible measurement method.
Judgment & Synthesis: combining evidence and trade-offs into a defensible recommendation.
Communication & Delivery: producing a clear, useful deliverable for the intended audience.
A polished answer can hide an incomplete workflow
Prompt quality matters, but it is only one transition in the work. A strong prompt can still produce an answer that cites the wrong period, carries an unsupported assumption or solves the wrong problem. Confidence is not proof either. NIST's AI Risk Management Framework emphasises context, measurement, documentation and defined human oversight rather than blind acceptance of system output.[3]
Fluency appears when someone knows which role AI should play at each stage: researcher, synthesiser, critic, calculator, editor or verification aid. It also appears when they know when not to use it, what must be checked independently and which uncertainty needs to survive into the final deliverable.
How can an organisation assess AI fluency?
A useful assessment needs to create opportunities for the relevant behaviour to appear. Give someone a realistic objective, imperfect evidence, appropriate AI support and a deliverable with a credible audience. Then examine the submitted work alongside the available process evidence.
That does not prove universal productivity or predict every future task. It produces task-bounded evidence about what the participant appeared to identify or address in the recorded work, the AI use and verification the available evidence establishes, what was missing and what they ultimately delivered. The report explains its score methodology, evidence availability, any missing ratings and the limits on interpretation. If evidence remains materially insufficient after bounded recovery, Two Prune withholds the overall score and produces a score-free limited-evidence report. Missing capture is never treated as poor performance or zero.
