The AI Fluency Framework for B2B Content Teams
- Harold Bell
- 6 days ago
- 12 min read
Key Takeaways
|
A content team gets access to a good model, output volume triples inside a month, and everyone agrees this is working. Two quarters later the same team is rewriting more than it ships, a client has caught a claim that was never true, and nobody can point to where the process broke.
In more than 16 years of building content programs for enterprise technology brands, I've watched that arc play out enough times to know it isn't a tooling problem. The team got faster at one skill and never developed the other three.
There's a framework that names all four, and once you have the vocabulary the diagnosis takes about ten minutes.
What is AI fluency
AI fluency is the ability to work with AI effectively, efficiently, ethically, and safely. It is defined by four human competencies rather than by tool knowledge, which is what makes it survive model releases that reset everything you knew about a specific system. |
The definition is deliberately not about the machine. It doesn't require you to understand transformer architecture and it doesn't require a prompt library. It describes what the person is doing.
How fluency differs from literacy
AI literacy is knowing what these systems are and roughly how they work. Fluency is the ability to actually get good work out of them. Literacy is knowledge, fluency is practice, and the gap between them is where most enterprise AI enablement programs die. You can brief a team beautifully on how language models function and still watch them delegate the one task that should never have left the building.
For content teams the stakes sit higher than for most functions, because our output carries a byline. A finance team using AI badly produces a bad internal forecast. A content team using AI badly produces a bad public artifact that gets indexed, cited, and quoted back at them. I've covered the strategic version of this in how AI changes content strategy for B2B, and the short answer is that the question moved from whether to use AI to where the human judgment has to sit.
What is an AI fluency framework
An AI fluency framework is a structured model for working with AI that names the distinct human competencies involved, so a team can diagnose which one is failing instead of concluding it's bad at AI generally. The dominant one is the 4D framework, built by Rick Dakan of Ringling College of Art and Design and Joseph Feller of University College Cork, with Anthropic partnering to produce a free course around it. |
The value isn't the taxonomy. It's the diagnosis.
Without a framework, a struggling team says something like our AI content isn't good enough, which points at nothing actionable. With one, the same team can say we're delegating work that should never have left the building, or our descriptions are fine but nobody's verifying output. Those two problems have completely different fixes, and telling them apart is most of the battle.
What the framework actually contains
Two layers. Four competencies, Delegation, Description, Discernment, and Diligence, which name what the human is doing. Then three modes, automation, augmentation, and agency, which describe the kind of working relationship you're in on a given task.
People remember the four and forget the three, which is a mistake. The modes are what tell you which competencies carry the weight. A one-shot formatting job and a multi-step agent run are not the same risk, and treating them the same way is how teams get surprised.
Why a framework beats a prompt library
Prompt libraries decay. Every model release resets a chunk of what you knew about coaxing good output from a specific system, and the library you built in March is folklore by September.
An AI fluency framework survives that, because it describes what the person is doing rather than what the machine needs to hear. Delegation logic doesn't change when a context window doubles. That durability is the entire argument for learning a framework instead of collecting tactics.
What are the four competencies
Delegation is deciding what to hand over. Description is communicating what you want. Discernment is judging what came back. Diligence is taking responsibility for the result. The framework was built by Rick Dakan of Ringling College of Art and Design and Joseph Feller of University College Cork, and Anthropic partnered with them to produce a free course around it. |
The framework has two moving parts, and people reliably remember one of them. The four competencies tell you which skill you're exercising. Three modes of working with AI, automation, augmentation, and agency, tell you what kind of working relationship you're in and therefore which competencies carry the weight.
Skip the modes and you end up applying the same process to a one shot formatting task and a multi step agent run, which are not remotely the same risk.
What should you actually delegate to AI
Delegate work where being confidently average is acceptable. Keep work where it isn't. That single test covers positioning, contrarian calls, original data interpretation, and the client relationship on one side, and structural drafting, schema generation, clustering, and repurposing on the other. |
Delegation happens before you type anything. What am I trying to achieve, which parts of this need me, and which parts don't. The framework asks for a pause almost nobody takes, because the tool is open and typing into it is free.
We use the Engine Optimization Matrix as our delegation map. Four engines, SEO, AEO, GEO, and LLMO, crossed against five levers: content, schema, distribution, authority, and citation. Twenty cells, and knowing which ones survive a handoff is most of the job. If the underlying system isn't built yet, start with a content strategy framework for B2B growth and layer the delegation decisions on top.
What delegates cleanly
Structural first drafts where the outline is already decided
Schema generation once the spec is written and the entity rules are fixed
Keyword clustering and gap comparison across a large set
First pass FAQ drafting from an existing body of content
Repurposing a long asset into shorter formats
What never delegates
Positioning, because it rests on a market read the model doesn't have. The contrarian call, because a model optimizes toward consensus by construction. Original first party data and what it means. And the client relationship, along with every judgment call inside it.
Here's a live example of the second one. We hold a deliberately contrarian position on llms.txt, because the evidence points to it correlating with noise rather than citation signal for answer engines, and the confirmed use case sits with coding agents. That call came from reading our own data against the consensus. Ask a model what to think about llms.txt and you get the consensus back, fluently, in seconds. The reasoning is in what is llms.txt and does it actually work, and the position exists because a human made it.
How do you describe work well enough for a model to execute it
Description means specifying three things: the product you want, the process you want followed, and the performance you want from the model as a collaborator. Teams that get the most out of it write those specs down as reusable artifacts instead of retyping instructions every session. |
Description is the competency most people collapse into the phrase prompt engineering, which undersells it badly. Prompting is the technique layer. Description is the discipline the technique serves.
Turn description into artifacts
The highest leverage move we made was moving description off the keyboard and onto disk. Our house style sheet is a Description artifact. It says no colons in H2 headings. It says the engine order is always SEO, AEO, GEO, LLMO. It says use the percent symbol rather than the word. None of that is clever prompting, it's a specification that gets reused, which is why it compounds instead of evaporating at the end of a session. For the technique layer underneath, prompt engineering for content marketing and few-shot prompting for marketing go deeper, and role prompting handles the performance side specifically.
Fetch before writing
Our second rule is that nothing gets drafted against an assumption about what a page currently says. We pull the live content first, every time, without exception. That rule exists because the most confident wrong output we ever produced came from a model reasoning fluently about a page it had never actually seen. Voice belongs in this category too, and most teams try to solve it at the editing stage instead. Brand voice AI covers why that's backwards.
Why is discernment the competency content teams skip
Because fluent and accurate are independent properties, and reading fluent prose makes them feel like the same thing. Language models are optimized to produce text that reads as confident and contextually appropriate. Accuracy is a separate property that has to be checked separately. |
Discernment is judging what came back. Not whether it reads well. Whether it's right.
The schema mistake we made
For a long stretch we hand injected BlogPosting JSON-LD into individual blog posts. That's what the consensus advice says to do, it's what a model will tell you to do, and every output we got looked correct. It was wrong. Our Wix template already emits BlogPosting site wide, so we were double emitting the entity on every single post.
Nothing in the output signalled the problem, because the problem wasn't in the output. It was in a fact about our own stack that neither we nor the model had checked. The corrected spec is now narrow: per post JSON-LD is FAQPage only, plus page specific types the template doesn't cover. That correction came from Discernment, not from better prompting. Both schema markup examples and FAQPage schema reflect the corrected approach.
The description discernment loop
The framework pairs these two competencies deliberately. You judge what came back, then feed that judgment forward into a sharper description rather than regenerating and hoping. Most people regenerate. Regenerating is not the loop, and the difference shows up as iteration that goes sideways for an hour instead of converging in three passes. At library scale, AI content audit walks through applying the same judgment across an entire archive rather than one draft at a time.
What does diligence look like when you publish under your own name
Diligence covers three things: being thoughtful about which systems you use and how, being transparent about AI's role with anyone who needs to know, and verifying the outputs you publish. For content teams the third one is the sharp end, because the byline does not move whatever the model contributed. |
A client who gets burned by a wrong claim in a published asset is not going to find the provenance interesting. That's also why transparency is a credibility asset in B2B rather than a liability. Being straight about where AI sits in your process reads as process maturity to a technical buyer.
I've written up the failure modes at length in the 7 deadly sins of AI-assisted content, and nearly every one of them is a Diligence failure wearing a different costume. Will AI replace marketers is the longer answer to the question sitting underneath all of this.
How do the four competencies map to automation, augmentation, and agency
Automation loads Delegation and Description. Augmentation loads Description and Discernment. Agency loads Delegation and Diligence, because a system running multiple steps on its own gives you far fewer chances to intervene mid process. |
Mode | What it looks like | Heaviest competencies |
Automation | You define a task, the system executes it. Formatting, conversion, routine generation. | Delegation and Description |
Augmentation | You and the system work together iteratively as thinking partners. | Description and Discernment |
Agency | The system works with meaningful independence across multiple steps. | Delegation and Diligence |
Most professional content work lives in augmentation, which is exactly where the description discernment loop does the heavy lifting. But agency is growing fast, and it shifts the weight back to the bookends. When you can't inspect the middle, the quality of your upfront delegation and your downstream verification is doing all of the work.
What happens when one competency is weak
The chain breaks at the weakest link. Perfect prompting cannot rescue a bad delegation decision, brilliant delegation produces nothing without a clear description, and the most accurate output in the world still creates exposure if nobody takes responsibility for it. |
Weak competency | What it produces | How it shows up in content |
Delegation | Work handed over that should never have left | Generic positioning, consensus takes, no distinctive point of view |
Description | Slow, frustrating iteration | Heavy rewriting, output that never quite matches house voice |
Discernment | Polished mistakes at speed | Confident wrong claims, invented statistics, broken technical guidance |
Diligence | Unowned work | No verification trail, credibility damage when something is caught publicly |
The most expensive profile
Row three is the one that costs real money. Strong Description paired with weak Discernment is the worst combination a content team can have, and it's also the most common, because writing clear instructions is adjacent to what content people already do well. Everything looks right. The velocity looks like progress. The correction cost arrives later and lands in public.
What should most B2B content teams actually do
Frameworks that stay conceptual change nothing. Here is how we operationalized it.
Write the delegation map first
List every task in your content operation and mark each one human, AI assisted, or AI led. The argument about the boundary cases is the actual work, and it's worth having out loud with the whole team in the room. Most of the value shows up in the twenty minutes of disagreement, not in the finished document.
Give discernment its own step
Not a vibe check at the end of editing. A distinct pass where the only question is whether the claims are true, run by someone who didn't write the draft. Then name one owner per published asset who verified it and stands behind it. On the tooling layer beneath all of this, AI marketing tools and workflows covers the stack and AI for B2B content marketing is the practical application guide.
Re-check your own priors quarterly
Our schema spec was wrong for months because nobody went back and asked whether the original decision still held. Set a recurring review over the decisions you made confidently a year ago, because those are the ones nobody re-examines. That discipline matters more than it used to: we've confirmed AI Overview citations on pages carrying zero organic ranking, which tells you answer engines are reading closely and judging on substance rather than position. How to show up in AI Overviews when you don't rank on Google has the data, and how to get cited by AI covers what makes a passage quotable in the first place.
What's next
If you can't say out loud which parts of your content operation are AI led and which are human only, that's the gap, and it's worth an afternoon. We build the delegation map with B2B technology clients as a matter of routine, and it surfaces more than the framework alone suggests it will.
Book a 30 minute working session and we'll walk your delegation map together.
Frequently asked questions
What is the 4D framework for AI fluency
Four competencies for working with AI: Delegation, Description, Discernment, and Diligence. It was developed by Rick Dakan of Ringling College of Art and Design and Joseph Feller of University College Cork, and Anthropic partnered with them to build a free course around it.
What is the difference between AI literacy and AI fluency
Literacy is knowing what AI systems are and roughly how they work. Fluency is the ability to actually get good work out of them. Knowledge versus practice. You can be highly literate about model architecture and still delegate badly.
Is the 4D framework the same as prompt engineering
No. Prompt engineering sits inside one competency, Description. The framework's argument is that prompting is roughly a quarter of the skill set, and most AI-assisted work fails on the other three rather than on the wording of the prompt.
Which competency do content teams get wrong most often
Discernment. Content teams are usually strong at Description because writing clear instructions is adjacent to what they already do. That strength becomes a liability when it isn't paired with real verification.
What are automation, augmentation, and agency
The three modes the four competencies map across. In automation you define a task and the system executes it. In augmentation you iterate together as thinking partners. In agency the system operates with meaningful independence across multiple steps. Most content work sits in augmentation.
What is the description discernment loop
You evaluate what the model returned, then feed that judgment forward into a sharper description rather than simply regenerating. Regenerating without adjusting the description is not the loop, and it's what most people do by default.
Does using AI in content production hurt your citation performance
Not inherently. What hurts is unverified, undifferentiated content, which is a Discernment and Delegation failure rather than an AI problem. We've confirmed AI Overview citations on pages with zero organic ranking, which suggests answer engines judge substance rather than provenance.
What should never be delegated to AI in content marketing
Positioning, contrarian calls, original first party data and its interpretation, and the client relationship. All of them fail when the output is confidently average, and confidently average is what a model produces by construction.
How long does it take to build AI fluency across a content team
The framework takes an afternoon to understand. Operationalizing it takes a quarter, because the work isn't learning definitions. It's writing down the delegation boundaries and description artifacts your team has been carrying informally in people's heads.
Where can I take the AI fluency course
Anthropic hosts it free through its academy, developed with Rick Dakan and Joseph Feller. Course materials carry a Creative Commons BY-NC-SA licence, so check the terms before reusing any of it in commercial work.
Comments