Framework
There is a clear, well-regarded answer to what it takes to work well with AI. The catch for a firm is that the answer was written for one person at a time.
The most useful account of what AI fluency actually requires comes from the AI Fluency framework, developed by Professor Rick Dakan of Ringling College and Professor Joseph Feller of University College Cork, and published through Anthropic's AI Fluency work. It names four competencies, the four Ds: delegation, description, discernment, and diligence.
In plain terms: delegation is deciding what work goes to a person and what goes to the model; description is communicating clearly enough to get useful output; discernment is judging whether what comes back is actually good; and diligence is taking responsibility for how the work is done and what goes out the door. It is the best short answer we have seen to a question most firms never quite frame: not "which tool," but "what does it take to use one well."
There is one thing to keep in mind about it, and it changes everything for a firm. The framework describes the fluency of an individual working with AI. That is the right unit for a course or a single professional. It is the wrong unit for a law firm.
Leave the four Ds to each lawyer and you get exactly the picture inside most firms today: everyone improvising their own division of labor, their own prompts, their own sense of what is good enough, their own private rules about what is safe to paste in. Some are excellent at it. Most are not, and no one can tell which is which from the outside. The result is inconsistent work product, exposure on confidentiality and privilege, and effort spent without efficiency gained.
The fix is not to train every lawyer to personal fluency and hope it converges. It is to build the four Ds into the firm itself, once, as method rather than talent, so the same standard holds whoever is at the keyboard. That reframing is the whole job, and it is worth taking the four in turn.
Delegation asks which parts of the work stay with a lawyer, which go to the model, and how the two hand work back and forth. For an individual it is a running judgment call. For a firm it has to become a settled, documented division of labor, by matter type, so the line sits in the same place whether the matter is in front of a partner or a second-year.
This is the hardest of the four to build from the inside, because the answer is specific to your matter mix, your risk tolerance, how partners are compensated, and where associates learn. We have written separately on why delegation is the one firms tend to outsource. For an implementation, the point is simpler: the line has to be drawn deliberately and written down, not left to each person's instinct.
Description is the skill of telling the model what you want well enough to get something useful back. Left to individuals, it is a thousand private styles, relearned every time someone new joins. Built for the firm, it becomes shared property: prompts, instructions, and templates grounded in the firm's own house style, precedent, and standards, so the system is briefed by the firm's accumulated craft rather than by whoever happens to be typing.
This is also where a firm protects what makes it distinctive. A general tool on default settings produces competent, anonymous work that looks like everyone's. Describing the work in the firm's own terms is how the output starts to look like the firm's, from the first draft rather than after heavy rewriting.
Discernment is judging whether the output is actually good, and it is the competency most quietly assumed. The evidence says it should not be. AI lifts straightforward, breadth-style work substantially, and degrades work that turns on judgment and synthesis: in one controlled study, professionals using AI on judgment-heavy tasks were materially less likely to reach the right answer than those without it, and studies of legal work show the same split, with confident prose masking missed issues.
For a firm, discernment cannot be a private feel for when something is off. It has to be an explicit standard, what "good enough" means for each matter type, who reviews what, and which tasks the model is not allowed near, so that quality does not depend on which lawyer happened to catch the error. In practice that means evaluating the tool against the firm's own past work on a matter type before trusting it there, so the bar rests on evidence rather than impression.
Diligence is taking responsibility for how AI is used and what it produces. Whatever goes out under the firm's name is the firm's: the accuracy, the ethics, the consequences do not transfer to the model. For an individual that is a matter of conscience. For a firm it is governance: a usage policy, confidentiality and privilege handling, a posture on client disclosure, alignment with the ABA Model Rules, and a named owner who keeps all of it current as the tools and the rules change.
This is the competency ad hoc use most reliably fails, and the one with a malpractice tail. It is also the one that most clearly belongs to the institution rather than the individual.
Read the four together and a sequence falls out. You cannot describe the work well until you have decided what to delegate. You cannot set a discernment bar until you know what the work is for. And diligence has to wrap all of it. That is the same order the work takes when a firm does it properly: codify how the firm actually practices first, then fit, brief, and govern the tool to that. The four Ds are a clean way to see why practice has to come before platform, and why fluency built into the firm, not scattered across its people, is what an implementation is actually for.
← Back to insights See the steps, in order →Personal fluency comes and goes with whoever is at the keyboard. We build the standard into how the firm practices, so it holds across the firm. A discovery call is a good place to start.
Schedule a discovery call →
Codified Counsel