Ask ten leaders whether they run an AI-fluent company and nine say they are getting there. Ask what a visitor could walk in and inspect to check, and the room goes quiet. That gap is the problem with how AI enablement gets assessed. Fluency is treated as a mood, measured by asking people how they feel about tools they have had for a year. Section’s 2026 AI Proficiency Report put 5,026 US knowledge workers through scored tasks instead of a survey and found 5.5 percent met its bar for proficient, against 73.5 percent it classed as experimenters. The better question is structural: which parts of a company exist when fluency is real.
What fluency means once you stop asking how people feel
The clearest definition on offer comes from Anthropic’s AI Fluency course, written with Rick Dakan and Joseph Feller and released under an open license. It names four competencies: Delegation, deciding whether and how to hand work to a model; Description, saying what you want well enough to get it; Discernment, judging whether what came back is any good; and Diligence, owning the result.
Read those four and the split between literacy and fluency stops being semantic. Literacy is knowing what the tool is. Fluency is a habit of judgment under time pressure on work that carries your name. The framework describes that well at the level of a person, and there it stops. It does not say what has to be true of a 180-person company for that judgment to survive a busy quarter.
Enthusiasm does not close that gap on its own. BCG’s 2026 AI at Work survey of close to 12,000 employees, managers, and leaders found 74 percent of frontline staff using AI daily or a few times a week, up 23 points in a year, while only 36 percent said their upskilling had been adequate. Access arrived and capability was left to chance, the same split that separates a license invoice from a capability budget.
The six parts where AI enablement shows up
Section’s study did not only score people. It compared those scores against features of the companies they worked in. Tested proficiency averaged 44 out of 100 where an agent rollout was paired with training specific to it and 25 where it was not, 43 against 26 where a formal AI strategy existed, and 43 against 25 where people had clear access to sanctioned tools. These are associations in a cross-sectional sample, not proof of cause, and 44 out of 100 is still a failing grade. They are also the closest thing this field has to an evidence-based parts list.
Six parts follow from that data and from what breaks in practice, each with an object you can look at.
| Part | What to inspect | Signal it is missing |
|---|---|---|
| An owner | A name on a calendar with hours reserved | A committee, a channel, nobody’s second job |
| Rebuilt workflows | A named task whose written steps changed | The old steps plus a chat window |
| A written layer | Context docs, saved prompts, review rules | Knowledge living in four power users’ heads |
| Managers who ask | A team ritual where the method gets shown | Encouragement with no follow-up |
| Hiring and onboarding | Week-one material that names the tools | New hires learning from whoever sits nearby |
| A measurement that can fail | A scored task repeated after a delay | A usage dashboard and a green trend line |
Score yourself on all six and the result is usually lopsided: two solid, one half-built, three absent. That shape matters more than any maturity label, because the absent parts are the work.
Part one: an owner with a name and calendar time
Fluency has an owner in every company where it holds. Not a sponsor, not a steering group. One person with reserved hours, a budget line, and the standing to tell a VP that a workflow is being changed.
The Financial Times hired an AI Fluency Lead from outside and stood up a cross-department group spanning data, technology, learning and development, delivery, and communications, as its published case study describes. A 200-person company cannot staff that and should not try. The small version is narrow: a named person, four hours a week, two quarters, with authority over one list of workflows. Section found tested proficiency averaging 42 out of 100 where a full-time head of AI existed against 33 where one did not, the weakest of its four comparisons, which fits what we see. The title matters less than whether the hours exist. Ownership assigned without time is the most common failure here, and the program guide covers who should hold it when no job title says so.
Part two: workflows that were rebuilt, not just assisted
The sharpest line between a fluent company and a familiar one runs through the workflow. In a familiar company every process still has the shape it had in 2023 and people paste pieces of it into a chat window. In a fluent one, a few processes were rebuilt with the model inside them and the old steps are gone.
McKinsey’s 2025 State of AI survey reported that only 21 percent of generative-AI adopters had fundamentally redesigned any workflow, and that redesign had the strongest link to earnings impact of the attributes it tested. BCG’s 2026 numbers show movement: 42 percent of firms now use AI to reshape workflows end-to-end or build new models on it, up from 22 percent a year earlier.
The inspection is simple. Name three recurring deliverables and ask whether the written steps changed. If the steps are the same and the model sits beside them, you have assistance. Assistance is worth having and it is not fluency, for the reasons set out in the difference between buying tools and building habits.
Part three: the written layer
Capability that lives in four people’s heads is a staffing risk wearing a success story. Fluent companies write things down, and the artifacts are dull: a context document per team describing the work and what good output looks like, a small prompt set someone maintains, a review rule naming who checks which outputs against what, and a short list of what must never go into a model.
This is the shift from using a tool well to producing something reusable, and the evidence that it is rare is blunt. In Section’s hands-on testing, 21 percent of workers could write an effective prompt and 3.8 percent produced workable instructions for an AI assistant. That second number predicts whether anything you build survives a staff change.
Keep it short enough to stay true: one page per team, owned by a person, revised when a tool or a process changes. A thirty-page policy nobody opens is not a written layer.
Part four: managers who ask for the method
This is the part buyers under-fund and the data treats as decisive. Section found that people whose managers both expect and require AI use scored 1.5 times higher on tested proficiency. It also found 65 percent of managers either setting no expectation at all or encouraging use without accountability, and 37.5 percent having done none of the three available things in the previous 30 days: show their own use, ask a team member to use it, or discuss goals for it.
BCG agrees from the employee side. Only 28 percent of frontline staff saw a strong link between what leaders said about AI and what their company did, and 66 percent got limited or no guidance on what to do with the time the tools saved.
What to inspect: a standing ritual where someone shows how they produced something. Fifteen minutes inside an existing meeting is enough. The question that makes a manager useful is not “is anyone using it” but “show me how you did that.”
Part five: hiring and onboarding that assume the tools exist
Fluency that is not built into intake decays by attrition. Every new hire resets a team’s average unless week one names the sanctioned tools, the review rule, and where the context document lives.
Some companies set a bar at the door. Zapier publishes a fluency rubric with four levels: Unacceptable, where someone avoids the tools; Capable, where they use AI on purpose and can explain how it helped; Adoptive, where it is a dependable, repeatable part of their work; and Transformative, where it has changed how they approach work and they pull others along. The post frames it as a self-assessment ladder rather than a pass-fail test, and says every teammate is expected to meet a baseline.
Borrow the ladder, with one caution. A rubric used in hiring selects for people who learned somewhere else and does nothing for the 140 people you have. Set the expectation with it and fund the training separately.
Part six: a measurement that can fail
The last part makes the other five honest. A measurement that can only go up is not a measurement. Usage dashboards count activity, not capability, and will report a healthy number for a team that cannot produce a sendable draft, as we set out in why usage dashboards overstate real capability.
A measurement that can fail has three features. It uses a real recurring deliverable from your own files. It is scored on whether a manager would send the output without a rewrite. And it is repeated weeks later with no warning, because capability that fades when the facilitator leaves was never capability.
The FT is instructive here too. Its internal fluency survey moved from 88 percent at AI-literate or above to 98 percent in six months, and the same page states plainly that the business impact has not been quantified and the evidence so far is anecdotal. A company willing to say that in public is further along than one quoting a number it cannot source.
What an AI-fluent company does not have
The absences are as diagnostic as the parts, and faster to check.
No center of excellence without a delivery mandate. A group that writes standards and ships nothing produces documents describing a company nobody works in.
No usage target tied to pay or standing. The moment a number affects someone’s review it stops measuring anything. You will hit the target and still not know what you bought.
No tool list longer than the workflow list. Four assistants and no rebuilt process is a procurement hobby.
No training that was delivered once. McKinsey’s 2026 survey found about 6 percent of firms attributing at least 5 percent of earnings to AI, barely moved while adoption climbed. Events do not move that line. The staged alternative is in the five stages from pilot to default behavior.
A thirty-minute diagnosis from documents you already own
No vendor, no survey. Open five files.
- The calendar. Open the calendar of whoever owns this. Are there recurring hours on it for the work? If not, the owner is a label.
- The process doc. Pull the written steps for one recurring deliverable. Were they edited in the last six months, and does the edit involve a model?
- The shared drive. Search for a context document, a saved prompt set, or a review rule. If the only hits are a policy PDF and a deck from a vendor session, the written layer does not exist.
- Last quarter’s onboarding pack. Does week one name the sanctioned tools and the review rule? If new hires learn this from whoever sits nearby, you lose ground on every hire.
- The last report that went up. An activity count is a license report. A scored task repeated after a delay is a capability report.
Five yeses is rare. Two or three is normal at a company of 50 to 500 people that bought licenses eighteen months ago, and it tells you where the money goes next. The fuller version is in what two years of engagements taught us about readiness.
What an AI enablement program builds first
Sequence beats breadth. Baseline one scored task per team so you know the starting point. Train against the gap that baseline exposes, on the team’s own documents and reporting cycle, because transfer only happens in the context the work lives in. Leave the session with the written layer drafted by the people who will use it, give managers the fifteen-minute ritual and the one question, and retest the same task weeks later.
Facilitated sessions run roughly $2,000 to $15,000 each at market rates depending on format and cohort size, which puts a multi-team rollout in the low hundreds of dollars per trained person. Set that against ATD’s 2026 State of the Industry figures: 16.7 formal learning hours per employee a year, against direct spend of $846 per employee, down from $1,254. The binding constraint is rarely budget. It is that nobody scheduled the morning or decided who owns the result.
If you ran the five-file diagnosis and did not like the answers, that is the conversation we have on a free AI-readiness call. Book one here and bring the five files.
Frequently asked questions
What does it mean for a company to be AI-fluent?
The capability is structural rather than personal: a named owner with reserved time, a few rebuilt workflows, a written layer of context and review rules, managers who ask to see the method, intake that names the tools, and a measurement that can come back negative. A company where four enthusiasts do impressive things is not fluent, it is lucky.
What is the difference between AI literacy and AI fluency?
Literacy is knowing what these systems are and roughly what they do. Fluency is reliable judgment in live work: deciding what to delegate, describing it well, assessing what comes back, and owning the result. Anthropic’s framework names those four Delegation, Description, Discernment, and Diligence.
How do I tell whether my company is AI-fluent without running a survey?
Inspect artifacts instead of asking people: the owner’s calendar, the written steps of one recurring deliverable, the shared drive, last quarter’s onboarding pack, and the last AI report that reached leadership. Self-report is the least reliable channel here. Section measured 57 percent of C-suite respondents saying AI was widely integrated against 18 percent of individual contributors inside the same companies.
Does everyone in the company need to be AI-fluent?
No, and aiming for it wastes money. Set a baseline everyone meets: the review rule, the sanctioned tools, and one task they can do unaided. Then go deep in the roles where the work is document-heavy and repetitive.
Should AI fluency be part of hiring?
It is reasonable as a screen, and it solves a different problem than training does. A hiring bar improves the people arriving and does nothing for the people already there. Zapier’s four-level rubric is a usable starting point, written as a self-assessment ladder rather than a gate.
Who should own this when there is no Chief AI Officer?
Whoever controls both the license renewal and the training budget, usually a COO, CFO, or VP of Operations. Four reserved hours a week for two quarters, with authority over one list of workflows, beats a committee of six senior people with no time between them.
How long does it take, and does fluency fade?
Expect two to three quarters before a method is the default on a few workflows, and assume decay without reinforcement. An unannounced retest weeks after training catches the fade while it is still cheap to fix.
What does it cost to build the parts we are missing?
Facilitated sessions run roughly $2,000 to $15,000 each at market rates depending on format and cohort size. The written layer and the manager ritual cost time rather than money. Custom automation, once fluency is real and the workflows are known, runs in the $25,000 to $150,000 range.
Key takeaways
- Fluency is a property of a company, not a feeling. Section’s scored testing of 5,026 US knowledge workers found 5.5 percent proficient against 73.5 percent experimenting, so self-report and tested ability measure different things.
- Six parts carry it: a named owner with reserved hours, rebuilt workflows, a written layer, managers who ask to see the method, intake that names the tools, and a measurement that can fail.
- The features Section found alongside higher tested scores were specific training, clear sanctioned access, a formal strategy, and a named owner. Associations, not proof, and still the best parts list available.
- Manager behavior is the cheapest lever and the most neglected: 65 percent set no expectation or no accountability.
- Diagnose from five documents you already own before buying anything. Two or three yeses is normal, and it names the part to fund first.
Arthur Wandzel
