How the AI Fluency Index would be reported

The first index publishes when the thresholds below are met, and not on a date. There is no quarterly release, no annual report, and no forthcoming edition. A recurring publication commits to producing a number on a day whether or not the data supports one, and the pressure that creates is the reason thresholds get quietly relaxed.

As of today the measured population is well under the individual floor, is drawn from a single-digit number of organizations, and includes zero workforce assessments that are not demonstration data. Every cut on this page would currently be suppressed. That is the honest state and this page will not describe it any other way.

A cut is any slice of the population narrower than the whole — an industry, or any other grouping. Every cut must clear all three of these before it is reported at all.

At least 20 individuals

The same floor the product already uses for the organization figure that goes in front of a board (20). A published index cut is more exposed than that, not less, so it is not set lower. Using one number rather than inventing a fourth also means the product has a single answer to how many people it takes before it will say anything.

At least 5 distinct organizations

The rule a headcount cannot give you. 20 people from one company is that company’s figure with an industry label on it. It is identifiable to anyone who knows they took part, and it is not an industry measurement in the first place — it is one employer’s culture.

No single organization above 50% of the cut

Without it, 20 people across 5 organizations can still be mostly one company, and both rules above are satisfied while neither does its job.

The unit being counted here is the organization, not the person, so the floor moves onto organizations: at least 10 organizations, and an organization is only counted once it has at least 5 respondents of its own.

The second half is what stops the figure being noise, and it is the same privacy floor every group in the product uses: below 5 respondents a group’s results are not shown, and a one-person organization would flip a share expressed over ten by ten points on its own.

A suppressed cut appears in the response. It is reported as suppressed, with the rule that suppressed it and the count that fell short. It is not omitted. A missing row reads as “there is no such industry”, which is a different claim from “too few to report” and one nobody made.

Suppression is complementary. Publishing a total and every cut but one leaves the suppressed cut recoverable by subtraction, which defeats the entire mechanism. When one cut is suppressed, the next smallest is suppressed with it, or the residual is not published.

Any figure must state which instrument produced it, and a request that does not name one is refused rather than answered across several.

This is not bookkeeping. 26 items were withdrawn from the bank and replaced, on 19 August 2026 and 25 August 2026 and 22 September 2026 and 26 September 2026, so Awareness, Instruction, Verification, Governance and Integration means from before those dates average different items from means after. A date range spanning either day pools two instruments into one number and reports it as though it came from one. Requiring the version makes that impossible to do by accident.

Every response carries the number of people behind it, the instrument version, the form, and the date range, in the response itself. Not in a footnote, not on this page, not in a covering note that can be separated from the figure it qualifies. A number that arrives without them is not one this product produced.

A dimension score is the mean of three answers, so it lands on thirds and almost never on a whole number. Two different rules turn that into a level, they disagree by construction, and both are correct for what they do — so any figure has to say which one it used.

• The distribution across the four levels rounds. A dimension at 2.7 counts in Applied. This matches the badge shown on every result and dashboard, so a person reading their own result and a reader of the index are counting the same way.
• The share under level 3 uses a hard cut at 3.0 and does not round. A dimension at 2.7 is under level 3. Rounding this instead would move everyone between 2.5 and 2.99 out of the count and silently change what the product has always meant by it.

So the same 2.7 is counted in Applied in one figure and under level 3 in the other. Badges round to the nearest level, so someone at 2.7 shows as Applied and is still under level 3. Both statements are true at once, and the phrase “under level 3” is used wherever that count appears, in those words, rather than any wording that implies the badge disagrees with itself.

These are the obvious ones to ask for. None can be produced honestly, and producing them would mean inventing data rather than reporting it.

• By organization size — no such field exists. Nothing in this product asks an organization how many people it employs. The two candidates are both unusable: one counts registered accounts, so a 900-person firm with 40 sign-ups reads as 40, and the other is a free-text display label. Deriving a size band from either would put a number in front of a reader that no organization ever stated. This cut needs a real field and somebody to answer it first.
• By job function — department names are not a taxonomy. What is recorded is the name of the team someone belonged to when they sat it, as their own employer writes it. Those labels are local to one organization and are not drawn from any shared list, so pooling them across employers pools things that are not the same thing. Cutting by them within a single organization is sound and is what the dashboard already does; cutting by them across organizations and calling the result a function is not.
• By anything a person said about their own work. The optional context step captures free-form description in the taker’s own words, consented on the explicit basis that it shapes their own report and goes nowhere else. It contains no industry, size or function field, and inferring one from it would use personal content for a purpose nobody agreed to. It is not a source for any figure on this page and is structurally prevented from becoming one.

By industry is the one cut that could be produced, from the industry recorded at the moment each assessment was submitted. It still needs a controlled list before publication: the field is free text today, so two employers writing “Legal” and “Law” are two industries in the data and one in reality.

Demonstration data and any paid or recruited panel are excluded by construction, on both the public and the workforce paths, and are marked as such on the record itself rather than filtered out by a query somebody has to remember to write. Only a real person who chose to take the assessment counts toward any published figure.

This matters more than it sounds. The population figure is the product for a measurement company, and demonstration data that can silently inflate it is the single easiest way for every number downstream to become false.