Methodology

The census asks people who work in international education: education agencies and counsellors, universities and colleges, language schools and pathway providers, associations, and the service companies around them. One person, answering for themselves and for the organisation they work in — the instrument separates those two throughout, because “I use AI daily” and “my organisation has deployed AI” are different claims.

This is a self-selected sample, not a probability sample. There is no sampling frame for this industry — no register of everyone who works in it — so no margin of error can honestly be quoted, and none is. People who agree to spend eleven minutes on a survey about AI are, on average, more interested in AI than people who do not. Read every figure as a description of the people who answered, cut by the kind of organisation they work in, and not as a projection onto the industry as a whole.

Four sections, around forty questions, roughly eleven minutes: AI in your work; connected and agentic AI; business conditions; policy and outlook. Sections are saved as they are completed, so a response that stops after section two still counts for section two.

Every question carries a stable identifier (s1_ai_frequency, and so on). Wording may be improved between editions; the identifiers and the stored answer values are frozen at the pilot and never renamed, because year-on-year comparability is the point of running a census rather than a poll. Where a question changes enough to break comparability, it gets a new identifier and the old series ends.

The questionnaire is published in English, Portuguese (Brazil) and Spanish. English is authoritative; the translations are reviewed by native speakers working in the industry, and answer values are language-independent, so a response is stored identically whichever language it was given in.

People reach the census four ways, and each is recorded:

• Partner invitations — an association, network or event invites its members, each with their own link. Partners see who started and finished; they never see an answer.
• The Brief’s own list — subscribers to the podcast newsletter.
• Referrals — a respondent passing the link to a colleague. Attribution is first-touch and fixed when the response is created.
• The public link — anyone arriving from social, search or a newsletter mention.

Source is stored on every response, which means the published figures can be tested for source skew: if a headline number moves materially depending on how people arrived, that is reported alongside it.

Automated submissions are challenged at the door. Each response is tied to a single token, one submission per email address, and an email address is verified as deliverable before results are sent to it.

Responses are flagged, not silently deleted, when:

• a section is completed faster than a person can read it — under 45 seconds. One fast section is someone with little to say about it; two or more is a pattern, and the response comes out of the published figures;
• answers contradict each other. Never using AI while reporting AI that acts without human approval is not a reading of the questions we can reconcile, and that response is excluded; milder contradictions — daily use with no tool named above monthly — are recorded and kept;
• a matrix filled in with one stroke: all fourteen tools marked as used daily is not a working week, and that response is excluded. Every tool marked “never” is perfectly plausible and is not flagged;
• several responses come from the same organisation. Colleagues are legitimate respondents; what they distort is the count of organisations, so they are flagged — by employer email domain, and a free-mail address never makes two people colleagues — and handled in the analysis rather than blocked;
• referrals are capped before they can be farmed: five credited referrals per respondent, at most two from any one organisation, and nothing is credited from a response the rules above excluded.

Flags are stored with the response and applied at analysis time. The count of excluded responses, and why they were excluded, is published with the results.

These rules were fixed before fielding opened:

• Headline figures are unweighted and always reported with their own base. With no frame to weight to, weighting would trade a visible bias for an invisible one.
• Every figure is reported at its own n, per section and per segment — not at the number who started. Later sections have smaller bases than earlier ones, and the difference is stated rather than smoothed over.
• Segment cuts are by kind of organisation (s1_org_type), the axis the instrument was designed around. Country, size and years in the industry are secondary cuts.
• Duplicate-organisation robustness check — each headline figure is recomputed keeping one response per organisation. Where that moves a figure by more than two points, both numbers are published.
• “Don’t know” is reported, not dropped, and is excluded from ranking questions rather than counted as a low score.

The census sets one scalar: The Brief AI Adoption Index, out of 100, for the industry and for each kind of organisation large enough to publish. It is what makes this a series rather than a report — the same number, computed the same way, next year and at each quarterly pulse in between.

The formula was fixed before fielding opened and is published here, which is the only thing that stops an index from being chosen to suit its first result. It is a weighted average of five answers, each scored from 0 to 1 and averaged across everyone who answered that question:

Five single-choice questions, on purpose: an index the quarterly pulse cannot recompute is an annual statistic with extra steps. “Don’t know” is left out of the average rather than scored as zero — uncertainty is not the absence of AI — and each component reports its own base, so the Index is published with the smallest of them beside it. If any component is under the minimum sample, no Index is published at all.

Changing any weight or scale would start a new series with a new version. It would not revise this one.

Nothing is published while the census is open. Half-collected numbers move, get quoted, and cannot be taken back.

No cut is published below ten responses. That floor applies everywhere the same way — the public report, the benchmark a respondent sees, and the private cut an association gets of its own members. Where a segment is too small, the page says so instead of falling back to a bigger group without telling you.

Respondents get their own results before anyone else. After the report is published, an anonymised dataset is released so the figures can be checked; it is reduced first so that no combination of answers can identify a respondent or a small organisation, and free-text answers are excluded from it.

Consent is asked in separate checkboxes, never bundled: results and your personal benchmark; contact from the sponsor; short quarterly follow-up surveys. Declining any of them does not change the survey.

Answers are stored separately from contact details, aggregates are published only above the minimum sample, and an association that invites its members can see participation but never an answer. The privacy notice sets out what is kept, for how long, and how to have it deleted.

If a published figure turns out to be wrong, it is corrected on the page with a dated note saying what changed. Figures are not quietly edited.

Questions about the method, or a request for a cut that is not published: survey@thebriefinternationaleducation.com.