Splitting the HK$500,000 AI Education Grant: Four Commitments, Two Years, and Evidence That Survives

November recess, in a primary school general office. The curriculum coordinator running the grant pushes a sheet of A4 across the desk. "Five hundred thousand. Three subjects, two grades each, six teaching exemplars, three demonstration lessons, three sharing sessions — all finished before 31 August 2028." The P4 mathematics teacher opposite reads it and asks one question: "So which lesson do we start with?"

What this money actually buys is not software; it is the fact that a group of teachers changed how they teach over two school years, and the evidence that they did. Which means the budget and the timeline should both be built backwards: work out what the October 2027 interim report has to show, then decide what to buy in November 2026.

What kind of money this is

The AI Education Grant comes from the Quality Education Fund, launched on 16 December 2025, offering eligible public-sector and DSS primary and secondary schools up to HK$500,000, once, to be spent by 31 August 2028.

"Once" is worth reading out twice in a meeting. It has two consequences. Anything with an annual price tag needs an answer to the question of who pays in year three. And the money is best spent on things that stay in the school after it is gone: a built-out set of materials and knowledge bases, a group of teachers who are fluent, a classroom routine that works. Licence fees leave nothing behind; those three do.

The four commitments

The proposal promises four outcomes and the reports account for them. They look like four numbers. They are four different kinds of work.

Three subjects, two grades

At least three subjects, at least two grades in each. One panel cannot deliver that, so the first move is naming names: which three panels, which grades, who is responsible. Do this at proposal stage and be specific — P4 and P5 Mathematics, P5 and P6 General Studies, S1 and S2 English.

There is a practical rule for choosing panels: find the teacher who is already marking compositions or generating practice questions with AI on her own initiative, and take her panel. She does not need convincing, only support. If two of your three panels start that way, year one will not stall.

Six teaching exemplars

At least six, with at least one for every subject-and-grade combination. What separates a useful exemplar from a decorative one is whether it records the part that failed. "Students were highly engaged" helps nobody. "The first prompt was too long and students read only the first two lines; the second version was three sentences and worked" is something a colleague can copy.

Give every exemplar the same four parts: lesson context (grade, subject, class size, duration), the exact prompt the teacher used, two or three pieces of student work including one that went badly, and a short reflection of two or three hundred words. A fixed format saves several evenings when the annexes are assembled two years later.

Demonstration lessons and sharing sessions

At least three of each. These are the commitments most likely to be crammed into the final term, and the ones crammed in are the ones nobody remembers. Attach them to things that already exist in the calendar: twenty minutes after the monthly staff meeting for a sharing session, the panel's co-planning slot for a demonstration lesson. Do not invent a new meeting.

The evidence needed is small: date, subject and grade, the list of colleagues who observed, one observation record, and the discussion points from the ten minutes afterwards. Write those five things the same day, not in reporting season.

Phasing the two years

Year one

September to October 2026: settle the platform and services, submit the proposal. November to December 2026: create accounts, import curriculum and materials, run the first round of teacher training. Keep that round narrow — ninety minutes, one tool, everyone leaves holding something built from their own teaching material. That beats a full day of overview.

January to June 2027 is the half-year that matters. All three panels get into classrooms, the first three exemplars are written, one or two demonstration lessons run. When the interim report is due in October 2027, you want real classroom data in hand rather than a plan.

Year two

2027-28 does two jobs: finish the remaining exemplars, demonstration lessons and sharing sessions, and spread what worked from three panels to more. The October 2028 final report has to answer the sustainability question, and the best answer is "five panels are using it and renewal is funded from recurrent expenditure" rather than a promise.

On the day you write the interim report, what you will be grateful for is not the platform you bought but the three lines somebody jotted down after each lesson.

A split in real numbers

Here is one secondary school's allocation of HK$500,000, taking the position that people should get more of it than licences do.

• Platform subscription, two school years: HK$200,000. Confirm uncapped whole-school accounts, all modules included, and no separate charge for AI usage. Per-seat or per-token pricing makes the year-three renewal impossible to forecast, which is exactly what the sustainability section needs.
• Consultant co-planning visits and teacher development: HK$150,000. Several visits a term, sitting with panels to build the exemplars, rather than delivering a talk.
• Relief cover, materials and record-keeping for demonstration lessons and sharing sessions: HK$50,000. This is the line item most often omitted, and omitting it is why nobody has time and the commitments slide into the last year.
• Evaluation: HK$50,000. Pre- and post-tests, student survey, teacher survey, data preparation.
• Hardware and peripherals: HK$50,000. Headsets, microphones, a few spare tablets — bought against an actual gap, not to use up a budget line.

The stance is deliberate: forty per cent on the platform, forty per cent on people. When a split puts seventy per cent into licence fees, it usually means nobody has been given time to change how they teach, and the report will be hard to write two years later.

How the evidence accumulates

What the system records for you

If every AI interaction already carries subject, grade, module and usage, then "three subjects, two grades" does not depend on anyone's memory — a usage export states it. Mastery of learning objectives, how often students interacted with a tutor, before-and-after results tracking: all of it accrues from ordinary daily use. One question worth asking during procurement is whether an administrator can export those records to CSV without help.

What somebody has to do by hand

Dates, audiences, headcounts, hours and feedback for training and sharing sessions need a human to log them. Observation records need writing. The reflection in an exemplar can only be written by the teacher who taught that lesson. No system substitutes for this part, but it can be compressed to almost nothing: ten minutes, same day.

A practical arrangement — one folder in the panel's shared drive, one subfolder per exemplar, four files in each, filed by whoever finishes. Two years later there is nothing to hunt for.

Common misreadings

• The claim: Buy the platform first; the exemplars and demonstration lessons will follow naturally. What actually happens: It runs the other way. Pick three panels and six specific topics, then ask which platform can do those six things. Schools that reverse the order tend to discover in year two that the features they bought do not line up with the exemplars they promised.

• The claim: Getting value from HK$500,000 means buying the fullest feature set. What actually happens: No report asks how many modules you licensed. It asks whether three subjects and two grades happened, where the exemplars are, and whether teachers changed. Converting thirty per cent of the budget into consultant visits and relief cover usually produces firmer evidence than extra modules.

• The claim: The report is an October 2027 problem; teach first, write later. What actually happens: The report wants records that accumulate lesson by lesson, and those cannot be reconstructed. Starting after the first lesson costs ten minutes. Starting in reporting season costs several weekends and produces something vague.

Three things to try on Monday

1. Write down three panels and six specific topics on paper — "P4 Mathematics × generating fraction word problems", "S2 English × writing feedback". If you cannot reach six, you have learned who to talk to next, which is more useful than another platform demo.
2. Create the shared folder and drop in the four-part exemplar template: lesson context, exact prompt, student work, reflection. One template file, ten minutes.
3. Put October 2027 and October 2028 into the school's administrative calendar along with what the interim report must contain, and set a checkpoint in June 2027. Dates in the calendar are the reason someone starts a year early.