Half a million AI-graded questions later, schools still await the rulebook

Revision Village’s proprietary AI marking tool has graded almost half a million questions and produced tens of thousands of hours of personalised feedback. That is past novelty. It is production infrastructure, running faster than the schools, tutors and regulators around it can adapt.

The platform, owned by Crimson Education since 2023, is used by about 85% of International Baccalaureate students across more than 700 schools globally, and 93% of IB students in New Zealand, according to the Herald. Chief executive Tim Vaughan knows the playbook. He previously helped grow Education Perfect past one million users before a sale that valued it at up to $455 million.

Vaughan pitches AI as a way to strip out repetitive marking so teachers can focus on students who need help, and argues personalised education at scale “has never actually been possible before”. He also warns that AI must not remove the cognitive friction of learning. “Learning is about kind of wrestling with ideas,” he says. Tauranga educator Barbara Cook of Steam-Ed Tauranga is more sceptical, asking what children lose when software does too much of the thinking.

That pedagogy debate matters. But the sharper story for business is about speed.

The state built its own machine, carefully

The public system has not been idle. Back in 2025, NZQA’s then-chief executive Grant Klinkum announced automated text scoring would do the first sweep of Year 10 writing tests, after a 2024 trial on 36,000 learners found 80% agreement with human markers.

It is now embedded. An August 2026 review by NZCER and NZQA covered 32,406 responses marked this year, finding 99.8% exact-or-adjacent agreement with human scores at rubric level and 84.5% of total scores within one point of human markers. That is production-grade validation.

Yet the rollout remains deliberate. NZQA’s September 2026 release says expansion to end-of-year NCEA assessments is gated on AI matching or beating human quality. Caution is defensible for a national qualification. But it means a private operator is moving at software speed while the public exam body moves standard by standard.

AI is not a cost-cutting story, yet

Here is the detail that should recalibrate boardroom assumptions. NZQA states plainly that “no cost savings are expected” from AI use in the foreseeable future, despite results arriving 3.5 weeks faster. Humans have not vanished either. For the May 2025 round, NZQA’s contracted marking workforce was 20,128 people, more than 36% of the number of students assessed.

The payoff is speed, consistency and capacity, not headcount. Any business modelling AI as a straight labour substitute should take note. The early gains come from doing more, faster, with human checking layered on top.

Cheap compute changed the economics

This is global, not local. In February 2026, UK assessment specialist Daisy Christodoulou reported that nearly half a million pieces of student writing had been assessed with AI models across England, the US, Australia and New Zealand. She credited the “dramatic reduction in cost” of standard models, which lets operators send the same work for multiple AI judgements to catch inconsistency. Accuracy did not suddenly leap. The price of checking collapsed.

Policy flipped inside a year

Against that pace, official guidance has wobbled. Joint Ministry of Education and NZQA guidance once told schools to avoid using AI to mark work. In October 2025 the Ministry reversed course, allowing AI-assisted marking with human oversight. Principals said the change pushed too much responsibility onto individual schools.

The Helen Clark Foundation’s senior researcher Sarah Bell says more than nine in ten school leaders and four in five teachers now use AI at school. Her verdict: “AI uptake is going nuts, governance is going nowhere, and risk is skyrocketing.” She points to Estonia and Singapore as countries that coordinated training and tools rather than reacting case by case.

The centre-right reader need not want a heavy-handed rulebook. But clear, stable guidance is what lets schools buy and use tools with confidence. Ambiguity is its own regulatory cost.

The lesson reaches beyond classrooms

Wherever value rests on slow, scarce human judgment, whether assessment, legal review or audit work, nimble private operators are already building tools that scale it. Revision Village did it in a niche international qualification and reached near-total penetration. Tutors and traditional providers now compete with software that marks overnight.

The next test is NCEA’s internally assessed work, which makes up most credits and where AI risk is highest. Whoever sets workable rules there first will decide whether New Zealand’s edtech exporters lead the transition, or watch it run ahead of everyone.

Sources

• NZ Herald: AI in education: Tauranga experts weigh the benefits and risks (2026-10-24)
• RNZ: Artificial intelligence exam-marking on the way for Year 10 writing tests (2025-04-09)
• No More Marking: Solving marking at scale! (2026-02-28)
• NZ Herald: Ministry revises AI marking rules as teachers warn of confusion (2025-10-16)
• Helen Clark Foundation: AI is already in every classroom. Policy needs to catch up (2026-10-20)