AI Saves Teachers Time. The World Still Needs 50 Million More

A worksheet is rarely the hardest part of a teacher's day. The hard part is noticing who has stopped following, choosing another explanation, and deciding when the plan itself should change. Artificial intelligence can help draft the worksheet. Whether that time goes back into the work that needs a person is a decision the organisation still has to make.

Two October releases make that decision harder to dodge, and they do not describe the same problem. On 6 October 2026, a school-randomised trial in England found that an AI lesson-planning assistant cut reported preparation time by about 49 minutes a week, without making the assessed resources worse. The day before, UNESCO said the world needs 50 million more teachers by 2030 for universal pre-primary, primary and secondary education. On 14 October, the OECD reported persistent shortages of teaching professionals across many of its countries. None of these figures is a licence to replace a teacher with software.

A measured saving, on one part of the job

Oak National Academy's Aila is an AI lesson assistant. A teacher chooses a year group, subject and topic. Aila suggests a learning objective aligned with the national curriculum for England, and activities to go with it. The teacher then refines the plan, with the tool or away from it.

The Education Endowment Foundation (EEF) commissioned the National Foundation for Educational Research (NFER) to test the effect on workload. The evaluation, published on 6 October 2026, was a school-randomised trial over ten weeks in the autumn term of 2025. It involved Key Stage 2 teachers, ages 7 to 11. NFER and the EEF news release describe 464 teachers in 108 schools. Schools were allocated to a group encouraged to use Aila, or a group asked not to use it. Teachers in both groups could still use other generative AI tools, such as ChatGPT or Microsoft Copilot. Both groups were asked not to pull new resources from the Oak National Academy website, so the comparison would isolate Aila rather than Oak's existing library.¹²³

The primary measure was time, recorded in a weekly teacher diary. NFER reports that the Aila group spent about 2.5 hours a week on lesson and resource preparation, against 3.3 hours in the comparison group. That is 76% of the comparison group's time: a difference of 0.8 hours, or 48.6 minutes a week. The EEF news release describes the same contrast as two and a half hours against 3 hours and 19 minutes, about 49 minutes, and calls it a quarter (24%). An independent panel of six teacher-reviewers assessed resource quality. NFER found no evidence that the Aila group's resources were worse.¹²

Those headline hours come from the published summaries. The evaluation report is more precise about who is in the average. Of the 464 teachers randomised, 416 were in the primary analysis: 214 in the Aila group and 202 in the comparison group. 48 teachers, 10.3%, completed no diary entry and were left out. The EEF project page states the same 416 when it describes the data included. Attrition was slightly higher in the comparison group, and the report says that difference was not statistically significant. The reasons given were capacity and teachers leaving the school.³

This is a controlled result, not a vendor's estimate of hours saved. It is also narrower than a claim that teachers became 24% more productive at the whole job. The measured saving covers preparation, during ten weeks. The published findings do not show that pupils learned more, that a school could run with fewer teachers, or that the saving would last for years.

Use was modest. NFER reports that Aila was used for an average of only 7% of lessons, and that both the frequency and the amount of use declined during the trial. The EEF adds that 81% of teachers in the Aila group used the tool at least once, and that fewer than one in five felt it had meaningfully changed their teaching or planning. The EEF also reports that teachers who used Aila more often saved more time, including when the first week and the half-term break were set aside. That is an association inside the trial, not proof that requiring more use would produce the same saving.¹³

Views were mixed. Teachers found Aila accessible and aligned with the curriculum, including when a topic was unfamiliar. Many still preferred other tools, or plans they already had. Aila is built to help someone plan a lesson from scratch, which asks for more involvement than adapting a plan they trust. Outputs often needed adaptation before they were ready for the room. The EEF says almost all teachers in the trial reported lesson planning as a source of stress, and felt they spend too much time on it. A shorter slot can matter even when the tool does not become a habit.³

The time saving was not evenly spread. Teachers in the Aila group who had lower confidence in subject knowledge, lower pedagogical knowledge, who were newer to the profession, or who had higher confidence with technology, were more likely to report a lower workload than similar teachers in the comparison group.¹³

The EEF news release also describes an earlier trial, published in December 2024, in which teachers using ChatGPT spent 69% of the time spent by a comparison group asked to plan without any AI tools. The EEF project page describes that earlier result as a 31% saving for Key Stage 3 science teachers. Use declined in that trial too. Both evaluations received a high security rating from the EEF. The comparison is not identical. In the Aila trial, the other group could still use generative AI. The 49 minutes is time saved on top of tools people were already allowed to use, not a comparison with a world that had no AI at all.²³

NFER's Ben Styles said the team wants future research on whether a newly qualified teacher, or someone teaching outside their subject, actually teaches better with a generated plan. He noted that standing in for experience could help, or could risk poorer lessons. That question is open. The October trial did not answer it.²

The shortage is people, and most of the need is replacement

UNESCO's Institute for Statistics, with the International Task Force on Teachers for Education 2030, released a fact sheet for World Teachers' Day. The article is dated 5 October 2026 and was last updated on 13 October. It says the world has 107 million teachers and needs another 50 million by 2030 for universal access to pre-primary, primary and secondary education. That is one additional teacher for every two currently working. The 50 million is a projection of recruitment for an education goal. It is not a count of vacant posts this morning.⁴

58% of the teachers needed by 2030 are replacements for people who leave, not posts created only to expand the workforce. Attrition is 7.2% in pre-primary education, 5.4% in secondary and 4.7% in primary. If people leave because of pay, workload, safety or a lack of support, a faster first draft of a lesson touches one piece of that story.⁴

On pay, UNESCO says pre-primary and primary teachers earn slightly less than other professionals with similar qualifications. In two out of five of the 68 countries analysed, teacher salaries stay below those of similarly qualified professionals across all education levels. A planning assistant does not negotiate a salary, mentor someone through a difficult class, or make an unsafe school safe.⁴

The same release describes the working week: about 19 hours in the classroom, and around 60 hours in total, much of it unpaid preparation, marking and administration. In primary schools, the global pupil-to-trained-teacher ratio is 28 to 1, the highest of any level. A tool that shortens preparation can matter inside that week. It does not change the ratio.⁴

Status is a retention issue, and the useful evidence is narrower than a single global percentage. In TALIS 2024, the OECD reports that teachers who agree the profession is valued in society are more than 10% less likely to want to leave teaching within five years, after accounting for motivation, contract type and satisfaction with employment terms, including salary. The share who feel valued is not one number. It is above 80% in Kazakhstan, Uzbekistan and Viet Nam, and below 10% in the French Community of Belgium, Croatia, France, Hungary, Portugal, the Slovak Republic and Slovenia. Software does not confer that status.⁵

An indicator of imbalance is not a vacancy count

On 14 October 2026 the OECD published Mapping Skill Shortages in OECD Countries, the 2026 edition of the Skills for Jobs database, covering 38 countries. It does not count empty classrooms. It builds a composite indicator from labour-market signals, including employment growth, wage growth, growth in hours worked and growth in online job postings, then ranks each occupation against other occupations in the same country. The latest year is 2024 for most countries. Chile and Norway stop at 2022. Korea, the Slovak Republic, Switzerland and Türkiye stop at 2023. Australia and New Zealand include 2025.⁶⁷

On that measure, shortages are concentrated in high-skilled occupations. Teaching professionals sit with health professionals and information and communications technology professionals among the largest and most persistent. Australia, France, the United Kingdom and the United States show persistent shortages of teaching professionals. More than 30% of teachers are aged over 50, and only around 10% are under 30. Teachers have, on average, more than 16 years of education, which is one reason the OECD says training more teachers is unlikely to close the gap quickly. A smaller school-age population may ease demand for initial schooling in many member countries. Demand for lifelong learning, including upskilling and reskilling, may keep or raise the need for teaching professionals over a longer period.⁶⁷

Do not add UNESCO's 50 million to this indicator. One is a global projection tied to access by 2030. The other is a relative ranking inside a set of countries, built from historical signals. A national average can also hide a severe local gap. The OECD says its results may differ from national skills assessments because the data and the method differ.⁴⁷

What both the UNESCO projection and the OECD ranking undermine is a comfortable assumption: that automating a slice of professional work automatically solves a shortage of qualified people. The same mistake shows up when a task gets easier and a job title is treated as if it had vanished. AI changes tasks inside jobs before it changes the titles.

Decide what the saved minutes are for

Picture a small nonprofit school whose pupils need very different kinds of support. This is a hypothetical example, not a reported case. Teachers use an approved assistant to draft outlines and differentiated exercises. A manager can book the shorter preparation time as a cost saving. The school can also spend it on feedback, conversations with families, peer coaching, or another pass for a child who needs more help.

The second choice is not automatically the right one. Some exhausted teachers would rather finish earlier. That is a real benefit. The point is that someone decides, in the open, whether the gain becomes less overtime, a better service, training time, or a change in staffing. Without that decision, the spare minutes often fill with new administration.

There is a quality trap as well. A tool can produce a plausible plan faster while moving the work from drafting to checking. Teachers still have to judge subject accuracy, whether the material fits the age group, accessibility, safeguarding and fit with the local curriculum. The NFER result is useful because it included a quality check, and that check was of resources, not of what pupils learned. A school using a different system should not assume the same outcome.

For a closer look at the gap between stretching a task and having mastered the work behind it, see why AI can let apprentices go beyond their training without that counting as mastery.

A small test is more useful than a programme

Small schools, charities and training teams do not need a grand AI programme to learn something. They need a short experiment with a baseline. Pick one recurring task, such as preparing lessons or adapting reading materials. Ask a willing group to record time spent and the changes they make to the output. Compare with a similar period, or a similar group, working without the tool. Classes, subjects and people will still differ. Say so.

Measure more than minutes. Are the materials accurate and usable? Do people keep professional control? Are learners who need more support served better, or only faster? Does the tool create a privacy or safeguarding risk? Do not put identifiable information about pupils into a system that has not been approved for that purpose. Include staff who are less confident, rather than treating the keenest users as the whole team.

The useful number for a leader is not the count of prompts. It is whether the task got easier without making the core service worse, and whether people actually gained time or capability. A shorter preparation slot can support retention. It cannot, by itself, prove that people stayed. That would have to be tracked for longer.

Judgment still has to be learned

AI may change the work of education specialists rather than only lighten it. A curriculum lead may spend less time on a first draft and more time testing materials. Learning-support staff may have more ways to adapt content, and more responsibility for judging whether the adaptation helps. A manager may have to set rules for approved tools, privacy, accessibility and evaluation. These are plausible shifts in the tasks inside a job. They are not evidence that new job titles have already appeared at scale.

Support should differ by person. A novice may want a structured starting point and still need a colleague's feedback. An expert may want control of the content and less of a fixed workflow. The risk for early-career teachers is that convenience replaces learning the craft. If they accept a generated sequence without discussing why it works, they gain speed and miss the judgment they will need when a lesson fails. If the draft is material for a mentor to pull apart, it can be a training object. The October trial did not test those longer effects. Both remain questions.

For organisations in the European Union, the legal line is more specific than a general warning about "AI in schools". This article is not legal advice. Annex III of the EU AI Act lists certain education and vocational-training uses as high-risk: systems intended to determine access or admission, to evaluate learning outcomes, to assess the appropriate level of education, or to monitor and detect prohibited behaviour during tests. A tool that drafts a lesson is not, by that description alone, one of those systems. Classification depends on what the technology actually does.⁸

A different rule is about use, not about a future high-risk label. The Commission's AI Act Service Desk says that using an AI system to infer the emotions of a person in the workplace or in an education institution is prohibited, except where the system is intended for medical or safety reasons. The same service desk says prohibited practices are enforceable from 2 August 2026. The rules for Annex III high-risk systems apply from 2 December 2027, following Regulation (EU) 2026/1744, the Digital Omnibus on AI. A camera that scores a teacher's "engagement", or a pupil's mood, is a different question from a worksheet generator. Data-protection duties already apply to information about pupils, whatever that timetable says. For a closer look at human oversight when software influences people at work, see who oversees the algorithmic manager.⁹¹⁰¹¹

Questions leaders actually ask

Does 49 minutes mean a school needs fewer teachers?

No. The trial measured preparation time for Key Stage 2 lessons over ten weeks. It did not test staffing levels or pupil attainment. UNESCO's 50 million is mostly replacement demand. A shorter planning slot does not recruit, train or keep the person who has to be in the room.¹⁴

Can the 24% figure be read as a productivity gain for the whole job?

The EEF describes the cut in preparation time as a quarter. NFER's ratio is 76% of the comparison group's preparation time. That is one task. Teachers in the comparison group could still use other generative AI. Use of Aila averaged 7% of lessons and fell during the trial. Fewer than one in five teachers said the tool had meaningfully changed their teaching or planning.¹²³

Should a small organisation wait for a perfect tool?

No. It should test one approved tool on one task, with a baseline, a quality check and a written decision about where any saved time goes. It should not paste information about pupils into an unapproved system, and it should not treat the keenest user as proof that everyone benefited.

How hilighter can help

hilighter does not plan lessons, mark work or decide how many teachers a school needs. For a training provider, a nonprofit or a department, the practical question is the same one this evidence raises: which tasks got lighter, which skills still have to be demonstrated, and who is accountable for the judgment that remains.

hilighter helps teams map skills by role, compare people with those requirements in Skills & Talent Insights, record proof of capability in Skills Evidence & Recognition and set personalized growth paths. Its AI Job Impact view is indicative. It is not a prediction that a role will disappear. A free plan covers up to three job roles, with unlimited employees.

Sources

1. Roy, P., Poet, H., Staunton, R., Aston, K., Thomas, D. and Styles, B., NFER, Lesson planning using AI lesson assistant, Aila: A Teacher Choices Trial, 6 October 2026.
2. Education Endowment Foundation, AI tool reduced primary teachers' lesson planning time by a quarter, but views on the tool were mixed, 6 October 2026.
3. Education Endowment Foundation, Aila Teacher Choices trial (project page: 416 teachers in the included data, Oak website restriction, 7% of lessons, 81% used at least once, subgroup pattern, and the December 2024 ChatGPT trial described as a 31% saving). The evaluation report (NFER for the EEF, October 2026) is the source for the analysis sample: 416 of 464 randomised teachers, 214 and 202 by group, and 48 with no diary entry (10.3%).
4. UNESCO, New UNESCO data: For every two teachers, the world needs one more, 5 October 2026, last update 13 October 2026. UNESCO Institute for Statistics and the International Task Force on Teachers for Education 2030.
5. OECD, Sustaining the teaching profession: Results from TALIS 2024.
6. OECD, Mapping Skill Shortages in OECD Countries, 14 October 2026, https://doi.org/10.1787/735f9485-en.
7. OECD, Occupational imbalances, skill shortages and policy implications, in the same report (latest year by country; teaching professionals, age and years of education).
8. European Commission, AI Act Service Desk, Annex III, point 3 (education and vocational training).
9. European Commission, AI Act Service Desk, Is emotion recognition prohibited in workplace and education settings?.
10. European Commission, AI Act Service Desk, When does enforcement start? (prohibited practices enforceable from 2 August 2026; Annex III rules from 2 December 2027).
11. Regulation (EU) 2026/1744, Digital Omnibus on AI, amending Regulation (EU) 2024/1689, including the application date for Annex III high-risk systems.

Sources last reviewed: 30 October 2026. The 49 minutes cover lesson and resource preparation in one English trial, not the whole teaching job, and the comparison group could still use other generative AI. UNESCO's 50 million is a projection to 2030, mostly replacement, not a vacancy census. The OECD figure is a relative indicator. TALIS describes intention to leave, not a measured resignation rate. EU dates follow the Commission service desk and EUR-Lex and are not legal advice. The December 2024 ChatGPT trial is described here as the EEF states it on the October 2026 pages; that earlier report was not re-opened for this article.