What teachers and school leaders actually need from AI

Teachers and school leaders already use AI every day. They use it for the easy work. The work that costs them the most time, and matters most for learners, stays manual.

We followed 33 educators through the first half of a 12-week research fellowship. They brought teaching experience from 13 countries, across primary, secondary and university, public and private, low-income and premium schools. We showed them no product. We asked them to map their real working year, log every AI interaction, then try to solve their hardest problem with the AI tools they already had, and finally narrow and re-test it. We kept every transcript and classified every failure.

Nothing in this brief is a hypothesis we brought in. Each problem was first named by an educator in their own words, then tested against their real AI use, then attempted with real tools, then narrowed and re-tested. Section 6 shows that trail problem by problem.

• The alignment gap is universal. AI use clustered on drafting: lesson plans, worksheets, circulars. Marking, compliance, student tracking and leader oversight got almost no AI.
• The hardest problems are solvable in principle. 28 of 33 frontier problems failed because of missing system design: no knowledge of the school, no record over time, no connection between tools. None failed purely because AI was not capable enough.
• Leaders are blind between observations. 11 of 16 leaders described the same gap: no single, current view of what is being taught, how students are coping, and what teachers do about it.
• Generic AI ignores the school. Tools did not follow the school's textbook, policy, format or local context unless forced, and even then drifted.
• The pattern holds across both groups. We only generalise a finding when the MENA and International cohorts produced it independently. All seven findings meet that bar.

Most education technology is built feature first. A team builds a tool, then looks for problems it can solve. That approach tends to produce more of what free tools already do well: fast drafting.

We inverted the order. Phase 1 (weeks 1 to 6) is problem discovery only, and participants see no product. Phase 2 (weeks 7 to 12) tests AI tools against the problems Phase 1 validated. Problems first, solutions second. The order protects against confirmation bias: participants cannot tell us what we want to hear about features they have never seen.

Two questions drove the work:

1. Which problems cost educators the most, and would they pay to solve them?
2. When AI fails on those problems, why does it fail?

Participants applied to the fellowship and were selected to balance role, school type and education system. Between them they bring teaching experience from 13 countries across South Asia, the Middle East and North Africa (MENA), Africa, Europe and North America. Many have taught in more than one country. We report two groups by where each cohort was run, not by nationality.

Roles and settings

Data completeness

We report gaps openly rather than filling them.

Four assignments, four kinds of evidence

Each assignment produced a different kind of evidence. Together they form a chain: a problem named in week one has to survive three more tests before it counts as a finding.

How we classified failure

A failed AI attempt looks the same on screen whatever the cause. So every attempt was tagged with one or more failure modes:

Guardrails

• No "AI can't" verdict without a re-attempt. Before accepting a capability ceiling, the research team or a stronger prompter tried again. Two verdicts reversed this way (Finding 2).
• Full transcripts, not screenshots. The back-and-forth shows whether someone hit a wall, gave up, or never knew how to ask.
• Authenticity checks. Submissions showing signs of AI-written inflation were down-weighted and verified in live sessions rather than taken at face value.
• Findings kept separate from interpretation. For every participant and assignment, what they reported or did is recorded apart from the research team's reading of it.
• Traceability check. For each frontier problem we checked whether it could be traced back to the participant's own words in Assignment 1 (Section 6.5).
• Log, don't delete. Problems AI could not touch stay on record.
• Privacy by design. Participants anonymised all student and school data. Findings identify participants by role and cohort only.

Before asking about AI, we asked participants to describe their real year and their real week. This is the workflow any AI has to fit into.

5.1 The year: assessment and reporting season is the pressure point

Of the 25 participants whose Assignment 1 named an overwhelming season:

One participant named both. The start of term was more often named the most wasted time: name tags, ID cards, student files, kit sorting, sensor and battery testing, account resets, repetitive planning and reformatting.

5.2 The hours (self-reported)

5.3 What a worst week looks like

Assignment 1 asked for a day-by-day journal of a real worst week. The same pattern repeated: teaching itself was rarely the problem. The week broke under everything stacked around it.

5.4 The one task they would eliminate

Assignment 1 asked each participant for the single task they would remove from their job. Almost none named teaching. They named the work around it.

5.5 A teacher's lesson cycle today

5.6 A leader's oversight cycle today

5.7 The toolkit, and what irritates them about it

Participants named a wide but disconnected stack: general chatbots (ChatGPT, Gemini, Claude, Copilot, Meta AI), education tools (MagicSchool, Brisk, Khanmigo, SchoolAI, Diffit), research tools (NotebookLM, Perplexity), design tools (Canva, Gamma), messaging (WhatsApp), parent apps, and records in Excel, Google Classroom, learning management systems and paper. Much AI use runs on personal accounts, outside any school policy.

Parent messaging "gets cluttered; hard to find a thread from a month ago."Early-years leader (International cohort)

"I always have to inject the real details."STEM program lead, on AI-drafted evaluation reports (International cohort)

Participants also flagged a quality risk on the easy side: students and teachers copying AI output without judging it. One leader called it "AI work slop" and warned AI could do "more harm than good" without judgment. Another teacher described editing AI-assisted reports so they would not be "caught as AI generated."

Every problem in this brief started as an educator's own words about their own work. This section shows how each one was carried through four assignments, and what happened at each step.

6.1 Stage 1, named: what educators said in week one

Before any discussion of AI, participants described their pains inside a map of their real year. These week-one statements are the anchor for everything that follows. A selection, verbatim:

6.2 Stage 2, checked: does AI touch this problem today?

In Assignment 2 participants logged a typical week of real AI use, then rated whether AI touches each of their top problems. The answer was mostly no, and participants explained why in specific terms.

Documented AI failures in real work, with their cost

Assignment 2 also asked for a recent, real AI failure. These are not hypotheticals; each happened during the participant's actual work.

6.3 Stage 3, attempted: the hard problem, across several AI platforms

In Assignment 3 each participant attacked one hard problem across at least three platforms and submitted the full chats. Platforms used across both cohorts included ChatGPT, Claude, Gemini, Copilot, MagicSchool, Khanmigo, SchoolAI, Brisk, Perplexity, NotebookLM and TeachBuddy AI (which some participants already used). Participants chose their own platforms.

6.4 Stage 4, narrowed and re-tested

In the MENA cohort, Assignment 4 asked participants to narrow to the most painful step, attempt it again, package what worked, and re-test it from a clean start. Of the 14 who submitted:

Three Stage 4 results changed what we know:

• A "can't" became a "can." The Grade 8 science teacher split her marking problem: keep the final mark human, use AI for the wrapper (partial-mark breakdown, error analysis, feedback phrasing). It tested clean from a fresh start.
• A ceiling turned out to be missing inputs. The senior school leader supplied the grading conversion and formula. Value-added analysis then worked and reproduced in a new chat.
• The reusable part is the school's context, not the prompt. The Head of Science's fresh test showed the prompt reproduced fine. What made it work was the curriculum sequence for both grades, which must be supplied every time.

In the International cohort, Assignment 4 was a narrowing round. Each participant confirmed or refined their frontier problem; the full attempt follows in a later assignment.

6.5 Traceability: did the frontier problem come from week one?

For each participant we checked whether the problem they ended Phase 1 on can be traced to their own words in Assignment 1.

6.6 Case journeys

Twelve journeys, anonymised, from first mention to final test.

6.7 Every participant, at a glance

One row per participant, anonymised. Frontier problem is where each ended Phase 1.

Each finding is labelled by where it appeared. Both cohorts means the MENA and International cohorts produced it independently. Theme counts are research-team coding; one participant can appear in several themes.

Convergence at a glance

At the end of Phase 1 the MENA cohort's problems were ranked by severity: importance to the educator × willingness to pay × whether a working solution is buildable. Evidence comes from Assignments 3 and 4; origin from Assignments 1 and 2.

The International cohort's top untouched pains

The strongest single use case across the International cohort: report-card comments generated from a full year of student data.

The MENA cohort was run after the International cohort to test whether its findings held in a different context. They did. Every core finding reappeared. The MENA cohort also sharpened four areas that matter especially for schools in the region:

The two problems ranked critical, data reconciliation and the per-student record over time, both sit at the leader level, where school-wide decisions are made.

Willingness to pay (WTP) below is self-reported during Phase 1, before any product was shown. Treat it as directional.

Leaders described how schools buy: head office mandates tools centrally, while section heads adopt informally. One set a clear bar: yes, if setup takes less than an afternoon.

Limitations

• Sample. 33 educators who applied to a fellowship. They are likely more engaged with AI than average.
• Self-report. Hours and willingness to pay are self-reported. Attempts and transcripts are direct evidence; workload figures are not.
• Coding. Failure modes, themes and traceability were coded by the research team, with a re-attempt rule for any "cannot be done" verdict.
• Uneven stages. The MENA cohort's Assignment 4 included a fresh re-test; the International cohort's was a narrowing round.
• Tool choice. Participants chose their own platforms; some already used TeachBuddy AI. Phase 1 did not introduce or evaluate any product.

What AI could not touch

• Legal limits: compiling past exam papers is blocked by copyright, not capability.
• Physical presence: invigilation, playground supervision, lab preparation.
• Human decisions: discipline outcomes, appraisals and hiring stay with people. AI can prepare the evidence around them.
• Infrastructure: in under-resourced schools, AI assumes stable internet, devices and printing that are not there. "No AI tool could fabricate genuine research results," as one lecturer put it about a study disrupted by absent students.

This research was made possible by the Founding Research Fellows below, who gave their time, their honesty and their real working weeks. Names are listed by cohort and are not linked to any finding or quote.