Can AI Lower Tutoring Costs?
High-dose human tutoring works in randomised trials, but it costs thousands of dollars per student a year and is hard to deliver at scale. AI tools are available on demand and look far cheaper to run, though published cost data are scarce. Whether they buy the same learning is a different question, and the trial evidence is early, short and mixed.
On this page
Better: Note and confidence
Low confidenceStructured tutors helped in short trials; unguarded chatbots hurtA Harvard physics trial measured larger gains with a purpose-built AI tutor than with an active-learning class, but over two lessons. At one Turkish high school, students who practised with an unrestricted chatbot scored 17% lower on an unassisted exam once access was removed.Faster: Note and confidence
Low confidenceAvailable on demand; time savings measured in one courseHarvard students using the AI tutor spent a median of 49 minutes against the roughly 60 minutes of class learning time the authors assumed. No school-system trial we found measured time to mastery at scale.Cheaper: Note and confidence
Low confidenceLower running cost likely; cost per unit of learning unmeasuredIn our reading, the one AI running cost we found, a working-paper estimate of about US$20 per tutor a year for a tool that coaches human tutors, is small next to the US$3,500 to $4,300 a year researchers estimated for high-dose human tutoring. We found no peer-reviewed trial comparing the full cost of AI and human tutoring for the same learning gain.Safer: Note and confidence
Governance-gatedOnly with learning guardrails, filtering and evaluationEngland now expects AI education products not to give full answers by default. South Australia's EdChat publishes uptake figures but no learning outcomes.The problem, in numbers
Tutoring is one of the few school interventions with strong randomised evidence behind it, and one of the most expensive to deliver. In Chicago, two large randomised trials of daily, two-on-one maths tutoring for secondary students raised test scores by 0.18 to 0.40 standard deviations.Source 3 The researchers estimated the program cost US$3,500 to $4,300 per student a year, even with paraprofessional tutors used to hold costs down.Source 3
Australia has tried tutoring at system scale. In its 2020-21 Budget, NSW committed $337 million to small-group tuition to help students catch up after learning from home during COVID-19.Source 2 In 2022, schools spent $250 million delivering it to 138,268 students, about $1,800 per student tutored on our calculation.Source 1 The state's evaluators compared participants with matched non-participants at a random sample of 282 schools and found the same average growth in both groups, so they could not attribute any academic gain to the program.Source 1 The auditor had already found that the costing assumed tutors working 35 hours a week for 52 weeks, and support was cut from 40 weeks to 20 per student.Source 2
In our reading, the cost case for AI starts here: human tutoring is expensive, it is hard to staff at the dose the trials used, and spending alone does not guarantee results.
What improves, and why
In our reading, an AI tutor costs little per extra session once it is built, and it is available at any hour. That changes who can get help. The UK government says its planned AI tutoring tools are aimed at disadvantaged pupils who cannot afford private tutors, up to 450,000 a year on free school meals in years 9 to 11, and that they will complement face-to-face teaching rather than replace it.Source 8
A second model keeps the human tutor and uses AI to coach them. In a US trial with 783 tutors, an AI tool that suggested teaching strategies cost about US$20 per tutor a year in computing charges on the authors' estimate, excluding salaries.Source 6 That figure comes from a working paper that has not been peer reviewed, so we cite it only for its cost accounting, not its effect on learning. In our reading, it shows how small the AI share of the bill can be when people still do the tutoring.
A cheap tutor is only cheaper if it buys the same learning. Most of the trials so far are too short to tell.
What the evidence actually says
Structured AI tutors can work, in narrow settings. In a randomised crossover trial in an introductory physics course at Harvard, 194 students each had one lesson with a purpose-built AI tutor and one in an active-learning class. Median post-test scores were 4.5 with the AI tutor against 3.5 in class, from a baseline of 2.75, and the AI group spent a median of 49 minutes against about 60 minutes of class learning time, as the authors assumed for a 75-minute class.Source 4 The authors stress the limits: one course, two lessons, introductory material, expert-written prompts and several months of development. They advise against AI entirely supplanting in-class teaching, and the paper does not discuss cost.Source 4
Unguarded chatbots can harm learning. In a field experiment with nearly a thousand students at one Turkish high school, access to GPT-4 during practice lifted practice grades by 48% with a standard chat interface and 127% with a version built to tutor.Source 5 When access was removed, students who had used the standard interface scored 17% lower on an unassisted exam than students who never had it. The tutoring safeguards largely mitigated that harm.Source 5 In our reading, this is the central caution for anyone pricing AI tutoring: a tool that does the work for students can look effective while it is switched on.
System deployments are measuring use, not learning. South Australia has been rolling out its EdChat assistant to public secondary students in phases since term 4, 2025, after a trial involving more than 10,000 students.Source 7 The department reports that 41% of students used it at least once in its first full year and that 94% of student interactions related to curriculum subjects, but it has published no learning outcomes or costs.Source 7 A randomised World Bank evaluation of a six-week AI tutoring program in Nigeria is still a working paper, so we do not treat its results as proof.Source 6
We found no peer-reviewed study that compared the full cost of AI and human tutoring for the same learning gain. The independent evidence on cost-effectiveness is thin, and our confidence is graded down accordingly.
Where it falls short
Short trials, long claims. The strongest AI result covers two lessons in one university course.Source 4 The human tutoring evidence covers full school years.Source 3
Running cost is not total cost. The Harvard tutor took months of expert development, and the US$20 coaching figure excludes salaries.Source 4,Source 6
Uptake is uneven. In our reading, an on-demand tool helps only students who use it: in South Australia, 41% of students had tried EdChat at least once in its first full year.Source 7
Answers are not learning. Without guardrails, practice gains reversed into lower grades once access ended.Source 5
Where governance decides the outcome
Whether AI tutoring is safe, and whether it is cheap, both depend on how it is designed and checked. England's product safety standards, updated in January 2026, expect education AI to avoid giving final answers or full solutions by default, to move from hints to detail, to track when learners offload thinking, and to avoid sycophancy or designs that prolong use. They also require filtering and activity logging, and expect no commercial use of personal data without a lawful basis.Source 8 South Australia's EdChat keeps data in Australia and uses a content filter that can notify the principal.Source 7 In our reading, neither framework yet requires a provider to show learning gains before schools pay for scale. For monitoring and consent questions, see our review of student safety online.
How to think about adoption
Treat AI tutoring as a trial, not a saving. Price the whole program, including setup, prompt design, teacher time and support, not only the licence or computing charge. Choose tools that withhold answers by default and test students without the tool, because the Turkish trial shows practice scores can mislead.Source 5 Compare against a matched group and measure learning, not log-ins; NSW spent heavily on human tutoring and still could not show a gain.Source 1 Use AI first where human tutors are scarce, or to support human tutors rather than replace them, and keep high-dose human tutoring for students who need the most help, where the evidence is strongest.Source 3
BFCS.ai does not fabricate figures. The AI learning effects here are measured in randomised trials, but those trials are short and confined to one course or one country; the human tutoring cost is a researchers' estimate, the AI running cost is a working-paper estimate, and the NSW per-student figure is our calculation. We found no peer-reviewed study that compared the full cost of AI and human tutoring for the same learning gain, and we say so where it matters.
Sources
- 1
NSW Centre for Education Statistics and Evaluation (CESE), "COVID Intensive Learning Support Program: Phase 3 Evaluation Report," published 22 December 2023. education.nsw.gov.au
Report: education.nsw.gov.au
Back to textWhat we took from it
In 2022 a total of $279 million was distributed to 2,186 government schools and program funds delivered small-group tuition (groups of 2 to 5) to 138,268 students reported as participating; schools spent $250 million implementing the program in 2022. Participants at a representative random sample of 282 schools, contacted to correct their data, were matched to similar non-participants: literacy participants grew "on average by the same amount" in Check-in reading scores, with no statistically significant differences at any year level; on average growth was the same for numeracy, with Years 5 and 6 showing slightly less growth (effect sizes of minus 0.08 and minus 0.11); participation "had no measurable effect" on absences in Term 4 2022; no cost-benefit analysis was completed. The per-student figure (about A$1,800) is our calculation: $250 million divided by 138,268 students is about $1,808. Page:
- 2
Audit Office of New South Wales, "COVID Intensive Learning Support Program," performance audit, Report 358, 15 December 2021. $337 million program announced in the 2020-21 NSW Budget ($289 million government schools, $31 million non-government schools, $17 million administration); government schools funded at $983 per targeted student (reduced from $1,109); the original costing assumed a full-time teacher would tutor 35 hours a week for 52 weeks and an average group size of five, and support was cut from 40 weeks to 20 weeks per student; the evaluation had not established methods to reliably assess whether the program closed the equity gap. audit.nsw.gov.au
Back to text - 3
Guryan, J., Ludwig, J., Bhatt, M. P., Cook, P. J., Davis, J. M. V., Dodge, K., Farkas, G., Fryer, R. G., Mayer, S., Pollack, H., Steinberg, L. & Stoddard, G., "Not Too Late: Improving Academic Outcomes among Adolescents," American Economic Review 113(3): aeaweb.org
Cost and design from the University of Chicago Harris School, "As Students Return to School, New Research Finds High-Dosage Tutoring Can Reverse Learning Loss," news release from the University of Chicago Education Lab, 15 March 2021 (released two years before the journal version): harris.uchicago.edu
Back to textWhat we took from it
738 to 765, 2023 (peer-reviewed; abstract read, full text paywalled). Two large randomised trials of high-dosage maths tutoring for secondary students in Chicago using paraprofessional tutors; tutoring raised maths test scores by 0.18 to 0.40 standard deviations.
two-on-one maths tutoring in daily 45 to 50 minute sessions in 9th and 10th grade; cost of $3,500 to $4,300 per participant per year (US dollars, researchers' estimate).
- 4
Kestin, G., Miller, K., Klales, A., Milbourne, T. & Ponti, G., "AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting," Scientific Reports 15, article 17458, 3 June 2025 (peer-reviewed). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Randomised crossover trial in an introductory physics course at Harvard: 233 enrolled, 194 analysed; each student had one lesson with a purpose-built AI tutor at home and one in-class active-learning lesson (the paper names GPT-4 as an example of the kind of model needed; its methods do not name the model used). Median post-test 4.5 for the AI condition against 3.5 in class (baseline 2.75); effect size 0.63 by linear regression, 0.73 to 1.3 standard deviations by quantile regression. Median AI time on task 49 minutes against about 60 minutes of learning in a 75-minute class. The authors note a single course, introductory material, expert-written prompts and several months of platform development, and advise against AI entirely supplanting in-class instruction; the article does not discuss cost.
- 5
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R., "Generative AI without guardrails can harm learning: Evidence from high school mathematics," Proceedings of the National Academy of Sciences 122(26), e2422633122, 2025 (peer-reviewed; open-access full text read on PubMed Central, PMC12232635). doi.org
Full text: pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Field experiment with nearly a thousand maths students in grades 9 to 11 at a single high school in Turkey, over four 90-minute sessions in the 2023 to 2024 school year; GPT-4 access during practice improved grades by 48% (GPT Base, a standard chat interface) and 127% (GPT Tutor, with learning safeguards); when access was removed, GPT Base students had a 17% reduction in grades compared with students who never had access (measured on a closed-book, unassisted exam: 0.054 on a 0 to 1 scale, about 17% of the control average), while the safeguards in GPT Tutor largely mitigated the negative effects.
- 6
Working papers, not peer reviewed, cited only for their design and cost accounting. povertyactionlab.org
De Simone, M. E., Tiberti, F. H., Barron Rodriguez, M. R., Manolio, F. A., Mosuro, W. & Dikoru, E. J., "From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria," World Bank Policy Research Working Paper 11125, 2025: ideas.repec.org
Back to textWhat we took from it
J-PAL evaluation summary of Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S. & Demszky, D., "Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise," working paper, November 2025: randomised trial of an AI tool that suggests strategies to human tutors, 783 full-time tutors (388 with the tool) and 1,013 students in grades 3 to 6 in a large Southern US school district, March to May 2024; total API cost for 388 tutors over two months was US$1,419.66, an estimated US$20 per tutor per year, excluding other program costs such as salaries (our check: US$1,419.66 over 388 tutors is about US$3.66 each for two months, about US$22 a year, consistent with the authors' rounded figure); the study did not measure end-of-year state test scores.
randomised evaluation of a six-week after-school program using Microsoft Copilot (GPT-4) for English with first-year senior secondary students in Edo, Nigeria; the authors' World Bank blog of 9 January 2025 notes attendance disrupted by flooding, teacher strikes and after-school work.
- 7
Department for Education (South Australia), "EdChat" (page last updated 24 August 2026): education.sa.gov.au
Department for Education (South Australia), "AI tool to roll out to all public high schools," 15 September 2025: education.sa.gov.au
Back to textWhat we took from it
a generative AI chatbot developed with Microsoft using the same model as ChatGPT, for learning and teaching, not available to the public; made available to public secondary students from term 4, 2025 through a phased roll-out; includes filters that block inappropriate material, with notifications going by default to the principal; data is stored in Australia and not sent back to the ChatGPT server; a 2024 external review of the proof of concept (with Telstra-Quantium and The Learning Future) focused on uptake and usage, with more than 10,000 students in the trial and proof of concept phases and 94% of student interactions related to curriculum subjects.
in its first full year 41% of students and 36% of educators had used the tool at least once (independent Insights Report); it reports 93% of student prompts in school hours related to curriculum subjects; no learning outcome or cost figures reported.
- 8
UK Department for Education, "450,000 disadvantaged pupils could benefit from AI tutoring tools," press release, 26 January 2026: gov.uk
UK Department for Education, "Generative AI: product safety standards," updated 19 January 2026 (applies to England): gov.uk
Back to textWhat we took from it
tools that could support up to 450,000 pupils a year on free school meals in years 9 to 11 alone, described as levelling the playing field for those who cannot afford private tutors, co-created with teachers from summer term 2026 and available to schools by the end of 2027; they will complement "high-quality face-to-face teaching" and not replace it; the government will robustly test the tools and develop benchmarks; no cost for the tools is given; the release cites evidence that one-to-one tutoring can accelerate learning by around five months (second-hand).
products should avoid giving "final answers, full solutions, or complete worked examples by default", use progressive disclosure from hints, track when learners offload thinking, not use sycophancy or design interactions to prolong use, and not use personal data for commercial purposes without confirmation of an appropriate lawful basis (all "expect" standards); filtering and activity logging are "must" requirements.