Four tasks of an online school where an assistant can help
A school with cohorts of students has four recurring jobs. Each has a different cost of error, so a different level of human involvement.
- The course programme. The assistant drafts a structure: topics, the goals of each lesson, a list of practical tasks. An error here is visible at once, because the course author reads the programme.
- Homework checking. The biggest time saving and the biggest risk: an error reaches the student as a grade or a comment in the school's name.
- Student support. Questions about access, deadlines, lesson format. An error in one line can promise a refund or a deadline that does not exist.
- Emails and announcements. Mailings about a cohort start, reminders. The same applies as for any shop; for an example calculation see the article on email campaigns.
Below we cover the two middle tasks, because that is where savings and risk are greatest. General rules for text requests, which suit the first and fourth, are in the article on a prompt for a business task.
Homework checking: what it costs
A conditional school: 300 students, each submitting 8 assignments a month, 2,400 checks in all. One check is the assignment text with grading criteria plus the student's answer as input (we take 3,000 tokens) and a detailed comment with a grade as output (500 tokens). The volumes are conditional; Ukrainian tokenises less economically than English, so for longer answers recalculate on your own.
Prices of three current models from the OpenAI API pricing page on 19 October 2026 (standard processing, short context, per million tokens): gpt-6-luna is $0.10 input and $0.50 output, gpt-6.1-sol is $2 and $10, gpt-6-astra is $10 and $50. The National Bank of Ukraine rate on 19 October 2026 is UAH 44.8625 per dollar.
| Model | One check | 2,400 checks a month |
|---|---|---|
| gpt-6-luna | $0.00055 (UAH 0.02) | $1.32 (UAH 59) |
| gpt-6.1-sol | $0.011 (UAH 0.49) | $26.40 (UAH 1,184) |
| gpt-6-astra | $0.055 (UAH 2.47) | $132 (UAH 5,922) |
Calculation: (3,000 × input price + 500 × output price) ÷ 1,000,000 × 2,400 × 44.8625. Volumes are conditional; OpenAI API prices and NBU rate on 19.10.2026.
The conclusion from the table: prices differ between models by hundreds of times, but even the highest sum is small at the scale of a school with three hundred students. So price is not the main constraint. The main thing is the quality of checking, which price does not show: it depends on how specific the criteria you gave are, and on the assignment itself. Test all three models on the same twenty assignments and compare with the teacher's grade, rather than choosing by price or name.
The sum does not include, for example, analysing images and video, checking code with execution, or long works of dozens of pages: for those, recalculate the token volume.
How many checks to read by hand
You cannot trust a model without checking, and rereading everything is pointless because then there is no saving. A calculation gives a sensible boundary. Let the share of faulty checks (a wrong grade or a comment that misleads the student) be p. If you reread n checks at random, the probability of seeing no error is (1 − p)ⁿ. To see at least one with 95% probability you need n ≥ ln(0.05) ÷ ln(1 − p).
| Error share p | How many to reread | What it means for 2,400 checks |
|---|---|---|
| 2% | 149 | about 48 errors in a cohort, you reread 6.2% of checks |
| 5% | 59 | about 120 errors, you reread 2.5% |
| 10% | 29 | about 240 errors, you reread 1.2% |
| 20% | 14 | about 480 errors, you reread 0.6% |
How to use it. Read the first 60 checks in full: that shows whether the error share is close to 5%. If you find more than three errors in sixty (over 5%), rewrite the criteria instead of enlarging the sample. If you find none, then with 95% probability the error share is below 5%, and afterwards you can reread 30 random ones every week. This calculation works only for a random sample, not "the first ones that came up": take checks from different students and different assignments.
Read separately all cases where a student appealed a grade and all grades that affect a certificate or a refund. A sample does not suit them: the cost of an error is higher than the cost of rereading.
Why an AI-text detector is not evidence
A school is easily tempted by an extra check: run the answers through an AI-text detector and return the "suspicious" ones. Look at what the author of such a tool said. In January 2023 OpenAI released its own classifier, and as of 20 July 2023 it withdrew it, citing low accuracy. In its own evaluation on a set of English texts it identified only 26% of AI-written texts as "likely AI-written" and wrongly labelled 9% of human-written texts.
Let us recalculate on a conditional group of 100 students where 10 really used AI (the share is conditional), at these rates:
| Situation | Calculation | Result |
|---|---|---|
| Students who used AI flagged correctly | 10 × 26% | 2.6 |
| Honest students flagged wrongly | 90 × 9% | 8.1 |
| Share of correct among all flagged | 2.6 ÷ (2.6 + 8.1) | about 24% |
So of every four "suspicious" students, three are honest. Modern detectors may have different rates, we have not checked them, and we do not claim they are the same. But the conclusion does not depend on the specific figure: when the share of those who really used AI is small, even a detector with a relatively low false-alarm rate produces many false accusations. So a detector cannot be grounds for lowering a grade or expelling anyone.
What works instead without having to assert who wrote it: an oral or written follow-up ("explain the third step in your own words"), tasks on material that was not publicly available, staged submission with drafts. This tests understanding, not the authorship of the text. Write the rules on AI in assignments into the contract or course rules before the cohort starts.
A student support bot
A bot in the school's chat is useful for questions whose answer is in the materials: how to log in to the platform, where the lesson recording is, when the next session is, what is in the plan. The choice between a builder, a ready assistant and custom development is covered in the article on a chatbot for business. For a school add three rules.
- Only from the course materials. Give the bot a base of answers (rules, schedule, instructions) and require it to answer only from it. To a question outside the base it says "passing to your curator" rather than inventing.
- It does not promise money, grades or deadlines. A refund, an access extension, moving a submission date, a grade and a certificate are a person's decision. The bot passes the request on and names the curator's response time, one you actually keep.
- It does not see more than it should. Do not connect payment data, grades and students' personal data to the bot until it is settled where they are processed: see the article on data for an AI assistant. If some students are minors, clarify the rules for processing their data with a lawyer; we have no conclusion for that case.
What to give the model and what not
- Are there written criteria by which two teachers would give the same grade?Without criteria the model grades by eye, and that is not reproducible. No → write the criteria first
- Did you read 60 checks by hand and find no error?Then the error share is below 5% with 95% probability. No → refine the criteria and repeat
- Does the grade affect a certificate, a refund or expulsion?An error here costs more than rereading. Yes → a person decides, the model gives only a draft
- Do you plan to use an AI-text detector as grounds for sanctions?False alarms outnumber correct ones. Yes → replace it with a follow-up and staged submission
A first step for a week
- Take one course assignment and write criteria for it: what is assessed, how many points for what, what counts as a mistake.
- Take 20 past answers the teacher has already checked. Give them to three models with the same criteria.
- Compare with the teacher's grades. Record how many match and where they diverge.
- Choose the model by those divergences, not by price.
- For a live cohort, read the first 60 checks in full, then 30 random ones every week.
Our position
A school should give homework checking against written criteria to a model, leaving to a person the sample and everything that affects a certificate, money and expulsion. The price of a check is too small to be a reason to refuse. Do not use an AI-text detector as grounds for sanctions. A support bot works while it answers from the course materials and hands money, grades and deadlines to a person.
The position does not apply to tasks a model cannot assess against criteria: creative work where the grade depends on the teacher's taste, oral answers, large projects of many files. It also does not apply if your students are minors or the grade data is subject to separate requirements: then a lawyer's consultation comes first. The sample calculation assumes random rereading and the same error share across all assignments; if one assignment is harder than the others, check it separately.
Prices and exchange rate checked on 19 October 2026 on the OpenAI API pricing page and at the National Bank of Ukraine; the information on the classifier is from OpenAI's page on its AI-text classifier (with the withdrawal note dated 20 July 2023). The arithmetic was recalculated in code. The number of students, assignments, tokens and the share of students who used AI are conditional. The article is not legal advice.