Language schools mostly use AI for back-office work, not teaching: drafting enquiry replies, marking placement tests for a first pass, building timetables, writing agent updates and brochures, summarising student feedback, matching homestays and producing marketing. Each is safe to hand over only with a named person checking the output.
The safe pattern is the same each time. AI produces a draft or a shortlist, a member of staff checks it against the facts in your own systems, and anything involving a student's welfare, visa status or payment never goes through an assistant at all. The eight jobs below are ordered roughly from lowest risk to highest, so a school can work down the list.
1. Enquiry replies: the job with the clearest payoff
A school's inbox is dominated by a few questions: course dates, fees, what is included, how to get a visa letter, how to reach the school from the airport. They arrive in several languages and often outside office hours. An assistant given your price list, term dates and policy text can draft a reply that a person reads and sends.
Here is how a prompt might look, saved once in a shared project (ChatGPT Projects, Claude Projects or a Gemini Gem; Google's help page says Gems are being replaced by skills, from 17 November 2026 on personal accounts and from March 2027 on Workspace business accounts):
You draft replies for the admissions team of a language school. Use ONLY the facts
in the attached price list and policy sheet. If the answer is not there, write
"[CHECK WITH ADMISSIONS]" instead of guessing. Reply in the language of the enquiry,
in a warm, plain style, under 120 words. Never promise a visa outcome.
Sign off as [your name], Admissions.An illustrative result for a student asking about a four-week general English course starting in March: a short greeting, the price and what it includes taken straight from the sheet, the start date, the enrolment link, and a line asking about their level. The reviewer's job is to check the figures against the sheet and to look for anything that sounds like a promise.
Risk: a confident wrong fee or date. Mitigation: the "do not guess" instruction, a current price sheet (an out-of-date one produces out-of-date answers) and a review step for the first month. Count the edits you make; once you are changing fewer than one reply in ten, you can reduce the checking to a daily spot check.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now2. Placement tests: marking a first pass, not deciding the level
New students need a level before they can be put in a class. AI can score a short written sample against your level descriptors and suggest a band, which a teacher confirms in a two-minute check. It does not replace the speaking interview, and it must not be the only input.
Say a school uses six levels and asks for a 100-word written answer plus a 15-question grammar test. The teacher pastes the answer, with the student's name removed, and asks the assistant to compare it with the descriptors and list the evidence for its suggestion. A useful output reads: "Suggests level 3: uses past simple correctly, inconsistent with present perfect, vocabulary limited to everyday topics; two examples quoted." The teacher confirms or overrides.
Risk: assistants tend to be generous with fluent-sounding writing, and they cannot tell if a student used a translation tool or someone else's help. Mitigation: check the model against twenty past students whose levels you already know before using it live, and keep the speaking check as the deciding step. A wrong placement is expensive: the student is unhappy, the class dynamics shift and a refund conversation may follow.
3. Timetables and room allocation: useful with hard constraints
Building a timetable is a constraint problem: teachers' availability, room capacity, levels that cannot clash, a two-week intake shift. A general assistant can draft one from a clear list of rules, and spreadsheet-based tools do it more reliably. For the broader picture, see language school software compared and what timetable generators solve and miss.
The practical use is a first draft you then adjust, plus quick answers to "what if teacher X is off for a week?". Give the assistant the rules in a table, and ask it to list any rule it could not satisfy rather than quietly bending one. Check by counting: every class has a teacher, every room fits its group, no teacher is in two places.
Risk: a timetable that looks complete but double-books a teacher in week three. AI assistants make arithmetic and consistency slips on long tables. Mitigation: validate the output in a spreadsheet with a simple duplicate check on teacher and time, before anyone sees it.
4. Agent and partner updates: the report that nobody enjoys writing
Schools that recruit through agencies send regular updates: attendance, progress, any problems. Producing these by hand for 40 students takes an afternoon. A template with fields filled from your management system, polished by an assistant, shortens it dramatically.
A before and after helps. Before: an admin copies attendance figures into a Word document and writes three sentences of progress per student. After: the figures are exported to a sheet and the assistant turns each row into a two-sentence note in a consistent, courteous tone, which a teacher or director of studies skims. Illustrative output: "[First name] attended 18 of 20 lessons this month and moved from level 3 to level 4 after the progress test. Speaking confidence is improving; extra practice on past tenses is recommended." The agent receives a consistent report without you writing 40 of them from scratch. Marketing to agents is a separate task, covered in agent emails and multilingual marketing.
Risk: invented detail. A model asked to make a report "sound positive" may add praise the data does not support. Mitigation: tell it to use only the fields supplied and to leave a gap rather than fill one, and have the teacher confirm any comment about a student's behaviour or welfare.
5. Brochures and web pages in several languages
Translation is where AI is strongest and where schools most often skip the check. A machine translation of your brochure into four languages reads fluently, and that fluency hides errors: a wrong word for "full board", a course name rendered as a literal phrase that means something else, a date format that confuses.
DeepL and the general assistants all produce usable first drafts. DeepL's paid Pro plans include glossaries (one on the Individual plan, five on Team, per its plan page) and state that Pro texts are never used to train its models, which suits a school that translates enrolment terms. The step that matters is a native-speaker review of every language you publish, ideally from a current or former student or an agent you trust. Keep a glossary of your fixed terms (course names, "homestay", "intensive", "exam preparation") and give it to the tool with each job so those words stay consistent. The workflow is set out in the brochure translation check.
Risk: published errors, and legal-sounding terms (cancellation, refund) translated loosely. Mitigation: native review, and keep the English terms and conditions as the version that governs. Ask an adviser if you are unsure how that sits with consumer rules where your students live.
6. Student feedback summaries
End-of-course surveys produce hundreds of free-text comments that nobody has time to read closely. An assistant can group them by theme and count mentions: "accommodation cleanliness: 14 comments, mostly negative; teacher quality: 31, mostly positive; afternoon activities: 9, mixed". That points the director of studies to the three things worth discussing at the next meeting.
Say a school gets 220 responses over a term. Export the comments without names, split the file if it is long, and ask for themes with an example quote and a count. Then check the counts by searching the spreadsheet for two keywords yourself, because assistants miscount. A neat trick is to ask it to list the comment numbers behind each theme so you can verify them.
Risk: averaging away the single serious complaint. A comment about a safeguarding concern or a homestay problem is one line in 220 and can be lost in a summary. Mitigation: run a second pass with the prompt "list any comment that mentions safety, harassment, health or discrimination, verbatim", and route those to a named person the same day.
7. Homestay matching and host communication
Homestay coordinators match students to host families using a mix of preferences: diet, pets, smoking, distance, host availability. A spreadsheet and a clear scoring rule do most of the work, and an assistant can draft the introduction messages to host and student.
The honest advice is to keep AI away from the matching decision for anything that touches welfare. Use it for the writing: a friendly arrival message to the host with the student's first name, arrival date and dietary needs, and a reminder of house rules. Matching still rests with the coordinator, who knows which hosts are reliable. Personal details here are sensitive, so use first names and the minimum necessary, and check your plan's data terms before pasting any of it.
Risk: a mismatch on something a student cares about, such as an allergy, and any leak of a student's personal details. Mitigation: a human confirms the allergy and safeguarding fields every time, and the assistant only ever sees the fields it needs to write a message.
8. Marketing: posts, newsletters and web copy
This is the job most schools try first and the one with the least to protect them from error, apart from claims. An assistant can draft social posts, a monthly newsletter to former students and an agent mailshot from your course calendar. The problems are the usual ones: a generic voice, and statements that are not true for your school.
Keep a "facts that are true about us" document and feed it in each time, so the draft uses your real class sizes, locations of activities and exam results. Anything with a claim (pass rates, accreditation, "the best") is written or approved by a person. The tutorial on five education workflows to automate first puts marketing in context with the other jobs.
Risk: an unverifiable claim in an advert, and bland copy that reads like every other school's. Mitigation: a claims check before anything is published, and your own photos and stories in place of generic ones.
Two checks that catch most of the damage
Across all eight jobs, two habits do most of the protecting. The first is a "facts sheet" per job: one page of current prices, dates, policies and level descriptors that the assistant must use and quote from. Most wrong answers are not the model inventing things from nowhere, but your own out-of-date information being used faithfully. Put a review date at the top of each sheet and a name next to it. A school that changes fees in January and forgets the sheet will send the old fees for months.
The second is a "reason" line. Ask the assistant to say where each fact came from, for example "price: sheet, line 4". A reviewer can then check in seconds rather than rereading the whole reply, and a line that cites nothing is the one to distrust. Say a teacher is checking twenty agent reports on a Friday afternoon: the cited source turns a fifteen-minute read into a five-minute scan of the lines that carry no source.
A risk table to pin up in the office
| Job | Who sees the output | Main risk | Check before it goes |
|---|---|---|---|
| Enquiry replies | Prospective student | Wrong fee or date | Compare with the price sheet |
| Placement first pass | Teacher only | Over-generous level | Speaking check decides |
| Timetable draft | Staff | Double-booked teacher | Duplicate check in a spreadsheet |
| Agent reports | Agent | Invented praise | Teacher confirms comments |
| Brochure translation | Public | Loose legal terms | Native-speaker review |
| Feedback summary | Management | Serious complaint lost | Second pass for safety words |
| Homestay messages | Host and student | Sensitive data leak | First names and minimum fields only |
| Marketing copy | Public | Unverifiable claim | Claims approved by a person |
The out-of-date fee sheet that quoted $40 too little
Say a school uses an assistant to draft enquiry replies and its price sheet still lists last year's intensive-course fee. For three weeks, six replies quote a fee that is $40 too low. Nobody notices because the replies read well and the reviewer is checking tone, not figures. The error shows up when a parent forwards the quote at enrolment and the office has to decide whether to honour it. The fix is cheap: a monthly review date on the sheet, and a reviewer who checks the numbers first and the tone second.
When review time eats the saving on enquiry replies
Review time is the figure most owners forget. If a drafted enquiry reply takes two minutes to read and correct, against six to write from scratch, you save four minutes a message. At 60 messages a week that is four hours. But if the reviewer is checking each one against three documents because the sheet is messy, review creeps to five minutes and the saving disappears. Time one real batch of ten before you decide a job is worth keeping, and tidy the facts sheet if review is slow.
A mid-sized school's first term with AI, in hours
Here is an illustration with invented figures. Say a school with 180 students at peak and a four-person office spends roughly 25 hours a week on enquiries, reports and timetable fixes. The owner starts with jobs 1 and 4 only.
- Week 1: collects the 25 most common enquiry questions, writes the price and policy sheet, builds the reply prompt above.
- Weeks 2 to 4: every draft is read before sending; the team logs each edit. About a third need changes, mostly dates.
- Month 2: edits fall to roughly one in ten. Agent reports move to the sheet-and-template method, cutting an afternoon's work to about 90 minutes.
- Month 3: the school adds feedback summaries and keeps placement, timetables and homestay as they were.
Illustratively, if the enquiry replies take three minutes instead of eight and the school answers 60 enquiries a week, that is five hours back each week, and the agent reports return another two or three. Those hours go into following up leads who did not reply, which is where enrolments are actually won. The sums depend entirely on your volumes, so run them on your own enquiry count rather than borrowing mine.
What stays out of AI altogether
Some jobs are outside the list for good reasons: safeguarding and welfare conversations, visa and immigration advice (staff can explain your own letter process but should not guess at rules), refund and complaint decisions, and anything involving under-18 students' personal information beyond what a plain message needs. Write these into a one-page staff rule so a new starter does not paste a complaint thread into a chatbot to get a reply quickly.
Choosing where the data goes
Most of the risk in the list is about data rather than accuracy. Business plans of ChatGPT, Claude, Microsoft Copilot (work accounts) and Gemini in Workspace do not train on your content by default, according to the vendors' pages as of November 2026. Personal plans put the job on you: ChatGPT Free and Plus share chats for training until someone turns off "Improve the model for everyone", so an admissions officer drafting on a personal account may be sending enquiry text on different terms from the rest of the office. ChatGPT Business and Claude Team both start at two seats, at $20 a seat a month billed annually or $25 monthly. If you teach students from the EU, the GDPR applies to how you handle their details, so keep a short record of which tools see which data and ask a qualified adviser about anything beyond that.
Start with the lowest-risk job, prove the review step works, and add one more job every few weeks. If you would rather map this against your own enquiry volumes and systems, that is what my AI implementation consultation covers.
Questions language school owners ask
Is it safe to put student names into an AI assistant?
Only on a plan whose data terms you have read. Business plans from the main vendors do not train on your content by default, but that is a policy, not a guarantee of confidentiality. Where you can, work from templates with placeholders and keep passport details, health notes and safeguarding information out of any assistant entirely.
Will AI replace my admissions or student services staff?
In a school of any size it removes the repetitive part of their week: the same fee, visa-letter and arrival-day questions. People still handle upset students, edge cases, agent relationships and anything involving welfare. Most schools find the staff time moves to follow-up and student experience rather than disappearing.
Which job should a small school start with?
Start with enquiry replies. The questions repeat, mistakes are visible and easy to correct, and a draft reply that a person reviews saves minutes on every message. Placement tests and homestay matching carry more risk and come later, once you trust your review step.
Further reads
- Tutoring Centre Admin With AI: Enquiries, Timetables, Invoices — Similar admin for a tutoring centre: enquiries, timetables, invoices.
- Best AI Tools for Tutors: What a Private Tutor Can Actually Use — What a small teaching business can actually use.
- Hostel Software: Can AI Handle Guest Messages in Several Languages? — Handling messages in several languages at volume.
- Tour Descriptions in Several Languages: An AI Workflow — A multilingual copy workflow that transfers to course pages.
- AI Receptionist for a Hostel: Night Calls in Several Languages — Out-of-hours calls in several languages.
- Tutor Scheduling Software: Bookings, No-Shows and AI Reminders — Reminders that cut no-shows.
- How to Start a Language School: Accreditation, Agents and AI — What a new language school costs in year one, how accreditation and agents fit together, and where AI saves admin time from the first term.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT business pricing page and Data Controls FAQ, Claude pricing page, Microsoft Learn (Copilot privacy), Google Workspace generative AI privacy hub, Google Gemini help (transition from Gems to skills) and DeepL Pro plan page, all checked November 2026.