The first evening of a new B1 course. Twelve adults, one teacher, a lot of goodwill. By the end of the lesson the teacher knows two things the placement test didn't: two students are really A2 and will struggle from week one, and one is closer to B2 and will be bored by week three. The next week goes into reshuffling, apologising and quietly adjusting the pace for a group that isn't really one level.
Most small language schools place students with some mix of a paper grammar test, a few questions at the front desk and a short chat if a teacher happens to be free. It works most of the time. When it doesn't, the costs are real: students who drop out because the class is too hard, students who ask for a refund because it's too easy, and teachers who spend the first weeks fixing placement instead of teaching.
This post is a blueprint for a better placement process: one that tests the right skills, uses AI where it helps (scoring drafts, summarising evidence, spotting borderline cases), and leaves the decision with a teacher.
Why one level matters so much
Read those numbers as a teacher would. A student placed in B1 who is really at A2 is sitting in a class that assumes around 150 hours of learning they haven't done. At three hours a week, that's most of a year. No amount of goodwill closes that gap in a ten-week course.
The opposite mistake is quieter but just as costly. Students who are placed too low don't complain in week one. They stop coming in week four.
The blueprint
Step 1: Map your courses to the CEFR, precisely
Placement can only be as good as the course map behind it. "Intermediate" means different things in different schools. Start by writing down, for each course, the CEFR level it starts from and the level it aims for, using sub-levels where your courses are finer than the six main levels.
| Level | What learners can do (short version) | Typical course | What placement should check |
|---|---|---|---|
| A1 | Simple phrases about themselves, very slow and clear speech | Beginners | Whether they're true beginners or "false beginners" |
| A2 | Routine tasks, familiar topics, short simple texts | Elementary | Past tense, everyday vocabulary, short exchanges |
| B1 | Main points on familiar topics, travel situations, simple connected text | Intermediate | Coherent speaking for a minute or two, writing a short message with reasons |
| B2 | Complex texts on concrete and abstract topics, fluent interaction | Upper intermediate | Arguing a point, handling unfamiliar topics, accuracy under pressure |
| C1 | Demanding texts, implicit meaning, flexible use for work and study | Advanced | Register, nuance, extended writing |
| C2 | Understands virtually everything, expresses themselves precisely | Proficiency | Usually placed by interview and exam history |
The descriptors in the second column are shortened from the CEFR's global scale. The last column is where your school's experience goes: what actually separates students who succeed in each course from those who don't.
Step 2: Ask before you test
A short questionnaire before the test saves time and improves the result: why they're learning, previous courses and certificates, how long ago they last studied, which schedule they need, and a self-assessment using a few CEFR "can do" statements ("I can describe my job and my daily routine", "I can follow a meeting on a topic I know").
Self-assessment isn't reliable on its own, but it's useful evidence. A student who rates themselves B2 and tests at A2+ needs a conversation, not just a number.
Step 3: A validated test for reading, listening and grammar
For receptive skills, use a validated, adaptive placement test rather than a home-made quiz. Adaptive tests adjust difficulty to the student's answers, so they're shorter and more accurate across levels. Established options include publisher and exam-board tests such as the Oxford Placement Test, Linguaskill or the free EF SET for English, and the Goethe-Institut's online placement test for German.
One firm rule: don't let a chatbot generate your placement test. Test items need to be calibrated against real learners. A model can write plausible questions, but it can't tell you whether they reliably separate A2 from B1.
Step 4: A short writing task, scored as a draft
Ask for a short piece of writing that fits the likely level: an email to a landlord, a paragraph about a recent trip, an opinion on a familiar topic. Here AI helps. A model can compare the text against a rubric aligned to CEFR descriptors (range, accuracy, coherence, task achievement) and produce a draft estimate with evidence:
Estimated writing level: B1, borderline A2+. Task achieved: the email explains the problem and asks for a repair. Range: everyday vocabulary, some repetition ("problem" four times). Accuracy: past tense mostly correct; frequent article errors; one sentence unclear ("I wait since Monday the man"). Coherence: basic linkers (and, but, because). Uncertain: only 90 words, so limited evidence for B1 range.
The teacher reads the text and the draft together, and decides. The draft saves time and makes the reasoning visible; it doesn't replace the teacher's judgment.
Step 5: A real conversation
Speaking is where placement most often goes wrong, and where a teacher adds the most. Five to ten minutes, in person or on video, following a simple structure: a warm-up about themselves, a description task, and one question a level above where they seem to be.
A model can transcribe the conversation and note features for the teacher (how long the student could speak without prompting, range of tenses, how they handled the harder question), but the teacher is in the room, and the teacher decides. If you record conversations, tell students why, get their consent and delete recordings once placement is final.
Step 6: Combine the evidence and decide
At this point you have a questionnaire, a receptive test score, a writing draft and a teacher's speaking judgment. The model can combine them into a placement proposal with one line of reasoning per skill. Most students will be straightforward. The matrix below shows which ones need a closer look.
Step 7: Check in week one, and learn from it
Placement doesn't end on day one. Ask teachers to confirm or flag every student's placement after the first week, and make moving between courses easy in that window. Then look at the flags each term: if students placed by one route are often moved, adjust that route. Over a few terms, your placement gets calibrated to your own courses, which no off-the-shelf test can do.
How it runs, end to end
- Inquiry and questionnaireStudent5 minGoals, previous learning, schedule, a few can-do statements.
- Adaptive testPlacement test20-30 minReading, listening, grammar and vocabulary, calibrated across levels.
- Writing taskStudent, then AI15 minShort text; the model drafts a rubric-based estimate with evidence and uncertainties.
- SpeakingTeacher5-10 minStructured conversation; optional transcript notes for the teacher.
- Placement decisionTeacherReviews the combined proposal, confirms or changes it, adds a note for the course teacher.
- Week-one checkCourse teacherafter lesson 2 or 3Confirms or flags; easy moves between courses in the first week.
Where the law comes in
Placement data, writing samples and especially voice recordings are personal data under the GDPR. Keep them in your school system, not in consumer chatbots, and delete what you no longer need.
When you don't need AI for this
A school with thirty new students a term and teachers who have time for a proper conversation with each of them doesn't need a model. A validated online test plus a structured ten-minute chat is already a good process. AI earns its place when volume is high, when placements happen outside teachers' hours (online enrolment across time zones, for example), or when you want consistent, documented reasoning across many teachers.
Tools that fit
Language school management systems handle enrolment, courses and payments; placement tests come from publishers and exam boards, or from your own calibrated item bank. The AI layer sits between them: it reads the questionnaire, drafts writing scores against your rubric, summarises speaking notes, combines the evidence into a proposal, and records the teacher's decision and the week-one outcome.
Questions school directors ask
Can AI score speaking on its own?
There are commercial automated speaking tests, and some are well validated for specific purposes. For placement in a small school, a teacher's conversation is still the most reliable way to judge how someone copes in a real exchange, and it doubles as the first contact with the school.
How accurate are AI writing scores?
Good enough to be a useful draft against a clear rubric, especially for mid-range levels, and less reliable at the edges and for very short texts. That's why the draft names its uncertainties and a teacher decides. Compare model drafts with teacher judgments for a term before relying on them.
What about online students we never meet?
Video conversations work well for the speaking step. For fully asynchronous enrolment, a recorded speaking task reviewed by a teacher within a day or two is a reasonable compromise.
Does the student see the reasons?
They should. A short explanation ("your reading is strong; speaking is where the B1 course will help most") makes placement feel fair and gives the student something to work on from day one.
Rule of thumb
Map your courses to the CEFR precisely, measure with validated tests, let AI draft and combine the evidence with its reasoning visible, and have a teacher make every placement. Then check in week one, and let the moves you make teach you where your placement goes wrong.
If placement at your school still means a paper test and a busy front desk, tell me how enrolment works today, and I'll suggest where to start. Translation agencies match texts to the right linguist in project intake, and music schools deal with a similar enrolment rush in school administration.
Building something with AI?
I help small businesses turn ideas into software that pays off. Tell me what you’re working on and get a free first assessment.