AI in education: what a training provider can actually automate

The conversation about AI in education has been stuck between two headlines for two years: the teacher replaced by a machine, and the student who outsources their homework. Neither describes the real problem of a language academy with 400 enrolments a year, a private vocational training centre, or a business school selling a €9,000 master's programme. The bottleneck there isn't pedagogical: it's administrative, it's about follow-up, and it's about margin.

A mid-sized training provider spends a striking number of hours on tasks that teach nobody anything: answering for the fifth time how much the course costs, chasing a missing ID document to close an enrolment, marking 120 exercises against the same rubric, and realising in December that the student who hasn't logged in for three weeks is never coming back. That's where AI has measurable room to work, and that's where to start before touching anything that affects a grade.

This guide walks through the four areas that work today in training providers —admissions and enrolment, automated tutoring, assisted marking and dropout detection—, what data each one requires, and where the line sits, especially when the student is a minor.

What does AI in education change, and what doesn't it?

The first thing is to separate two things that get constantly conflated: using AI to teach and using AI to run the centre. The second is less glamorous and is the one that pays for itself within an academic year.

What genuinely changes today:

• Response time to prospects. A question about prices, dates or credit transfers gets answered in seconds and at any hour, which is when people actually shop for training.
• The cost of administrative friction. Incomplete paperwork, half-finished payments, mishandled funding claims: repetitive work that can be guided and checked automatically.
• The granularity of follow-up. Moving from "that group is doing so-so" to "these nine students haven't been active for 12 days and seven of them enrolled in the final stretch of the deadline".
• The marginal cost of feedback. Giving a first reasoned response on an exercise stops depending on whether the teacher has a free Sunday.

What doesn't change: who grades, who decides whether a student progresses, who calls the parent, and who designs the programme. A model doesn't know whether the student is going through a rough patch or whether this week's weak exercise is a blip or a trend. That reading belongs to the tutor, and it's what sustains the value of a centre against a recorded course.

If AI in your centre doesn't free up teaching hours or reduce dropout, it isn't contributing: it's decorating.

How do you automate enrolment and admissions?

This is the lowest-risk, fastest-return case, because the data involved is administrative and the process is already defined. Three layers, from least to most ambitious:

1. Handling enquiries. An assistant built on your own catalogue —programmes, entry requirements, dates, prices, credit transfers, delivery modes— answering on the website and on WhatsApp. The key isn't the model, it's that it answers from your documents rather than from something it made up: that's the pattern we cover in RAG for business. If the assistant can't find the answer, the right behaviour is to say so and hand off, not to improvise an entry requirement.

2. Collecting and validating paperwork. Prior qualifications, ID, proof of payment, funding documentation. The heavy lifting here is reading documents and checking that what's on the paper matches what's on the form: extraction, field verification, and a flag when something doesn't add up. That's document data extraction applied to the registrar's office.

3. Chasing half-finished enrolments. Most centres lose enrolments in the gap between "I'm interested" and "paperwork complete". A flow that detects what's missing, from whom, and chases it with a specific message ("we still need proof of your prior qualification") recovers a share of that leakage without anyone reviewing a spreadsheet every Monday.

What to watch: the assistant must never promise places, prices or credit transfers that depend on a decision by the centre. An assistant that hands off too often is better than one that creates a commitment you later have to walk back.

Can AI tutor a student?

It can accompany, not tutor. The distinction matters because it's where most money gets wasted in this sector.

What works reasonably well:

• Answering content questions 24/7 about the course's own material: explaining a concept again, offering a different example, generating extra practice exercises.
• Conversational practice in language learning, where the value lies in repetition without embarrassment and immediate correction, not in assessment.
• Guiding the student through the pathway: what's due this week, what's still outstanding, what to review before the exam.

What doesn't work, however hard the vendor sells it: an "adaptive tutor" that decides a student's pathway from half a dozen interactions. Real personalisation needs enough signal —attempts, timings, error types by content area— and most centres don't have it yet. Without that history, what you're buying is a content sequence with a marketing layer on top.

A practical rule: the content assistant should be scoped to the course material and should say it doesn't know when asked outside that scope. An assistant that answers everything blurs the line between what the centre certifies and what comes out of a general-purpose model.

How much marking can AI assist with?

This is the question that generates the most heat in the staff room, and the honest answer has two halves.

You can assist the first pass: applying a rubric, flagging objective errors, detecting missing required sections, summarising the most repeated mistakes across the group, and drafting a comment for the teacher to edit. On non-graded practice exercises, you can give feedback directly to the student with an explicit note that it's automated.

You should not delegate the grade. For three reasons that are operational rather than philosophical: grading has legal consequences (appeals, credit transfers, qualifications), the model isn't stable between runs, and the centre has to be able to explain why a piece of work scored 6 and not 7. If you can't reconstruct the reasoning, you can't defend the grade.

Where the real saving sits: in high-volume, low-stakes formative marking —weekly exercises, self-assessments, language practice— and in the time teachers spend writing the same comment forty times.

| Type of work | Role of AI | Who decides | |---|---|---| | Self-assessment and practice | Full automated feedback | Nobody: it isn't graded | | Weekly exercise with a rubric | First pass + draft comment | Teacher reviews and publishes | | Final project or exam | Formal checks and summary | Teacher grades in full | | Grade appeal | None | Board or management |

Before putting any of these systems into production, build your own test set: 30-50 exercises already marked by your staff, run them through the tool and measure the deviation. It's the same logic as AI model evaluation applied to a teaching case: if the deviation on your real material is high, the tool doesn't go in, however flawless the demo was.

How do you spot dropout before it happens?

In money terms, dropout is the biggest problem a training provider has: the student who falls away in November has already consumed the acquisition cost and stops paying the remaining instalments. And it's the case where the data you already hold goes furthest.

The signals that genuinely anticipate a drop-off are boring, and they live in the virtual campus and in administration:

• Days without logging in, and above all the trend: going from daily access to once a week.
• Late or missing submissions, two weeks running.
• Falling attendance at live or in-person classes.
• A failed direct debit, which almost always precedes withdrawal.
• Students who enrolled in the final stretch of the deadline or with a recorded hesitation in the CRM.

From that you build a risk model which is, underneath, the same problem as customer churn prediction in any other subscription business. Two warnings that decide whether it's useful at all:

The window matters. Predicting that someone will drop out "at some point during the course" doesn't let you act. Predicting high risk in the next three weeks does, because a tutor's phone call fits inside it.

Without an intervention protocol, the model is pointless. The risk list has to reach a named person, with a deadline and a defined action. If the report stops at the management meeting, the model is just documenting dropout more precisely.

What data does each case require?

The short answer to "where do I start" is almost always: with whatever you already have the data for. This table is the filter we use in a first conversation with a centre:

| Use case | Minimum data needed | Difficulty | What it returns | |---|---|---|---| | Admissions assistant | Up-to-date catalogue, requirements and prices | Low | Instant answers, fewer repeated enquiries | | Paperwork validation | Digitised documents + form fields | Medium | Enrolments closed sooner, fewer complaints | | Course content assistant | Course material, organised and current | Medium | Questions answered out of hours | | Assisted marking | Written rubric + marked history | Medium | Teaching hours freed up | | Dropout risk | 2-3 intakes of activity, submissions, attendance and payments | High | Timely intervention, enrolment retained |

Notice that every row depends on the data being organised and trustworthy. In centres with three disconnected systems —virtual campus, admissions CRM and billing software— the first job isn't AI, it's data quality and integration. It isn't the part that makes the brochure, but it's the part that decides whether the rest works.

What does data protection require when the student is a minor?

Education handles two categories that raise the bar: data on minors and, sometimes, data on learning support needs that counts as health data. The minimum before putting any of the cases above into production:

• A clear legal basis for the processing, and a particularly careful one for dropout risk profiling: using activity data to offer support is not the same as using it to make decisions that affect the student.
• Plain-language information for families and students: what is analysed, what for, and who sees it. Finding out through rumour destroys trust faster than any technical error.
• A data processing agreement with the vendor, covering where the data sits and an explicit commitment not to use it to train general models.
• Minimisation: calculating dropout risk doesn't require the content of the assignments, activity metadata is enough.
• No automated decisions with a significant effect on the student without human involvement.

If you want the detail on how all of this gets documented, we develop it in the guide on GDPR and AI. And if the centre is going to open up tool use to teaching staff —which is what ends up happening, with or without permission— it's worth having a one-page AI use policy first, setting out what can and can't be pasted into a chat window.

Mistakes training providers keep making

• Starting with the pedagogy. The adaptive pathway is the flashiest case and the one that demands the most data. Start with admissions and dropout, which pay for themselves within a course.
• Buying the campus platform's AI module without testing it. Demos go well with demo content. Ask for a trial on your exercises and your rubric.
• Leaving the assistant unscoped. An assistant that answers everything will eventually give out wrong enrolment information, and the complaint lands on the centre, not the vendor.
• Leaving teaching staff out. If teachers believe assisted marking is the first step towards replacing them, the project stalls. The honest argument is the time it gives back, and it has to be backed with evidence. That's change management, and it starts before the purchase.
• Measuring usage instead of outcomes. The indicator isn't how many questions the assistant answered: it's how many enrolments closed sooner, how many hours were freed and how much dropout fell against last year.

Frequently asked questions

Can AI grade work at a training provider?

It can propose a mark against a rubric, but a teacher has to sign it off. Grades have academic and legal consequences, and the centre needs to be able to explain the reasoning in an appeal — which doesn't hold up if the decision came from a model with no human review.

What do I do if students use AI to write their assignments?

Assume they will and redesign assessment accordingly: more oral defence, more in-class work, more staged submissions where the process is visible. Generated-text detectors have a false positive rate far too high to base an accusation against a specific student on.

How much data do you need to predict dropout?

As a reasonable reference, two or three complete intakes with platform activity, submissions, attendance and payment status. With less history you can still set up simple rule-based alerts —days without logging in, failed submissions, bounced payment— which already catch a good share of cases without any model.

Is it worth it for a small academy?

For an academy with fewer than 200 students, the clear cases are the admissions assistant and rule-based inactivity alerts; both are cheap to set up and show up in enrolment numbers. A trained dropout model is hard to justify until there's enough volume and history.

---

The pattern we see in the centres that do this well is nothing spectacular: they get the admissions and campus data in order first, automate the administrative side, and only then move into anything that touches the classroom. If you're weighing up where to start at your centre, in the audit we review which cases have enough data today and which should wait, with your systems in front of us. And if you'd rather talk it through first, let's talk.