AI for training providers
Ask ChatGPT for a 4 session course on working safely at height. 3 minutes later you have a programme, learning goals per session, a case and a 20 question test. Then the same HR manager opens your 14,500 euro quote for an in company course and asks the question you hear more often this year: what am I actually paying for?
The honest answer is that the course material was never the work. The work is per learner: the assignment someone has to read and grade, the feedback someone has to write, the exam that has to be taken and recorded, the certificate that has to be right and expires in 3 years, and the file your quality mark auditor wants to see. That part costs the hours, and it grows with every enrolment. That is exactly where AI for training providers does the most, as long as an assessor with a name keeps deciding.
This article is for owners and directors of private training providers, exam bodies and certification bodies with 10 to 200 people: trade courses, safety and technical training, healthcare, logistics, finance. You put thousands of learners through per year, you work with a fixed group of trainers and assessors, and your growth stops at the number of assessors you can find and train. It is not for you if you are a solo trainer running 20 courses a year, if you are building an e-learning platform for consumers, or if you are a school under the education inspectorate.
In this article you will learn where the hours per learner really go, why a course generator does not change your revenue, which part AI does and which part fixed rules do, what ISO 17024 and the AI Act say about who decides, a worked example with 8,000 learners a year, what not to automate and how to start without putting your quality mark at risk.
Key takeaways
• Writing course material is a one time cost per course. Grading, feedback, exam administration and certificates are costs per learner, and they grow with your revenue.
• A course generator makes a generic course. Your value is your assessment model: what you consider a pass, which feedback works, what an auditor accepts.
• The heart of the automation is a learner file: 1 line per step per learner, with assignment, grade, feedback, exam, certificate and validity.
• AI reads the assignment and drafts a grade and feedback from your model. Fixed rules guard completeness, deadlines and validity. The assessor decides and signs.
• ISO/IEC 17024:2026, published on 31 March 2026, asks certification bodies to show human oversight of AI. The AI Act lists AI that assesses learning outcomes as high risk, with obligations from December 2027.
• In the worked example an assessed learner goes from 90 to 30 minutes. With the same team the provider goes from 3,000 to about 5,000 assessed learners a year.
• The register of certificates and expiry dates becomes a product employers pay for in year 2.
• Leave the pass or fail decision, the observation at a practical exam and every suspicion of fraud with a person.
AI for training providers: why a course generator does not change your revenue
The market is full of tools that write courses. You enter a topic, pick a level and a duration, and get a programme with slides, exercises and a test. For a company that wants an internal training quickly, that feels like progress. For you as a training provider it feels like competition on the part you already did fastest.
Because how long does it take to write a module when you have taught that course for 12 years? You have material, you update it when a standard changes, you let a colleague read along. That is a few days per year per course. The weeks of your team are somewhere else.
They are in the 40 practical assignments that arrive on Friday and have to be graded by Monday. In the feedback that differs per learner, because the mistake differs. In the exam that has to be scheduled, taken, marked and recorded. In the certificate that needs the right name, date and code, and in the email from an employer who wants to know which of his 60 employees have to renew next quarter. And in the file you open every 2 or 4 years for the audit of your quality mark.
A course generator does nothing about those hours. Worse, it makes the material cheaper and leaves the cost per learner untouched. That is why training providers with an AI tool produce new courses faster and still cannot handle more learners. The volume is not in the writing. It is in grading and administration, and that happens per person.
Your customers are not standing still either. According to Statistics Netherlands (CBS), 27% of Dutch companies with 10 to 50 employees used AI in 2025, up from 19% in 2024. The HR manager reading your quote has more and more often tried something themselves. That changes the question you have to answer: not "can you deliver a course", but "what do you do that we cannot do ourselves".
Where the hours per learner really go
Ask a training manager where the time goes and you rarely hear "course material". You hear: marking, feedback, administration, and chasing evidence. That work has a few fixed shapes, and they come back with every learner.
• Intake and placement: which course fits, which exemptions apply, which prior training is missing. For in company requests also the quote with a tailored programme.
• Reading and grading: the practical assignment, the portfolio, the reflection report or the open questions, against the assessment model.
• Writing feedback: what went well, what is missing, what the learner has to fix. Different per person, but with the same 20 mistakes coming back.
• Exam: schedule, take, mark, set the result, handle the appeal.
• Certificate and register: issue it with the right details, register it with the sector register, track its validity and offer the refresher on time.
• Quality file: evaluations, trainer qualifications, test matrices, complaints, the evidence your quality mark or accreditation asks for.
Grading and administration is where it goes wrong. Not because it is hard, but because it is scattered and different per person. The assignments sit in the LMS, the grades in an Excel per trainer, the feedback in emails, the certificates in another system and the expiry dates in the planner's head. Whoever wants to know how far a learner is has to ask the assessor. And when that assessor is ill or leaves, the colleague starts over.
We saw the same pattern at a subsidy advisory: 5 advisors worked 5 ways in 5 tools. The craft was the same, the recording was not. With training providers it is stronger still, because every trainer knows their own course and brought their own way of grading.
Your assessment model is the value, not your course material
Here is the shift most training providers miss. A generic AI can write a generic course. Every employer can do that now. What an employer cannot do is apply your judgement: when a learner's risk inventory is good enough, which 3 mistakes in a welding report mean someone really does not understand it, which feedback gets someone working instead of defending.
That judgement now sits in the heads of your best assessors and scattered across thousands of earlier grades. As long as it sits there, no AI can use it. And no new assessor either, except by shadowing for a year.
So the first step is not buying an AI tool. It is recording your standard: per course what you expect per criterion, which evidence is sufficient, what the typical mistake is and which feedback you give then. That is the same foundation we describe in the article on setting up an AI knowledge base. Without that standard, AI mostly gives you a faster intern who writes nicely.
With that standard your sales conversation changes too. The employer no longer pays for a course ChatGPT also writes. They pay for your standard, applied to their people, with a named assessor under it and a certificate an auditor accepts.
The learner file: what AI does, what rules do, what the assessor does
The heart of a good automation is not a chatbot in your LMS. It is a learner file: 1 line per step per learner, from intake to recertification. What was handed in, when, what AI made of it, what the assessor decided, which feedback was sent, which certificate was issued and when it expires. Everything AI does updates that file. Everything the assessor does happens from that file.
What AI does
• Reads the submitted assignment and prepares a proposal per criterion from your assessment model: pass, fail or doubt, with the passage it is based on.
• Drafts the feedback in your tone, from your earlier feedback on the same mistake.
• Checks whether a portfolio is complete and which evidence is missing, and writes the email to the learner.
• Drafts a programme and a quote for in company requests from your own catalogue and earlier quotes.
• Reads the evaluations and summarises per trainer and per course what keeps coming back.
What fixed rules do
• An assignment only goes to the assessor when it is complete and handed in on time.
• Every grade has a named assessor, a date and a verdict per criterion. Without that, no result.
• A certificate is only created when exam, payment and identity check out, and never by AI.
• Every validity has an end date. 6 months before it, a signal goes to the employer and the planner.
• In case of doubt, a fail and every suspicion of fraud, the file goes to a second assessor.
What the assessor does
• Reads the AI proposal next to the assignment and decides per criterion. AI proposes, the assessor ticks the box.
• Adjusts the feedback where it is too generic and signs it.
• Takes the practical exam, has the difficult conversation and handles the appeal.
That division is not a detail. AI is good at reading and sorting, and weak at knowing when something is just not enough. Fixed rules are boring and therefore reliable. The assessor is expensive and therefore used where judgement is needed. That is how we built a vehicle tax file processor too, where a file went from 4 to 8 hours to under 5 minutes: AI reads, code applies the rules, a person signs.

Want to put this on paper for your own courses before you talk to anyone? The free AI product requirements template helps you put the steps per learner, the links with your LMS and the cost of a wrong grade in 1 place. That gets you quotes you can compare.
ISO 17024 and the AI Act already say who decides
For exam bodies and certification bodies this is not taste, it is a requirement. On 31 March 2026 ISO published ISO/IEC 17024:2026, the standard personnel certification is accredited against. The new edition has its own clauses on AI for the first time. According to the e-Assessment Association, certification bodies have to monitor and validate AI outcomes, show human oversight and show their people can work with the tools. The American accreditation body ANAB has set 31 March 2029 as the final transition date, with a self attestation due on 1 October 2027.
The AI Act adds a layer on top. AI used to assess learning outcomes or to decide access to a course is listed in Annex III as high risk, for education and vocational training at all levels. The obligations for that category apply from December 2027. Whoever lets an AI tool say pass or fail by then has a problem with the auditor and with the law at the same time.
The good news is that the setup above produces exactly what both ask for: a log of what AI proposed, who looked at it, what that person decided and when. A learner file built this way is your evidence of human oversight. A loose chatbot next to your LMS never delivers that evidence.
Worked example: a provider with 30 trainers and 8,000 learners a year
An example, not a client. A safety and technical training provider has 30 trainers and assessors and 12 office staff. 8,000 learners come through per year. 3,000 of them hand in a practical assignment or portfolio that has to be graded. 5,000 only take a closed question exam. On top of that, 300 in company requests come in per year.
The minutes per assessed learner, today and with AI and a learner file:
• Intake and placement: 10 minutes becomes 5, because exemptions and prior training are checked automatically.
• Reading and grading: 45 minutes becomes 15, because the assessor gets a proposal per criterion next to the assignment.
• Writing feedback: 20 minutes becomes 5. The assessor adjusts and signs.
• Exam and administration: 10 minutes becomes 3.
• Certificate and register: 5 minutes becomes 2, and the refresher is offered automatically.
Total: from 90 to 30 minutes per assessed learner. That is 1 hour less, times 3,000 is 3,000 hours a year. For the 5,000 exam only learners the administration goes from 25 to 10 minutes, which is 1,250 hours. The 300 in company requests go from 3 hours to 1 hour per quote, which is 600 hours.
Together that is 4,850 hours a year. An assessor delivers about 1,400 productive hours. That is the capacity of almost 3.5 assessors, without a vacancy. Put differently: with the same team the provider can handle about 5,000 assessed learners a year instead of 3,000, if sales brings in that volume.
What does it cost? In this example a Blueprint of 7,500 euro, a foundation for the assessment model and the knowledge base of around 20,000 euro, and 2 automations: the grading assistant with feedback and the learner file with certificates and validity, 65,000 euro together. That is 92,500 euro, plus 2,500 euro a month for operations. Suppose the provider fills only a third of the new capacity: 650 extra assessed learners at an average of 650 euro. That is 422,500 euro of revenue a year, without an extra assessor. Payback is then under a year.
Those numbers are an example. Your own numbers come from measuring your own courses, not from an article. But the ratio shows up at almost every provider: grading costs more than writing.
What not to automate
There is a limit, and for training providers it is a sharp one. You sell a certificate an employer trusts to mean someone can stand on a scaffold safely. A wrong result costs more than a few hours.
• The pass or fail decision. AI may propose, the result belongs to a named assessor. The standard and the law ask for that, and so do your customers.
• The practical exam. Observing whether someone performs an action safely is human work. A camera with AI grading it is precisely the use that falls under high risk.
• Fraud. AI may flag that 2 reports look suspiciously alike. An accusation comes from a person, after a conversation. Reliable detection of AI written text does not exist, by the way, so do not build your policy on a detector.
• The appeal. Whoever disagrees with a result gets a second assessor, not a second algorithm.
And then the data. You hold names, dates of birth, sometimes a citizen service number for a sector register, grades and sometimes medical statements. Those do not belong in a free chatbot. A trainer who pastes learner portfolios into a public AI tool to mark faster has written the first complaint of the year personally. A proper setup keeps the data in the EU, separated per client, with a log of who saw what. That is also the first thing an employer with a procurement department will ask you.
How to start
Do not start with the tool. Start with 1 course and 50 graded assignments.
• Pick the course with the most graded assignments, not the course you like most.
• Take 50 recently graded assignments and measure where the minutes went. Use your time records or a week of shadowing, not your memory.
• Record with your 2 best assessors what you expect per criterion, which mistakes come back and which feedback you give then. That is your standard.
• Build the grading assistant for that 1 course and let AI only propose and draft. The assessor works as always, but with a proposal next to it.
• After 3 months, measure the minutes per assignment and how often the assessor accepts the proposal. Only then come the file, the certificates and the next course.
The order matters. Whoever starts with the course material speeds up the part that was already fast. Whoever starts with grading touches the hours that grow with every learner. And whoever measures how often the assessor deviates from the proposal has the evidence of oversight the auditor will want to see.
From course to certificate register
There is a step further, and for many providers it is the real opportunity. Once the file exists, the employer can look into it themselves. Which of their people are qualified, what expires when, who has to renew, and enrol with 1 click. The refresher then stops being a yearly calling round by your planner and becomes an ongoing service with a fixed price per employee per year.
That is the difference between a provider that sells course days and a provider that sells a product. The trainer and the assessor remain necessary. But the volume comes from the platform, not from extra people on the phone. Our article on building a client portal for service firms explains why such a portal starts where the hours are, not at an archive. This is usually a year 2 step. First the file has to work inside your own team.
Who builds it, what happens after launch, and if month 1 disappoints
Codelevate turns Dutch B2B service companies into software companies: we map the expert process you run by hand, automate it so the same team handles the volume, and build the platform your customers pay for, with 2 engineers by name from the blueprint to year 2.
Who works on it? The same 2 engineers who read your assignments and recorded your assessment model with your assessors. No rotating team, no handover after the sale.
What happens after launch? From month 4 it runs in operations: monitoring, an engineer with a name on the alert, the AI cost per learner with a cap, a monthly 1 page report and updates when a standard, an exam requirement or a model changes. The response time is set in the Blueprint. A new test matrix or a change in a sector register is then a planned update, not a new project.
And if month 1 disappoints? Exit after month 1: pay for what is delivered, keep everything. Your assessment model, your file and your code stay yours.
Want to go further on your own first? The free AI product requirements template helps you map your own courses before you have a conversation.
Send us what you have
Send us what you have: an overview of your courses and numbers, an anonymised graded assignment or just your time records. 1 of the 2 engineers reads it with you in 20 minutes, and you get 1 of 3 answers in writing the same day: do it in house, hire, or the Blueprint.



