The pressure is real and mostly external: investors ask, competitors announce, and customers have started expecting it. Meanwhile the company is eight people, three of whom write curriculum, and none of whom have shipped a machine-learning system.
The good news is that the framing is wrong. Almost nobody in education is training models. The work is retrieval, evaluation and product design on top of models that already exist — application engineering, not research. And there are four ways to get there, with genuinely different costs.
Route 1: run it manually first
Before building anything, answer a week of student questions by hand from your own material, and log what was actually asked.
This is the highest-return week available, because it tells you whether you have an AI opportunity or something else. If students are asking things your content already answers and they could not find — retrieval will help enormously. If they are asking about deadlines, refunds and where the recording is, you have a support and navigation problem that a tutor will not fix and may obscure.
Teams skip this step and build the wrong thing surprisingly often.
Route 2: buy a white-label platform
Fastest route to a real product. You bring content and brand; the platform brings tutoring, assessment generation, analytics and the unglamorous infrastructure — payments, notifications, mobile, uptime.
The objection is usually that white-labelling means giving up the product. That depends entirely on what your product is. If a specific software interaction is your differentiator, this is a genuine loss. If your differentiator is the curriculum and the teaching, the platform is infrastructure, and infrastructure that already works is not something you were winning on.
Two things to settle in the contract: what you can export and in what format, and whether pricing is a flat fee or a per-learner royalty. A royalty means your best year is also your most expensive.
Route 3: bolt an AI layer onto what you have
Keeps your existing platform, adds retrieval on top. Reasonable if the platform is otherwise fine and you want to test appetite before committing.
The trap is running two systems that both think they own the course material. Content gets revised in one and goes stale in the other, and nobody notices until a student quotes a superseded answer back at a tutor.
Route 4: build it
Right when the AI behaviour is the differentiator — when there is a specific interaction nobody will build for you and it is why customers would choose you.
The cost is not the first version; a competent engineer will get something working quickly. The cost is permanent: models get deprecated, embedding models improve and require reindexing, content changes, and evaluation never finishes. Budget for the ownership, not the build.
The part nobody costs properly
Evaluation. Knowing whether AI answers are actually good is not an engineering problem — it needs someone who knows the subject to read outputs and judge them, repeatedly, forever.
That is the thing you already have and a vendor does not. Whichever route you take, keep subject experts in the loop on quality. It is the single highest-leverage use of their time in an AI rollout, and the first thing that gets dropped when a launch date slips.
What not to ship
Unreviewed generated questions. Generation is a fast first draft. A distractor that is accidentally also correct will surface in front of a paying student at the worst possible moment.
A tutor with no source citation. If a teacher cannot check where an answer came from, nobody can defend it when it is wrong.
Autonomous marking of written work. Assist a human marker; do not replace one. The reputational cost of a wrongly graded assessment is not recoverable by a support ticket.
An AI feature with no answer for "I don't know". A system that always produces something will eventually produce something confidently false, and that is the answer everyone remembers.
Sikgen AI is route two: a white-label AI LMS with tutoring, assessment generation and analytics already built, under your brand and domain. If you are weighing routes, talk to us — we will say plainly if building or bolting on suits you better.
Frequently asked questions
Do we need machine-learning engineers to ship AI features?
Almost certainly not. Nobody in education is training models from scratch — the work is retrieval, evaluation and product design on top of existing models. That is application engineering. Where teams get caught out is evaluation: knowing whether answers are actually good requires subject expertise, which you already have.
What is the cheapest way to test whether our users want AI features?
Run it manually before you build it. Answer a week of student questions from your own material by hand and note what they actually ask. If the questions are things your content already answers and students could not find, retrieval will help. If they are about deadlines and pricing, you have a support problem, not an AI opportunity.
Is white-labelling a platform giving up our product?
It depends what your product is. If the software is your differentiator, yes. If your content and teaching are the differentiator, the platform is infrastructure — and infrastructure that already works is not a competitive loss.