How to Create AI-Powered Personalised Learning Paths for Your Students

Personalised learning is the most oversold idea in EdTech and one of the most useful when scaled down. The pitch — a unique path for every learner — is unachievable for a tutor with sixty students. The workable version has four paths, not sixty, and runs on data you likely already have.

It is a rule that changes what a student does next based on what they have demonstrated. Not a bespoke curriculum per learner — a small number of branches with clear entry conditions. Practically: after each assessment, students are routed to remediation, standard practice, or extension.

The distinction matters because the ambitious version is what causes tutors to abandon the idea. Sixty individual paths is a full-time job. Four well-designed branches, assigned automatically from a test score, takes about an hour a week and captures most of the benefit.

This is also closer to what good teachers have always done informally. The difference is that the informal version lives in your head and degrades as batch size grows, while an explicit version survives fifty students and can be handed to someone else.

Three things, all cheap to collect: performance per topic rather than per test, time-to-completion, and where students stop. Aggregate scores are nearly useless for personalisation because they hide which specific concept failed — a student scoring sixty could have three different problems.

Tagging questions by topic is the highest-return administrative task in this entire article. Without it you know a student scored 12/20; with it you know they can do everything except simultaneous equations, which is an actionable statement rather than a number.

Use four tiers and route by assessment outcome. Most students sit in the standard tier; the branches handle the ends of the distribution. This gives you the substantive benefit of differentiation while keeping the number of things you maintain small enough to actually maintain.

1. Foundation — the prerequisite is missing. Send them back one concept before continuing; do not re-explain the current topic louder.
2. Standard — the main sequence. Most students, most of the time.
3. Consolidation — the idea is understood but not fluent. More practice, no new content.
4. Extension — mastered. Harder application problems, not simply more of the same.

The routing rule can be blunt and still work. Below forty percent on a topic goes to Foundation, forty to seventy to Consolidation, above eighty-five to Extension. Refine the thresholds after a term of watching what happens; do not agonise over them beforehand.

In generating the branch material and diagnosing patterns, not in deciding who goes where. Producing four versions of every practice set used to be prohibitive; AI makes it routine. That production cost was the real reason differentiation stayed theoretical for most tutors.

A prompt that does most of the work: “Here are five practice questions on simultaneous equations at standard difficulty. Produce a Foundation version assuming the student cannot yet isolate a variable, a Consolidation version with more of the same difficulty, and an Extension version requiring students to set up equations from a word problem.” One request, three tiers.

AI is also good at pattern-spotting across a batch. Paste anonymised topic-level scores and ask which concept is failing most consistently and what misconception would explain the specific wrong answers. That second half — the diagnostic hypothesis — is frequently more useful than the statistics.

Batch the routing weekly rather than continuously, and automate assignment rather than messaging students individually. One review session per week, four groups, four sets of material. The workload is roughly constant regardless of whether the batch is fifteen or fifty.

The failure mode is manual communication. If routing means sending fifty personal WhatsApp messages, you will do it twice and stop. It has to be a group assignment inside whatever platform holds your students, so that placing someone in Consolidation is one action rather than a conversation.

This is worth checking before you commit to a platform: can you group students and assign different content to different groups? Many tools built around one-directional content delivery cannot, which quietly makes personalisation impossible regardless of how good your design is.

Directly. NEP 2020 pushes assessment toward demonstrated competency rather than recall, and personalised routing only functions when you assess by competency. If your tests produce a single number, you cannot route; if they produce a per-competency profile, routing is almost automatic.

So the policy direction and the practical mechanism reinforce each other. Writing outcomes as observable competencies — calculate, derive, distinguish, construct rather than understand or know — gives you both alignment and the tagging structure that personalisation depends on.

The same structure, with one addition: language is itself a branch condition. A student may understand the concept and fail the assessment because the question was in English. Diagnosing which of those two failed is essential before routing, or you will remediate the wrong thing.

A cheap diagnostic: re-ask a failed question in the student's stronger language. If they answer correctly, the gap was linguistic and Foundation material on the concept will not help — they need the vocabulary, not the mathematics.

Check too that your students can read the interface itself in their language, not just your content. A learner navigating an English-only portal to reach Tamil material is carrying an unnecessary load that will show up in your data as a comprehension problem.

Measure the spread, not the average. Personalisation that works narrows the gap between your strongest and weakest students while the mean improves. If only the average moves, you have probably helped the students who were already fine and missed the ones the system was for.

Track two numbers per term: the proportion of students who move up a tier, and completion among students who started in Foundation. Those two capture whether the branches are functioning as ladders or as sorting boxes. If nobody ever leaves Foundation, the material is too hard or the prerequisite mapping is wrong.

Tag your existing assessment questions by topic. Nothing else in this article is possible without it, and it is a single afternoon's work for a term's worth of quizzes. Every subsequent step — diagnosis, routing, AI-generated tiers — depends on having topic-level rather than test-level data.

Then pick one topic your students reliably struggle with, build the four tiers for that topic alone, and run it. One topic teaches you more about your thresholds and your students than a term of planning will.

About ninety minutes, concentrated in one sitting. Fifteen minutes reading the assessment data, twenty deciding placements, thirty generating or selecting the tier material, and the rest assigning it. The cost is roughly flat whether the batch is fifteen students or fifty.

1. Monday — pull last week's topic-level results and flag anyone who moved more than one band.
2. Monday — place students into the four tiers. Override the rule wherever you know something the data does not.
3. Tuesday — generate or pull the tier material for the week's topic. One prompt produces all four versions.
4. Tuesday — assign by group, not by individual message. If this step needs fifty messages, the system will not survive the month.
5. Friday — glance at who did not open anything. That list matters more than the scores.

That last step is the one most tutors omit and the one that catches problems earliest. A student quietly not engaging is a stronger warning than a poor score, and it is invisible unless you look for it deliberately.

Four recurring failures: routing on aggregate scores instead of topic-level data, treating tiers as permanent labels, building more branches than you can maintain, and never telling students why they were placed. Each is a design error rather than a technology limitation.

• Aggregate routing — a student scoring sixty could have three entirely different gaps. Without topic tags you are guessing.
• Permanent labels — a tier is a description of this week, not of the student. If nobody ever moves up, students correctly conclude it is a sorting box.
• Too many branches — every additional tier multiplies material to maintain. Four survives a term; nine does not.
• Silent placement — a student moved to Foundation without explanation assumes they have been written off.

The last one causes the most damage and costs nothing to fix. One sentence — “you are on the extra-practice track this week because of simultaneous equations; you will move back once that is solid” — converts a demotion into a plan. Indian students in particular read unexplained streaming as a permanent judgement, and so do their parents.