Personalisation for education and SaaS onboarding

Personalised onboarding applies the same machinery as a product recommender to a different question: not "which item?" but "which step next?" A learner who has mastered fractions should not sit through the fractions module; a new SaaS admin who has already connected their CRM should not be nagged to connect it. The fixed onboarding path most products ship is written for an imaginary average user, and every real user wastes time on parts of it. This article explains how a next-step model works for education and for SaaS, what signals it needs, the constraints that keep it honest, and how to prove it raised completion or activation rather than just moved people around.

Why one path fails for everyone

Onboarding designers know their users arrive with different goals, skills and contexts, and they usually handle that with a branching questionnaire on day one. That helps, but it locks the path at the moment the product knows least. Behaviour on day two is far more informative than answers on day one. A user who skips a step, repeats a step, or completes a task another way is telling the product what they need. Personalisation reads those signals continuously and re-sequences what remains, which is why it outperforms a static branch.

Education and SaaS: the same model, different constraints

How a next-step model works

Represent the journey as a graph

Every lesson or onboarding step becomes a node with prerequisites as edges. The model chooses among the steps whose prerequisites the user has met, which keeps the sequence sensible no matter what the model prefers. Content teams and product managers own the graph; the model only ranks within it. This separation is what lets non-technical teams add a lesson or a setup step without retraining anything.

Score candidates from behaviour

For each eligible step, the model estimates the probability that showing it now moves the user toward the goal: mastery of the next assessment in education, activation in SaaS. Inputs are the user's history of steps, outcomes and timing, plus static context such as role or course. Early on, with little history, the model leans on what worked for similar users, which is the same cold-start blending a product recommender uses. As history accumulates, the user's own signals dominate.

Decide when to intervene at all

A user who is progressing does not need a prompt. Interrupting them costs attention and trust. Part of the model's job is to score "do nothing" as a candidate, and to time interventions for moments when a user has stalled: a return visit with no progress, repeated failure on an exercise, a setup step abandoned halfway. That timing logic is where a next-best-action design and onboarding meet.

Adaptive learning paths: what to watch for

Education has a trap that retail does not. A model that optimises short-term engagement will learn to give learners what they find easy, because easy content is completed and rated well. Mastery requires productive difficulty. The remedy is to define the objective as performance on later assessments, not immediate completion, and to constrain the model with a spacing rule so revisited material returns at increasing intervals. Learners should also see why a step was chosen; a short explanation such as "because you found the last two exercises difficult" keeps the system understandable to students and teachers. Where a school or university needs the system to run on its own infrastructure and keep learner data in-country, the education industry page covers how we deploy.

SaaS onboarding personalisation: activation first

In software the objective is the activation moment: the first report generated, the first ticket resolved, the first invoice sent. Everything before it is friction and everything after it is retention. A personalised sequence orders setup steps by which ones most shorten the path to that moment for a user of this role on this plan, and it delays feature introductions and upgrade prompts until after activation. The strongest signal for what to show next is often what a user tried and failed to do, which is why onboarding personalisation works best when it lives inside the product, as an in-app copilot does, rather than only in email. The copilot we built for a field-service SaaS used the same in-product surface to guide users through setup and to answer questions in context.

User journey AI beyond the first week

The same graph-and-scorer design continues after onboarding. In education it becomes the revision scheduler and the recommendation of the next course. In SaaS it becomes the feature-adoption programme, the tip shown when a user starts a task the long way, and the moment to suggest an upgrade because the user has hit a limit rather than because the calendar says so. Treating onboarding as the first phase of a continuous journey model avoids building two systems that disagree.

Measuring completion and activation honestly

The metric is completion or activation, measured against a randomised control that receives the static path. Sequencing changes can shift when users do things without changing whether they do them, so a lift in step completion that does not show up in activation is not a result. Run the test for long enough to cover the typical time to activation, and split the result by role, plan and cohort. The method is the one set out in why you must run a controlled test; the one addition is that in education the assessment several weeks later is the outcome, and the test must run that long. Research on onboarding UX from the Nielsen Norman Group is a useful complement on what to show and when.

A worked example

A B2B SaaS product had an eight-step onboarding checklist and an activation rate that had not moved in a year. Analysis of event data showed that users in one role were abandoning at a step that was irrelevant to them, and that users who connected an integration first activated much faster than those who followed the checklist in order. We modelled the checklist as a graph with dependencies, trained a next-step scorer on the history of which steps preceded activation for which roles, and served it inside the product's existing copilot panel, with the static checklist kept for a randomised control. The personalised path put the integration first for the roles that benefited, skipped the irrelevant step, and stayed silent for users who were progressing. Activation moved in the personalised arm and held over subsequent cohorts. The product team then extended the same model to feature adoption after activation. The build followed the pattern in our personalisation engines service.

Team and timeline

A personalised onboarding build needs a product manager or learning designer who owns the journey graph, an ML engineer for the scorer and blending, a data engineer for the event pipeline, and a front-end engineer to surface the next step in the product. On a product with clean event data, a first version takes six to eight weeks including the control. It is delivered under personalisation engines, from $21,000 / ₹13.6L; where it lives inside a product assistant it is combined with SaaS copilots, from $19,500 / ₹12.8L. Care Plans on the pricing page cover retraining as the product and the courses change.

Before you start: a checklist

• Define the goal event: assessed mastery for education, the activation moment for SaaS
• Map every lesson or onboarding step as a graph with prerequisites and dependencies
• Confirm step completion, outcomes and timing are tracked as events under one user ID
• Decide which contexts (role, plan, course, cohort) the model may use
• Set the rule for when the model stays silent
• Keep the static path as a randomised control for at least one full activation cycle
• Agree how the product or content team adds steps without a retrain
• Plan the explanation shown to the user for why a step was chosen

Questions clients ask

• Is this different from a branching questionnaire? Yes; a questionnaire fixes the path on day one, a next-step model re-sequences continuously from behaviour.
• Will it hide hard lessons from students? Not if the objective is later assessment performance and spacing rules are enforced; that is a design decision, not a model property.
• Can we use our existing onboarding tool? Usually; the model decides the next step and the tool displays it, in the same way messaging tools receive picks.
• How much history do we need? Enough completed journeys to learn which sequences precede the goal; similar-user blending covers the first weeks.
• What if activation does not move? Then the sequence was not the bottleneck; the control test tells you that early and cheaply.

Related reading

See why AI copilots inside SaaS beat standalone chatbots for the in-product surface and event pipelines for the tracking foundation. Prices for every service are on the pricing page.

Sequence for the person in front of you, stay quiet when they are progressing, and measure the goal rather than the steps; that is onboarding personalisation that earns its place.