The first minutes in a new SaaS product decide whether someone comes back. Classic onboarding tries to lead every user down the same path with tours and checklists. That fails because people arrive with very different goals: one wants to import an existing data set immediately, another first wants to see the product with sample data. AI can help take that difference seriously — if it is used for the right tasks.
Ask instead of guessing
The most effective personalisation starts with a simple question: what do you want to achieve with the product? An open answer in the user's own words is more comfortable than picking from ten predefined categories, and a language model can reliably map it to a small number of starting paths. That is more honest and more accurate than inferring a profile from behavioural data the user never confirmed.
The mapping should stay visible and correctable: «We have set up your workspace for team project planning — is that right?» Anyone who wants a different setup switches with one click.
Filling the empty state
An empty product is the most common reason for early drop-off. AI can generate a sensible starting state from the stated goal: a first project with typical tasks, a template for the described use case, example categories that fit the user's business. This content must be clearly marked as a suggestion and removable with one click; otherwise it feels like someone else's data in your workspace.
Making imports easy
For many B2B products, importing existing data is the real activation moment. Spreadsheets from other tools rarely have the expected structure. A language model can recognise and map columns, interpret date formats and flag inconsistencies. The user confirms the proposed mapping before anything is imported. This step often saves more time than any other AI feature in onboarding, because it takes a tedious task off the user's hands instead of merely explaining something.
Help where it is needed
Instead of a tour that presents ten features at first login, a help function can answer questions in context: «How do I invite my team?» or «Can tasks repeat?» It should be grounded in your own documentation through search, so answers are correct and link to the right help page. A help function that invents features that do not exist does more harm in onboarding than none at all.
Where personalisation backfires
Personalisation becomes a problem when it is opaque or restricts users. Features hidden based on an estimated profile mean people never find them. Onboarding that changes on every visit confuses. And personalisation based on data the user does not expect to be used — such as analysing content uploaded for other purposes — undermines trust and can be problematic under data protection law.
Measuring whether it works
Onboarding improvements are measurable: the share of users who reach a defined activation step, the time it takes, return visits in the first week. AI-assisted variants should be tested against the current version, not simply introduced. Often individual building blocks — such as the import assistant — have a large effect while others, such as automatically generated welcome texts, barely make a difference.
Data protection and transparency
The stated goal and imported data are sent to a language model. That belongs in the privacy policy, the model provider must be contractually bound, and only necessary data should be sent — for column mapping, column names and a few sample rows usually suffice. Under the EU AI Act, users must be able to recognise when they interact with an AI system unless it is obvious.
A sensible starting point
- One open question about the goal, mapped to a few starting paths, visible and correctable.
- Suggested starter content, clearly labelled and easy to remove.
- An import assistant that maps columns and asks for confirmation before importing.
- Contextual help based on your own documentation.
- Activation measured against the previous version.
Innopulse uses these onboarding building blocks in its own SaaS products and builds them for clients so that they work measurably rather than just looking modern.