Cold-Start Motivation: How AI Should Engage a User Before It Knows Them
Cold-start motivation is the problem of helping a new user experience value before the AI has enough history to personalize confidently.
Cold start is a motivational problem as much as a recommendation problem
Recommender-system research has long examined how to learn useful preferences from new users who have little or no interaction history.[1] AI products face the same data scarcity, but they also face a motivational version of the problem: the user has not yet experienced enough value to justify giving the system more time, data, or trust. Asking twenty preference questions can improve the model while destroying onboarding. Asking nothing can produce irrelevant output that makes personalization look weak. The design challenge is to collect just enough evidence to create an early success, then let that success earn permission for deeper adaptation.
Treat the first model as low confidence
New-user signals are provisional. The system should expect to be wrong and update quickly as real preferences emerge.
The first interaction should set expectations before it tries to impress
Microsoft’s human-AI guidelines recommend making clear what a system can do and how well it can do it during initial interaction.[2] That is foundational for motivation because uncertainty increases perceived effort. A new user should quickly understand what the product is for, what input is needed, and what a useful first outcome looks like. Instead of presenting a blank chat box, the system can offer a small set of goal-oriented starting paths. Instead of promising magical personalization, it can explain that recommendations will improve with feedback. Clear expectations turn the absence of history from a hidden weakness into a normal part of the relationship.
A useful first win is better than a complete profile
The fastest route to trust is usually a meaningful outcome, not a long setup flow that promises future personalization.
Early personalization should support autonomy, competence, and relevance
Self-determination theory highlights autonomy and competence as important conditions for high-quality motivation.[3] A cold-start flow can support both without pretending to know the user. Offer meaningful choices rather than a personality quiz, and choose questions that immediately change the experience. Asking “Do you want a fast answer, a guided walkthrough, or a challenge?” gives the user agency and gives the system actionable information. The first recommendation should then produce a visible benefit—a useful plan, solved setup problem, or small accomplishment—so the user experiences competence before being asked for more profile information.
Ask questions with visible consequences
Users are more willing to answer when they can immediately see how the answer changes recommendations, tone, pacing, or task structure.
A prompt works only when the behavior is feasible in the moment
The Fogg Behavior Model describes behavior as occurring when motivation, ability, and a prompt converge.[4] Cold-start onboarding often fails on ability: the product asks users to define goals, import data, connect integrations, configure preferences, and learn a new interface before any payoff appears. AI can reverse that sequence by lowering initial effort. It can infer defaults from the current task, ask one question at a time, explain examples, and postpone optional setup. The design target is not minimal clicks for its own sake; it is making the first valuable behavior easy enough to perform before motivation decays.
Delay optional complexity
Advanced controls are valuable once the user has a mental model. Before that, they compete with the core reason the user arrived.
Too much choice can make a personalized product feel less personal
Iyengar and Lepper’s well-known choice research found that extensive assortments can attract attention yet reduce subsequent choice and satisfaction in some contexts.[5] A new AI user can experience the same overload when presented with dozens of personas, modes, templates, models, and settings. Cold-start design should compress possibility into a few understandable paths while keeping escape hatches for advanced users. Progressive disclosure lets the system learn from actual behavior before exposing every control. The paradox is that a product with powerful personalization often needs a deliberately simple first screen.
Use active learning questions whose answers change the experience immediately
The best cold-start questions have high information value and high user relevance. “What is your industry?” may help segmentation but feels bureaucratic if nothing visible changes. “Should I optimize this for speed, depth, or learning?” both reveals intent and shapes the next response. Teams should score onboarding questions by how much uncertainty they remove, how sensitive the information is, and whether the answer creates immediate user-visible value. This prevents the profile from becoming a data-collection project disguised as personalization.
Let behavior replace questionnaires as quickly as possible
After a few meaningful actions, observed preferences become more informative than speculative setup choices. The product can learn which explanations are expanded, which suggestions are dismissed, how much help is requested, and what goals are repeatedly chosen. But behavioral learning should remain cautious. New-user behavior is noisy because people are still exploring the interface. Early signals should therefore have lower confidence and shorter persistence than stable preferences confirmed across multiple sessions.
Cold-start success means earning the next interaction
The first session does not need to perfectly model the person. It needs to deliver enough value, clarity, and control that the user willingly returns or continues. That changes the optimization target. Instead of maximizing how much profile data is collected up front, optimize the sequence of evidence and value: explain the capability, ask one consequential question, produce a useful result, invite correction, then learn. Over several cycles, personalization becomes deeper because the product earned the data rather than demanding it before proving usefulness.
Works Cited
- 01
- 02Microsoft Research — Guidelines for Human-AI Interaction microsoft.com
- 03
- 04
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.