Brief · Start here 5–7 min read No jargon Open research

AI systems see everything.
They understand almost nothing.

Netflix processes 500 billion events per day. Most of that data answers one question: what did users do? This research builds a framework for the harder question — why.

YouTube detects that you watched three political videos and floods your feed for weeks. It read your behaviour correctly. It misread your state entirely — situational curiosity became a permanent profile update. That distinction is what this framework addresses.

AC
Arunamirtharaj Chandrasekar
SP Jain EMBA · Applied Research Project

73% of companies cannot use their behavioural data for anything beyond basic segmentation.

That number comes from Forrester. McKinsey found that companies excelling at personalisation generate 40% more revenue than the average. The distance between those two numbers is a gap most teams have never formally named.

The problem is not a lack of data. Netflix has too much data. The problem is a lack of interpretation. When a student pauses for 30 seconds before answering a question, the system records the pause and moves on. It cannot tell confusion from careful thinking. When a customer abandons a product flow at step three, the system logs the drop-off. It cannot tell frustration from distraction from simply running out of time.

AI systems are optimised to predict what users will do next. They are not built to understand why users are doing it in the first place.
The core claim of this research

This matters because what and why require different system architectures. Predicting the next click requires historical data and a good model. Understanding motivation requires a different kind of observation - one that most deployed AI systems have never attempted.

Four layers. Each one sits beneath the last.

The ABSD framework maps the inferential path from what users do to why they do it. Each layer is more useful than the one below it - and harder to reach.

L1
Action
What users do. Clicks, submissions, timestamps, navigation. This is what most AI systems capture today - at high fidelity and massive scale.
Netflix logs 500 billion events per day at this layer. The data is there. The interpretation is not.
L2
Behaviour
How users do it. Patterns across actions over time. Hesitation before a decision. Rapid switching without completing tasks. Returning to earlier content. These are not events - they are sequences. They require temporal aggregation.
A user who clicks help three times in five minutes after failing the same task is showing a pattern. The event log shows three clicks. The behaviour shows something else.
L3
State
What users feel. Confusion, confidence, frustration, flow. Not directly observable. Inferred from behavioural patterns using contextual weighting. Affective computing research achieves 70–79% accuracy in controlled settings.
The same hesitation pattern means different things in different contexts. Before a complex financial decision it may indicate prudence. Before a simple form field it indicates confusion. Context determines which.
L4
Drive
Why users do it. The underlying psychological needs - competence, autonomy, relatedness - that shape how people respond when their states are challenged. Stable over time. Shaping every response to difficulty.
Two users both reach a frustration state. One persists. One disengages. The difference is not the state - it is the drive beneath it. Same signal, different person, different system response needed.

The key insight is directional. From a generative perspective, Drive shapes State, State produces Behaviour, Behaviour manifests as Action. The user's motivation generates their behaviour. AI systems must work in reverse - inferring upward from the observable layer to the motivational one.

Seven patterns that bridge the layers. All of them are probably already in your data.

These signals appear across JITAI research, affective computing, and learning analytics. They are not invented for this framework. The contribution is naming them clearly and mapping them to the states they indicate.

Hesitation
Extended pause before action. Indicates uncertainty or low confidence. The most common signal - and the most frequently ignored.
Thrashing
Rapid switching without completing. Indicates overwhelm or decision paralysis. Often misread as high engagement.
Regression
Returning to earlier content. Indicates confusion or the need for consolidation. A positive signal - the user is self-correcting.
Acceleration
Speeding through content. Could be boredom, overconfidence, or disengagement. Context determines which - and which intervention is right.
Persistence
Continuing after repeated failure. Strong competence drive. The signal most systems misread - interrupting someone who is in flow is worse than not intervening at all.
Avoidance
Skipping, minimal engagement, early exit. Often anxiety, not disinterest. Consistent avoidance over time is the clearest churn predictor in the taxonomy.
Flow
Steady progress, appropriate pace, low errors. The optimal state. The correct system response to flow is no intervention - preserve the conditions, not disrupt them.

Netflix and Khan Academy are best in class. Both stop at Layer 1.

Two case studies examined through the ABSD lens using publicly available documentation. The finding is consistent: class-leading AI systems are excellent at capturing actions and partially capable at detecting patterns. Almost none attempt state inference. None model drive.

Netflix processes 500 billion events per day and uses deep neural networks to rank content recommendations. The system logs when you pause, when you rewind, when you abandon a film at the 40-minute mark. It cannot tell whether the pause was anticipation or confusion. It cannot tell whether the abandonment was dissatisfaction or distraction. Both signals get filed as consumption data.

Khan Academy tracks exercise attempts, hint usage, and mastery thresholds per skill. Khan Academy is a genuine Layer 2 system - behavioural patterns feed the mastery model. But the 45-second hesitation before a question a student has answered correctly before carries no different weight than a 4-second answer. The cognitive process is invisible. VanLehn's 2011 research showed that AI tutors underperform human tutors precisely because they lack this psychological understanding. That gap remains accurate today.

The signals that would enable Layer 3 inference are present in the data of both systems. The framework to interpret those signals is not.
Cases analysis - this research

What this is not.

×
Not a surveillance proposal. ABSD does not advocate for tracking users without their knowledge. The framework's application requires transparency about what signals are collected and how they are interpreted. The ethics page documents this in detail.
×
Not a claim that AI can read minds. The confidence gradient in the framework is explicit. Action is observed at 96% fidelity. Drive inference is longitudinal and probabilistic. The framework is honest about where inference stops being reliable.
×
Not a new technology proposal. ABSD is a conceptual framework. It organises existing research from four traditions - SDT, affective computing, JITAI, and learning analytics - that had not previously been connected. It is a map, not an engine.
×
Not domain-specific. The framework is designed to apply across CX, EdTech, healthcare technology, and any domain where AI systems mediate human decision-making. The signals and layers are the same. The contextual modifiers change.

Open. Practitioner-led. Live now.

This research is being conducted in the open. The survey is live. Expert interviews are in progress. Findings update daily.

53 references across six source clusters - Self-Determination Theory, affective computing, JITAI research, learning analytics, CX and industry reports, AI and recommendation systems. The literature synthesis found minimal cross-citation between the four primary traditions. ABSD is the first framework to connect all four.

The research question is empirical, not theoretical: do practitioners recognise the signal gap described by ABSD in their own systems? If they do, does the framework provide vocabulary and structure useful enough to act on?

The survey takes five minutes. The voice interview takes twenty. Both are open to any practitioner who builds AI products in CX, EdTech, or adjacent domains.

Does this resonate with your experience?
The survey captures exactly this question - and your behavioural patterns as you answer are part of the research.