Interactive Interview · Predicting the Why
← Home
Welcome
Welcome · ~15–20 minute conversation

Thanks for joining the interview.

This page walks you through the 15 questions you'll be asked, with a short brief and useful context for each. Use it to read ahead, take notes as we talk, or both — your notes save automatically to this browser only.

The project

Predicting the Why

An EMBA Applied Research Project at SP Jain. The paper proposes a framework called ABSD for how AI systems can understand why users behave the way they do — not just what they do.

Why you

Practitioner perspective

The framework is grounded in research but needs people who actually build AI products to pressure-test it. Your domain knowledge and instincts about user behaviour are what we want to surface.

Reference card

The ABSD framework, in one minute

01
Action
What the user does. Clicks, purchases, navigation. The surface layer — what AI systems already capture.
02
Behaviour
How the user does it. Patterns across actions — hesitation, retries, navigation strategies.
03
State
What the user feels. Confusion, confidence, frustration, flow — inferred from behavioural patterns.
04
Drive
Why the user does it. Underlying psychological needs — competence, autonomy, relatedness.

What to expect

Length
15–20 min
Format
Semi-structured
Sections
5 · 15 questions

No question has a right answer. We're interested in how you actually think about user behaviour in your work — patterns you've noticed, things that surprise you, intuitions you couldn't fully defend. Skip anything that doesn't apply.

SEC 1Opening & Context

Could you briefly describe your current role and the AI system or product you work on?

What we're asking

Just an orientation. We want to know your role, what kind of AI system or product you work on, and roughly the kind of users it serves.

Helpful direction
  • Don't overthink scope — a sentence or two is fine.
  • If you wear multiple hats, mention them.
  • If you can name the product you'll later reference, do.
Your notes (private to this browser)unsaved
SEC 1Opening & Context

At a high level, how does your system currently track or respond to user behaviour?

What we're asking

This is the baseline. What kind of data does the system actually collect, and what does it do in response? Stay descriptive — no need to defend the design.

Useful frames from the paper

Most systems today operate at the Action layer — they log events and respond to them. The paper argues this leaves three deeper layers (Behaviour, State, Drive) largely unaddressed.

action layerevent trackingbehavioural patterns
Your notesunsaved
SEC 2Signal Recognition— RQ1

What behavioural signals do you find most informative for understanding the user's experience?

What we're asking

We're trying to surface the signals you have learned to read — whether or not they're tracked formally. The thing you notice in user-session replays. The pattern that makes you say "something's wrong here."

For context — signals in the paper
  • Hesitation (long pauses, typing-then-deleting)
  • Thrashing (rapid switching between options)
  • Regression (going back to completed content)
  • Acceleration (skimming through quickly)
  • Persistence (continuing after failure)
  • Avoidance (skipping, early exit)
  • Flow (steady, low-error progress)

You may have others not on this list — that's exactly what we want.

Your notesunsaved
SEC 2Signal Recognition— RQ1

Which signals do you currently track? Are there signals in your data your team does not use?

Especially: hesitation, navigation switching, re-engagement after drop-off, help-seeking, or changes in interaction pace.

What we're asking

Two parts: what you actively track today, and what's sitting in your event logs that nobody has looked at yet. The gap between "available" and "used" is often where the most interesting questions live.

Why we're asking

Forrester (2022) found 73% of companies struggle to use behavioural data for anything beyond basic segmentation. We want to test whether that matches what you see, and what's blocking deeper use.

Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

When you see a user hesitate or pause, what do you infer? How confident are you?

What we're asking

The same pause can mean different things — confusion, deliberation, distraction, cognitive load. We want to know how you read it in your specific domain, and how much you'd bet on that reading.

A worked example from the paper

A 30-second pause before a complex financial trade is prudent deliberation. The same pause before a simple checkout is probably confusion.

Same signal, different state — because context changes everything.

Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

Looking at the four layers — Action, Behaviour, State, Drive — does this map to how you think about user understanding?

What we're asking

Genuinely — does this structure feel natural, or does it impose distinctions that aren't useful in practice? We're looking for friction as much as agreement.

The structure, in one breath

Action generates trace data. Behaviour is the pattern in that trace. State is what the user is feeling. Drive is the deeper need shaping how they react to challenge. Each layer is generated by the one above (from the user's side) and must be inferred from the one below (from the AI's side).

04
Drive
Autonomy, competence, relatedness. Stable.
03
State
Confusion, confidence, flow, frustration. Momentary.
02
Behaviour
Hesitation, thrashing, regression. Pattern.
01
Action
Clicks, submissions, navigation. Observable.
Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

Give us an example where a pattern clearly indicated a state. And one where the same signal was ambiguous.

What we're asking

Two short stories. The clean read — and the time you got it wrong, or couldn't tell. The ambiguity is often more revealing than the certainty.

Useful framing

In the paper this is called context-dependent signal interpretation. Narayanan & Georgiou (2013) described behavioural signals as having "individual and contextual heterogeneity" — they refuse to be deterministic.

Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

What contextual factors most affect how you interpret signals?

What we're asking

What changes the meaning of the same behavioural signal? Domain? User history? Time of day? Task complexity? The thing that, when you check it, reframes everything.

Three categories in the paper
  • Domain — task type, industry, interaction norms.
  • User — individual baseline, history, expertise.
  • Temporal — time of day, journey stage, recency of system changes.

If you'd add a fourth category, that's exactly the kind of input we need.

Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

When your system assumes a user is frustrated or confused — how do you validate whether that was correct?

What we're asking

What's the feedback loop between inference and ground truth in your system? Surveys? Support tickets? Conversion outcomes? Just a hunch?

Why we're asking

The paper argues that state-aware systems risk operating in a closed inferential loop — unchecked assumptions compounding over time. Whatever you do to break that loop is what we want to learn from.

Your notesunsaved
SEC 3Signal-to-State Inference— RQ2

Does your team design for the user's underlying motivations — competence, autonomy, relatedness?

What we're asking

This is the deepest layer — Drive. Does your product try to support these needs, even implicitly? Or does the conversation never get there?

The three needs from SDT

Autonomy — feeling in control of your choices. Competence — feeling effective at what you're doing. Relatedness — feeling connected to others.

Forty years of research show these predict engagement and wellbeing across domains.

autonomycompetencerelatedness
Your notesunsaved
SEC 4Framework Utility— RQ3

If you had reliable state-level inference — what would change in how your system responds?

What we're asking

The hypothetical world where the system actually knows the user is confused, or frustrated, or in flow — what would it do differently?

Possible directions
  • Trigger autonomous agent intervention
  • Adapt UI complexity in real time
  • Alter tone or pacing of content
  • Route to a human agent
  • Something else entirely
Your notesunsaved
SEC 4Framework Utility— RQ3

What are the biggest practical barriers to implementing state-aware features?

What we're asking

The honest reasons it hasn't happened. Technical? Data? Organisational? Ethical? Just no-one's prioritised it?

Common candidates
  • Engineering complexity
  • Data quality / availability
  • Org priorities & KPIs
  • Privacy & ethics
  • No framework to guide it
Your notesunsaved
SEC 4Framework Utility— RQ3

Does the ABSD framework give you vocabulary you'd actually use in product or design discussions?

What we're asking

This is the utility test. The framework isn't trying to be novel — it's trying to be useful. If the layer-language gives you something to point at in meetings, it earns its place. If not, it's just clever.

Test it

Imagine your next product review. Someone says "users are dropping off after onboarding." With ABSD vocabulary, would you reframe the conversation differently?

Your notesunsaved
SEC 4Framework Utility— RQ3

What's missing from the framework? What would you add, remove, or restructure?

What we're asking

This is the most valuable question in the whole interview. The framework will only be useful if practitioners shape it. What does it not yet cover that it should? Where does it overclaim? What feels missing?

Permissions granted
  • It's fine to say "I'd collapse layers 2 and 3."
  • It's fine to say "Drive is too abstract to be useful."
  • It's fine to say "There should be a fifth layer for X."
  • It's fine to say "the words don't work for me."
Your notesunsaved
SEC 5Closing

In your domain, where would understanding user motivation have the highest impact?

What we're asking

If a system in your space genuinely understood why users were doing what they're doing — where would that change the most? One use case, sharply drawn.

Why we end here

This question turns the whole interview forward. We're trying to find the highest-leverage applications of the framework, and you've just spent 15 minutes thinking about where the gaps are.

Your notesunsaved
Thank you

That's the conversation. Genuinely — thank you.

The framework only sharpens when practitioners push on it. Your notes are saved in this browser. If you'd like to share them with the researcher, export below.

Notes never leave your browser unless you choose to share them.