Shaping an AI Financial Coach Through Behavioral Research
Generative research on impulse spending that defined the product strategy, feature architecture, and AI capabilities for a new financial coaching platform.
Project Overview
- Role
- UX Researcher · Team of 4
- Context
- Stanford · Design for Behavior Change
- Methods
- Literature Review · Competitive Analysis · Screener Design · Diary Studies · Interviews · Thematic Analysis · Behavioral Modeling · Assumption Testing · AI Intervention Study · Usability Testing
My Research Ownership
Study Design & Execution
Authored the screener (N = 39), baseline diary protocols (N = 9), intervention methodology (N = 7), and pre/post interview guides.
Synthesis & Modeling
Led thematic analysis and authored the baseline persona and behavioral journey map.
Assumption & Intervention Testing
Independently ran primary Assumption Test (N = 14) and co-designed the AI field intervention study.
Product Translation & Evaluation
Built story maps and dashboard wireframes; designed and ran the formative usability evaluation (N = 2).
In Summary,
Traditional budgeting tools act reactively—logging expenses after the money is already spent. Sage began instead as a generative question: what actually drives an impulse purchase, and could an AI coach meet people during an unplanned purchase? This work was foundational discovery research meant to shape a new product from the ground up based on a new approach to financial wellness that tackled unplanned purchases rather than targeting impulses.
The research redefined Sage as a tool built to help users recognize patterns, set intentions, and decide for themselves what they wanted to change rather than one that simply tries to "stop impulse purchases".
01 — Framing the Behavior
What actually drives an impulse purchase—and where could an intervention realistically fit?
Before designing anything, we needed to understand what research already said about impulse spending and where existing tools intervened. My literature review on FOMO and the team’s broader review converged on one pattern: spending is driven not only by financial knowledge, but by environmental cues, emotional states, social pressure, and the friction built into purchasing environments.
What existing tools assume
I assembled the final 2×2 positioning map — researching HoneyDue, Penny, and manual money-management methods — along two axes: timing (reactive ↔ preventative) and effort (passive ↔ active). The gap: most tools either explain spending after it happens or demand sustained manual budgeting, leaving an opening for a lower-friction intervention operating closer to the decision itself.
02 — Establishing a Behavioral Baseline
Revelation 1
Impulse spending is heavily contextualized.
Study Design
screener responses
participants selected
pre-study interview
in-context purchase logging
post-study interview
Why a diary study
Convenience purchases happen so often that low-friction, in-the-moment logging captures the behavior without swaying it — a truer baseline than retrospective recall.
Each entry captured
Item, cost range, planned vs. unplanned, impulsivity, guilt, context, reasoning, and emotion. I authored the screener, outline, comms, guide, and both interview scripts.
Synthesis Process
Key Baseline Findings
★ Key Conceptual Insight
Finding 01: Whether spending felt problematic depended more on identity and values than on impulsivity alone.
Purchases tied to family, health, social connection, convenience, or self-improvement were often judged more leniently—even when they were spontaneous or expensive. This proved that “unplanned” did not automatically mean “undesirable,” highlighting the key difference from self-identifying "impulse" spenders and people who frequently made unplanned, and, retroactively, unwanted purchases.
Finding 02: People rationalized unplanned purchases once context made them feel reasonable.
Spontaneous purchases were reframed as necessary, productive, or goal-aligned. Hunger, time pressure, social situations, discounts, and “while I’m here” logic appeared repeatedly across all.
Finding 03: Logging made invisible spending visible — but awareness cut both ways.
Some participants felt they grew more deliberate seeing purchases accumulate; others felt validated that their spending still matched their self-image.
Modeling the Impulse Cycle
I translated the synthesis into a behavioral journey map locating where an intervention could interrupt automatic behavior without assuming every unplanned purchase should be prevented.
The journey map combines what participants said, thought, did, and felt at each stage. Click to enlarge.
03 — Challenging Our Biggest Assumption
Revelation 2
Not everyone wanted the behavior eliminated.
That assumption underpinned the entire initial product. If people did not recognize their impulses or perceive impulse spending as something worth “fixing,” a restriction-oriented intervention would be solving the wrong problem.
I selected this as a high-risk assumption, designed and ran a rapid test, analyzed the responses, and documented the implications for the product.
8 of 14
expressed at least some desire to stop or reduce impulse spending — and only 5 of those gave a strong “yes.”
Rapid directional test · n = 14 · convenience sample
Interpretation & Impact
The result was not evidence that most impulse spenders do not care about changing. The sample was small and convenient. But it was enough to challenge a foundational assumption: we could not responsibly design Sage around the idea that impulse spending was universally experienced as a problem by those who spent impulsively.
From fixing behavior → supporting agency: I recommended reframing Sage away from a product that “fixes” impulse spenders and toward a financial-wellness guide that could support users whether they wanted to reduce, understand, or simply become more intentional about their planned and unplanned spending.
04 — Intervention Testing
Revelation 3
Awareness created short-term changes in feelings towards spending, not short-term behavior.
Setup
- 7 participants
- 5-day field intervention
- Pre/post interviews + transcripts
The daily loop
Each morning, participants told a ChatGPT-based agent what they planned to buy, adding plans through the day. Each evening, it prompted reflection on unplanned purchases — context, emotions, reasoning.
AI stance
Framed as a reflective partner, not an enforcer: built to surface patterns and ask “why,” never to shame. I built and designed the agent's persona & designed the study's goal, methodology, selection criteria, guide, and comms.
Intervention Findings
01 · Awareness changed faster than behavior.
Nearly all participants became more conscious of their spending patterns, but most kept making similar purchases over the five days.
Implication: Awareness became an intermediate outcome — not proof of completed behavior change.
02 · Morning planning mattered more than evening reflection.
Declaring planned purchases introduced friction before spending. Several paused mid-day because they’d already stated their intentions that morning.
Implication: Intention-setting became a core interaction, not a setup step.
03 · Social spending is weighed differently.
Unplanned purchases in social settings produced little regret because they carried relational value. A purely individual “control your spending” model would miss that meaning.
04 · Reflection needs to know when to stop.
A non-judgmental tone was appreciated, but excessive follow-up felt repetitive or manipulative. Reflection depth should adapt to the user and the purchase.
05 · Emotional logics
Unplanned purchases clustered around mood regulation, social compliance, FOMO, and reward logic. The AI agent was most useful when it surfaced a pattern the participant hadn’t consciously articulated.
05 — Research → Product Decisions
Rather than translating findings directly into isolated features, I used the research to define several principles for how Sage should behave. Each one traces a straight line from evidence, to principle, to the product response it informed.
Evidence
Not all participants saw impulse spending as inherently undesirable.
Principle 01
Preserve Agency, Quality over QuantityProduct Response
Sage helps users reduce unplanned spending — it does not automatically block/punish purchases or set budget constraints.
Evidence
Morning planning introduced useful friction before some purchases occurred.
Principle 02
Intervene Before AutopilotProduct Response
Daily intention-setting becomes central to the experience.
Evidence
Some users valued deep probing; others found persistent AI questioning exhausting.
Principle 03
Calibrate ReflectionProduct Response
AI coaching adjusts depth and tone rather than applying one reflection pattern to every purchase.
Evidence
Awareness rose more consistently across the study than purchase reduction did.
Principle 04
Treat Awareness as ProgressProduct Response
Success is measured by awareness and intention, not by short-term spending cuts alone.
Story Mapping & Wireframes
I built the story map for the central persona, the Guilt-Driven Impulse Buyer, to feel the system from the user’s side across planning, spending, reflection, and long-term awareness — connecting behavioral findings to the MVP rather than jumping from themes straight to screens.
Visual direction: to counter the guilt and avoidance the research surfaced, I helped write the rationale for a non-punitive tone with soft boundaries that avoid triggering balance-checking anxiety.
06 — Evaluating Sage
Evaluating Sage
Study
Formative medium-fidelity usability study · n = 2. I designed the goal, script, participant intro, task structure, and reflection.
Tasks tested
- Onboarding + bank integration
- Logging a planned purchase
- Exploring a spending insight
- Reviewing spending with the AI + limits
What it measured
Whether users grasped the plan → spend → reflect model, and whether the integrated AI agent felt like a meaningful part of it.
The redesign the evaluation forced
The clearest change was structural — where the AI coach lived in the product. That before/after shift and the three research-driven changes it drove:
Before
The dashboard treated the AI as one destination among several. Logging a planned purchase quietly redirected users into a chat they hadn’t asked for — and some read that chat as customer support and avoided it.
After
The AI coach becomes the primary landing experience, introduced during onboarding, and immediately prompts users to articulate the day’s planned purchases — making plan → spend → reflect the spine of the product.
Three research-driven changes
01 · The AI didn’t feel central
Testers were confused when logging redirected them into the AI; one read AI in finance as “customer service.” → Made the coach the primary landing experience, introduced early in onboarding.
02 · The model wasn’t legible
Users didn’t grasp plan → spend → reflect. → Onboarding now teaches the daily loop up front.
03 · No system status
Users were unsure where they were or whether purchases saved. → Added progress indicators, submission feedback, and state persistence.
07 — What We Know, and What We Don’t
What it established
Impulse spending is strongly contextual; people judge purchases by identity, values, and social meaning; and, reframed as planned vs. unplanned, reflective logging can raise awareness while intention-setting creates useful friction before some convenience purchases.
What it did not
Studies were small, self-reported, and five days long. Awareness rose more consistently than behavior changed — so this does not establish that Sage causes sustained reductions. The rapid tests used small convenience samples.
What I’d test next
A longer longitudinal field experiment with distinct baseline and intervention periods, wider financial circumstances, and behavioral measures beyond self-report — unplanned-purchase share, impulse-to-purchase delay, abandoned purchases, and whether reflection stays useful past novelty.
In Closing
Final Product Direction
Sage became a financial-wellness product centered on awareness, intention-setting, and reflective coaching — not spending restriction.
The five-day study didn’t prove long-term behavior change; it proved that awareness and intention are the mechanisms worth building — and testing — next.