Product Design · Human–AI Interaction
NETO: guidance without taking control.
Designing a banking companion that explains the next step, makes its limits visible and leaves every consequential action with the person.
I created all of NETO’s UX and design. I used Claude to program the experience, providing the requirements, directing iterations and reviewing the implementation against my design intent. The public demo uses rules; live model integration is unfinished.
Guided prototype · Fictional data · No installation required
- 3 taskssupported in the guided demo
- Human controlbank actions remain with the person
- Rules firstAI interpretation explored separately
- In progresslive model integration

Research · Product framing · Interaction design · Agent contract · Browser evaluation
- My role
- Full UX and design ownership; requirements, implementation direction and review
- Context
- Original training research · Independent AI exploration in 2026
- Build approach
- Claude-assisted coding, guided by my requirements and review
- Current stage
- Guided demo + separate agent prototype; live integration pending
01 · Framing the product
Help someone take the next step without taking over.
The original research surfaced fear of making a costly mistake, unfamiliar banking language and the importance of human support. I translated those findings into a product question: could contextual guidance make an unfamiliar task easier to understand while leaving every banking action with the person?
I used a simulated transfer as the main interaction to explore. It contains several points where the person needs to understand what happens next: choosing a recipient, entering an amount, checking an account and reviewing before confirming. Accounts and transaction activity provide simpler starting tasks.
02 · UX research & synthesis
Understanding hesitation before designing assistance.
I investigated why older adults hesitated to use banking apps, combining desk research, a survey and interviews. The work connected fear of financial mistakes with practical barriers: unfamiliar language, visual effort, unclear navigation and the need for human support.
Frame the unknowns
I used a certainties, assumptions and doubts matrix to shape the survey questions, alongside desk research and a banking accessibility benchmark.
Listen across methods
A 20-question survey gathered 73 responses, including 40 in the 55–80 target group. Interviews by phone and WhatsApp explored experiences, fears and support needs.
Translate findings into direction
Personas, empathy maps and journeys made different levels of confidence visible. The recommendations prioritized plain language, readable interfaces, confirmation and step-by-step guidance.


How the UX work shaped the 2026 prototype
Fear of a costly mistake
Keep review and confirmation separate.
The original recommendation for confirmation became an explicit interaction boundary: NETO explains the review screen, while the person confirms in the bank.
Unfamiliar language and paths
Give one contextual instruction.
Plain language and step-by-step guidance informed short instructions beside a visible target, with repeat and pause controls.
Visual effort and need for support
Provide presentation choices and clear limits.
Readable text and adjustable presentation address visual effort. Human support remains a research need; the prototype does not provide a live handoff.
The original study was a training project and ended in research and recommendations. The independent 2026 prototype builds on that foundation. Recruitment through personal contacts and online groups limits representativeness; the study did not validate this later interface.
Explore the full research and recommendations ↓03 · Defining the AI’s job
Use AI for interpretation. Keep the interaction boundaries explicit.
The product does not need a model to know the order of a fixed practice flow. The proposed value of AI is understanding how someone asks for help and explaining a step in a different way. The current demo uses rules so the interface can be tested independently of model quality.
| Approach | Useful for | Boundary in NETO |
|---|---|---|
| Deterministic guide | Known screens, allowed targets and progression through a task. | Runs the playable prototype; predictable but limited to supported tasks. |
| Language model | Interpreting a request and reformulating an explanation. | Agent prototype exists separately; its live connection to the interface is unfinished. |
| Autonomous actions | Completing a task for the person. | Excluded. NETO does not execute clicks or confirm a banking transaction. |
One step instead of a long answer
A single instruction limits the amount to interpret at once. The tradeoff is a slower interaction; pause, repeat and direct bank controls keep the person in charge of the pace.
Context beside the bank
An overlay keeps guidance close to the interface. It can also obscure the content it is meant to explain. Compact mode, repositioning and Hide address that tradeoff; mobile behavior still needs further testing.
Voice as another route
Browser speech can read instructions aloud, while text stays available. Voice availability varies, and microphone input is optional. Audio must be controllable rather than required to complete a task.
A clear boundary before confirmation
The guide explains the review screen and waits. The bank’s confirmation remains a separate action by the person; a conversational “yes” is not permission for NETO to send money.
04 · The interaction
A companion that explains, points and waits.
NETO identifies a supported control, explains the next step and observes what changes after the person acts. The assistant can be paused or hidden without taking the bank away.




Explore the guided prototype
Try accounts, transaction activity or a transfer using fictional data. Speech output is available where the browser supports it.
Demo status: the playable environment uses deterministic responses, not a live model. No real bank or payment service is connected. Browser speech and optional voice recognition are separate from the unfinished AI integration.
05 · Designing the agent contract
A model response should not become an unrestricted action.
The Claude-assisted implementation includes a separate TypeScript agent prototype using the OpenAI Agents SDK and a Zod output schema. Its instructions specify a patient tone, one action at a time, no invented controls and no unsupported assurance that an operation is safe.
- Limit the context
The browser adapter identifies known screens and available target names. The request format contains the question, task, step and status rather than a copy of the banking page or its field values.
- Interpret the request
The intended agent output classifies requests such as accounts, activity, send, repeat, pause or suspicious. If a control cannot be identified, the target can be null instead of a guessed instruction.
- Validate before rendering
The response contract contains a version, intent, message and optional target key. The client checks the format and available targets; the guide remains responsible for progression.
- Wait for evidence of progress
The interface observes bank changes after the person acts. Pause and cancellation invalidate pending work so a late response does not silently advance the task.
Example of the response contract
Illustrative response format used to constrain the agent experiment.
{
"version": 1,
"intent": "send",
"message": "We can open Send money. Nothing is sent yet.",
"targetKey": "send"
}Where implementation stops today: the AI agent runs as a separate command-line prototype. The extension expects a local endpoint that the agent does not yet implement. The guided web demo lets me evaluate interaction behavior without presenting that missing connection as complete.
06 · Trust & failure states
A reassuring tone is not a safety guarantee.
Human oversight must happen where an action becomes consequential. NETO can explain where to review; only the person can decide whether the recipient, amount and account are correct and use the bank’s confirmation control. An assistant saying “done” is not evidence that a transfer succeeded.
| Failure scenario | Required response | Current evidence / limit |
|---|---|---|
| Unknown or duplicate control | Do not point to a guessed target; explain that guidance is unavailable. | The adapter returns no target unless it finds one matching, visible, enabled control. Coverage is limited to the practice bank. |
| Delayed answer after Pause | Keep the task paused and ignore an obsolete response. | The client includes cancellation and timeout handling. No live-model end-to-end reliability result exists. |
| Password, code or account details | Avoid sending the message to a model; ask the person to leave out private details. | Keyword and digit checks exist. They can miss sensitive text or block harmless requests; this is not comprehensive data-loss prevention. |
| Suspicious link or payment pressure | Stop reassurance; direct the person to an independently verified bank channel. | A scripted warning and a simple keyword guard exist. No fraud-detection accuracy has been measured; no human support agent is connected. |
| Valid schema, misleading explanation | Reject unsupported advice and preserve the last known task state. | A schema checks structure, not truth. A semantic evaluation set and a tested safe fallback are still needed. |
| Microphone or speech unavailable | Keep written guidance and manual controls usable. | Voice is optional. Browser recognition may use online processing; permission and device behavior need explicit review. |
Minimize context, but test what escapes
The adapter avoids reading banking field values. The intended model request contains a short question and allowlisted screen state. That reduces exposure; it does not establish provider retention, consent or production privacy controls.
A companion must disclose its limits
The character and patient language may encourage trust. The interface must still identify itself as assistance, never impersonate a bank employee, and avoid implying that it has verified a transaction or can connect the person to live support.
07 · Design-led implementation with Claude
I owned the UX and design, and directed the implementation.
I created the UX and design for NETO, then used Claude to program the experience. I supplied the requirements, guided the work through iterative feedback and reviewed the resulting interface and behavior. My responsibility was to keep the implementation aligned with the intended experience.
Translate design intent into requirements
I specified how the experience should work: one instruction at a time, the bank kept visible, controls to pause or repeat, and consequential actions left with the person.
Direct the coding through feedback
I used Claude for programming and guided the iterations, checking its output against my requirements and giving direction when the implementation needed to change.
Review the working experience
I reviewed the interface and interactions in the browser. The review overlay is a concrete example: the initial implementation competed with the banking details, so the correction needed to preserve both readable guidance and the information being checked.
Keep acceptance and validation distinct
My implementation review checks whether the prototype reflects the design intent. It does not replace usability testing with older adults or establish the reliability of a live model.
The before-and-after examples below document implementation corrections. The public guided demo and the separate agent experiment remain distinct; their live integration is not complete.
08 · Finding and correcting failures
Browser testing changed the design.
Internal review exposed problems that were not apparent from reading the code. These examples document implementation and interface corrections; they are not usability-study results.



Large text revealed layout limits
At 26 px text, native selectors introduced horizontal scrolling inside the panel. Constraining their width removed the internal overflow in the checked narrow-screen state. This does not replace a full zoom or screen-reader test.
The demo lost a core modality
The first web adaptation disabled speech. Reviewing the preview exposed the regression. The browser speech module and visible Enable voice / Mute voice controls were restored. Playback-start was verified in the browser; voice availability and audible output still depend on the device.
09 · Outcomes & next decisions
A testable interaction, with explicit evidence gaps.
Implemented
A bounded guided prototype
A fictional practice bank, NETO’s interface, three supported tasks, adjustable presentation, pause and speech controls, plus a separate structured AI agent prototype.
Internally checked
Core interaction behavior
Accounts and activity navigation; a transfer through review; pause, resume and Escape; narrow-screen layout and enlarged text. NETO did not confirm the banking action during these checks.
Not yet established
User and AI outcomes
No evidence yet that the prototype improves confidence or independent completion. Live-model reliability, full error recovery, screen-reader use and real 200% zoom remain to be evaluated.
What would change my decision?
I would compare the deterministic guide with a model-assisted version on the same tasks. If interpretation adds unsupported instructions or does not improve understanding and recovery, I would keep the rules-based experience. Speed alone would not justify the added uncertainty.
Proposed release gate: no assistant-triggered confirmation; unavailable targets fail closed; pausing invalidates delayed replies; private inputs are handled without transmission in the checked scenarios. Passing a finite test set would not establish universal safety.
What I would measure next
- Task understanding: can the person explain what the next action will do before taking it?
- Recovery: can they pause, correct a mistake and resume without the assistant advancing unexpectedly?
- Guidance reliability: how often does the model propose an unavailable target or unsupported explanation?
- Incremental AI value: does model-based interpretation or reformulation improve the experience compared with the deterministic guide?
These are proposed evaluation questions, not measured improvements. The next integration should be judged against the guided baseline using ambiguous requests, missing controls, changing screens, delayed responses and sensitive-information prompts before a participant study.
My main learning
The value of a banking companion depends on boundaries as much as conversational quality. Even a correct instruction becomes unhelpful if its panel covers the task, its voice is unavailable or the person cannot tell who will perform the next action. The next stage is to test those boundaries with people and then evaluate whether a model adds enough value to justify its uncertainty.
Explore the original research: 73 responses, including 40 participants aged 55–80
These figures describe the original survey, not a usability evaluation of the prototype. Research artifacts and methods are preserved below.
01 · Challenge
Banking apps were useful, but trust was fragile
Many older adults in the study described banking apps as confusing, difficult to navigate and risky to use without support. The project explored what made people hesitate and what would make digital banking feel more trustworthy.
The original research phase produced insights, personas, journeys, benchmark analysis and recommendations. It did not produce a tested product prototype; the later development phase is described above.
02 · Method
The study combined desk research, survey data and interviews
Desk research
Official sources and market references helped frame the growth of smartphone and banking app use among older adults.
Quantitative survey
A 20-question survey gathered 73 responses, including 40 respondents within the 55-80 target group.
Qualitative interviews
Interviews with older adults explored fears, support needs, language barriers and real experiences with banking apps.
03 · Artifacts
Personas helped make different levels of digital confidence visible
The research produced two personas and empathy maps representing different relationships with technology: one person with lower digital familiarity and another with more experience using apps.
04 · Findings
The strongest signal was not age. It was perceived risk.
The clearest pattern was fear of making a mistake with money. That fear connected to several practical barriers: small text, complex language, too many options, unfamiliar symbols, difficult login flows and weak feedback after actions.
Fear of costly mistakes
Interviewees described insecurity around tapping the wrong option or moving money unintentionally.
Banking language created distance
Acronyms, English terms and technical wording made key actions harder to understand.
Support mattered
Access to a real person was repeatedly connected to confidence, welcome and trust.
05 · Recommendations
The recommendations focused on clarity, reassurance and control
- Use larger text, stronger contrast and simpler layouts to reduce visual effort.
- Replace acronyms and technical terms with plain, direct language.
- Add confirmation before important actions and clear feedback after each step.
- Offer step-by-step guidance with images, especially for unfamiliar tasks.
- Keep access to human support visible instead of treating the app as a full replacement for help.
06 · Synthesis
The CSD matrix helped turn desk research into better questions
I used a certainties, assumptions and doubts matrix to decide what the survey needed to ask. It kept the research grounded in what was already known while making uncertainty visible.
07 · Reflection
Contribution and next steps
The research helped frame the prototype around confidence, language and control. Its recommendations informed a later exploration; they do not establish the effectiveness of the new interface.
- Connected behavior, confidence and accessibility instead of treating age as a single design problem.
- Used research artifacts to make needs visible before proposing solutions.
- Translated the research into recommendations ready for prototyping and validation.