Smart Assistant
An in-app AI assistant designed to help safety professionals complete complex tasks faster, through natural conversation without leaving the page they're working on.
Company
ecoPortal
My role
Product designer
Tools
Figma and Claude Code
Problem statement
ecoPortal wants to bring an AI assistant into its health and safety platform, designed specifically for teams meeting AI for the first time.
The goal is to support the safety officers, site managers, and frontline staff who already run their daily work inside the platform, but who are curious about AI yet unsure whether to trust it, unclear on what it can actually do, and cautious of getting it wrong in work that feeds compliance and real-world risk. These users bring deep safety expertise, little experience with AI chatbots, and low confidence in how to prompt one or judge its answers.
The assistant should make AI feel approachable, transparent, and safe to use in a regulated environment, while earning trust, showing how it reached an answer, and giving the organization a clear read on whether it actually helps. Success means guiding users from "I'm not sure I can trust this" to "I know what to ask, and I can rely on what it gives me."
Research findings
| Metric / Information | Business / Industry Problem | User Need | Opportunities / Features |
|---|---|---|---|
| Frontline workers who use AI regularly report saving around 8 hours a week. | Manual report and audit work eats hours, and competing platforms are already adding AI. Standing still means falling behind. | Cut the manual load and catch gaps, without leaving the platform they already use. |
|
| Only 26% of people trust AI, and just 14% trust it to act on their behalf. | An untrusted feature goes unused, and a poor AI experience can erode confidence in the wider platform teams already rely on. | A trustworthy assistant that shows its reasoning, so users can verify before acting on high-stakes calls. |
|
| 77% of AI users paste data into prompts and around 22% of those include PII or payment data. An estimated 10 to 50% of chatbot interactions are abusive or off-topic. | In a regulated platform, sensitive or careless inputs create data-leak and compliance risk, and off-topic use wastes the tool. | Guardrails that keep them safe when they enter the wrong thing, with a nudge back to what the assistant is for. |
|
| 60% of users never send a first message; prompt suggestions can roughly triple engagement; people decide whether to engage in under 5 seconds. | A capable assistant no one knows how to start with gets abandoned, stalling adoption and wasting the build. | A clear starting point and examples, so people new to AI know what to ask and what it can do. |
|
| Thumbs up/down is the standard signal behind model tuning, and in aggregate it shows teams what is working and what is not. | Without a feedback signal, the organization can't tell where the AI helps or fails, or justify keeping it. | A quick, low-effort way to flag a good or bad answer, and trust that it improves the tool. |
|
Key design decisions
Conversational UI is product design, not UI design
Designing the Smart Assistant required thinking beyond screens, about what the AI would say, when it would speak, what it wouldn't claim to know, and how errors would be communicated.
Designing for ambiguity
Some parts of the assistant's capabilities were still being defined while the design was in progress. Designing flexible, durable patterns that could absorb product decisions made later without going back to square one was key.
The suggestions system as a UX philosophy
It wasn't clear whether users would know the assistant's capabilities. Instead of displaying a blank input field, contextual suggestion chips were provided that surfaced what was possible, giving users a confident starting point.