People Assist AI
Replacing fragmented HR portals with a conversational AI assistant so every employee gets a direct, trusted answer in seconds, not minutes.
Enterprise HR · AI
Knowledge Management
Type
UX Designer + Motion Designer
Scope
2023–2024 · 6 months
Platform
2 UX · PM · Devs · HR Team
Domain
Web (Desktop) + Mobile App
Overview
What is PeopleAssist AI?
An enterprise employee experience platform that replaces keyword-search HR portals with a conversational AI assistant letting every employee ask a natural-language question and receive a direct, source-cited answer drawn from the organisation's existing knowledge base.
The organisation's knowledge base was already comprehensive. Policies, benefits, payroll guides, mobility documentation, referral processes it was all there. The problem was that finding the right answer required employees to navigate multiple portals, open several articles, and reconcile information manually. When that failed which it did often they raised a support ticket. Every ticket was a self-service failure that cost both the employee and the HR team time.
PeopleAssist AI addresses this by shifting the interaction model from document retrieval to conversational assistance. Employees type or speak a question in plain language. The AI interprets intent, searches across knowledge repositories, generates a concise answer, and cites the sources all in a single, trusted response.
My Role | UX Designer + Motion Designer — end-to-end |
|---|---|
Team | 2 UX designers · PM · Developers · HR team |
Platform | Web (Desktop) + Mobile App |
Tools | Figma · After Effects |
Timeline | 2023–2024 · 6 months |
Status | ✓ Shipped · Live enterprise users |
"The organisation already had the answers. The problem was that employees had to work harder than the question deserved to find them and every extra minute of searching was a minute of trust quietly walking out the door."
Objectives & Goals
Target Audience
01 · Research
Qualitative Research
Before a single wireframe was drawn, the team went to the real source employees using the existing system every day, and HR teams fielding the tickets when self-service failed. Research ran through internal surveys and HR email feedback channels to capture both the employee experience and the operational cost of the status quo.
Research Methods
01 · Internal Employee Surveys - Structured surveys distributed internally across employee populations. Covered frequency of use, types of questions asked, how often employees found what they needed, and what they did when they didn't. Looking for: breakdown points in the current experience; what "success" looked like vs. what prompted a ticket or a workaround.
02 · HR Email Feedback Analysis - Analysis of HR support email threads to identify the most common question categories, the volume of repeat queries, and which topics generated the most back-and-forth before resolution. Looking for: ticket taxonomy, frequency of deflectable vs. complex queries, and which content gaps were driving the most repetitive support load.
03 · Existing Portal Audit - A structured walkthrough of the existing keyword-search portal testing representative queries drawn from the most common HR email categories. Measured number of navigation steps, article opens per query, and reading time before finding a usable answer.
Key Findings - what was observed
01 · Employees were doing the AI's synthesis work themselves The portal returned a ranked article list. Employees opened 3–5 documents, scanned each for the relevant sentence, and reconciled across them manually. Survey responses consistently described the portal as "time-consuming" even when the right content existed.
02 · One failed search was the escalation threshold Survey data showed employees rarely tried more than one or two searches before abandoning the portal and raising a ticket or emailing HR. Escalation felt faster and more reliable than retrying search.
03 · Most HR support emails were answerable by existing KB content HR email analysis showed that the dominant question categories benefits, referrals, mobility, payroll, expense setup all had existing documentation. The content existed. The retrieval mechanism consistently failed to surface it cleanly.
04 · Employees needed to see the source before acting on an answer - Source visibility was not optional it was the condition for the answer to be usable at all. Policies change; acting on outdated information had real consequences.
05 · Mobile was a genuine operational gap, not just a convenience ask - A significant survey cohort needed HR information while travelling, on-site, or between meetings. The existing portal was functionally unusable on mobile, disproportionately affecting the workforce with the least access to other support channels.
06 · Employees naturally phrased needs as questions, not search terms - Open-text survey responses were overwhelmingly conversational "How do I refer a candidate?", "What is the bonus eligibility policy?" not keyword fragments. The mental model was already conversational. The system forcing keyword input was the design mismatch.
Key Insights - what the findings mean for design
Findings describe what was observed. Insights state the design implication each must trace forward to a specific decision.
"I don't use the portal anymore. I just email HR directly it's faster. The portal never gives me a straight answer, it just gives me more things to read."
— Representative employee, Operations function · Internal survey response
Define
Synthesising Into a Problem
Synthesising research into a single, precise problem statement is what makes every subsequent decision purposeful. The bridge sentence: research revealed what was broken. The existing portal audit confirmed why it persisted. The problem statement is the single sentence that focuses everything that follows.
Employees across all functions need a way to get a direct, trusted answer to an HR, policy, or workplace question in seconds, from any device because the existing keyword-search portal distributes information across systems without synthesising it, forcing employees to perform manual research that leads to frustration, abandoned self-service attempts, and preventable HR support tickets.
How Might We…
HMW 1 - How might we turn a fragmented, multi-system knowledge base into a single conversational assistant that gives employees a direct answer not a reading list?
HMW 2 - How might we make AI-generated answers feel trustworthy enough to act on specifically for compliance-sensitive HR and policy topics where getting it wrong has real consequences?
HMW 3 - How might we design the escalation path to a human as a first-class feature so employees never feel stranded or penalised when the AI answer isn't sufficient?
HMW 4 - How might we build a mobile experience that serves employees' high-urgency questions on-site and in transit without compromising the depth available on desktop?
HMW 5 - How might we reduce HR support ticket volume through better self-service without creating an experience where employees feel blocked from accessing human support when they genuinely need it?
User Personas
Empathy Map — Priya M. (Primary User)
Priya M. - Senior Engineer · Primary User · Derived from survey responses and HR email pattern analysis · Desktop + Mobile
03 · Ideate — User Flows
Six flows mapped before a single screen was designed. Every journey branches. Every flow shows what the system does automatically, where users make decisions, what happens when something fails, and how recovery works.
Flow 01 · Primary Query Employee opens PeopleAssist AI → types question in natural language → System queries knowledge repos → If answer found: show answer card + cited sources → Employee reads and is satisfied → Task complete ✓. If not satisfied → escalation path.
Flow 02 · Escalation to HR AI answer shown with source citations → Employee not satisfied → Taps "Still need help?" on answer card → Pre-filled request form with query context → System routes to HR specialist queue → Request raised ✓.
Flow 03 · Source Verification Employee receives answer → taps a source chip → opens the originating KB article → verifies the relevant section → returns to PeopleAssist and acts. Branch: flags the answer as potentially outdated, routing to KB review queue.
Flow 04 · Follow-up Conversation Employee receives initial answer → asks a follow-up question in context → AI maintains session context → returns a contextually-aware second answer → suggests related topics. Branch: new unrelated topic vs. continuation vs. escalation.
Flow 05 · HR Admin Oversight (Marcus) HR Business Partner reviews self-service analytics dashboard → sees top unanswered query categories → identifies KB gaps → triggers content update request → monitors deflection rate trend over time.
Flow 06 · Mobile Field User (Raj) Raj on-site → opens mobile app → taps a suggested question chip → receives answer → bookmarks it for offline access. Branch: poor connectivity → app serves cached answer → "View latest version when connected" prompt.
04 · Information Architecture
Structural Thinking
The IA reflects the shift from a browsable document directory to an answer-first system. Navigation is organised around what employees need to do not how the HR team categorises its internal content.
Design Principles
01 · Answer first, source second - The primary surface always shows a synthesised answer never a list of documents. Source citations appear as secondary, verifiable chips beneath the answer. The answer is shown first, always. The sources are for verification, not navigation.
02 · Visible uncertainty beats false confidence - When the AI cannot return a high-confidence answer, it says so explicitly and immediately offers escalation. A confident-sounding incorrect answer on a policy topic erodes trust permanently.
03 · Escalation is a feature, not a failure - The path to a human is always one action away and never feels like giving up. Employees should feel the system supports both paths with equal care: self-service and human-assisted are both valid outcomes.
04 · Mobile is a first-class platform - The mobile experience has different interaction components suggested chips over typed input, bookmarked offline answers, a compact request tracker. Any design decision that makes mobile feel like a compromised desktop is rejected. Raj's experience is as considered as Priya's.
05 · Style Guide — Visual Language
A purposeful colour system designed for clarity and communication. Every colour carries semantic meaning before a word is read: blue for primary actions, green for confirmed answers, orange for escalation, violet for AI-automated steps, crimson for errors.
06 · High Fidelity — Designed Screens
Three responsive breakpoints each designed for a different user in a different context. Desktop for Priya at her workstation. Tablet for Marcus reviewing analytics in a meeting. Mobile for Raj on-site at a customer facility.
Click Here to view High Fidelity Figma Files Link
Before & After
07 · Motion — Why Motion Matters Here
An AI assistant has a trust problem that static screens cannot solve alone employees cannot see the system thinking. Motion makes the invisible process visible, and transforms an unfamiliar technology into something that feels competent, transparent, and worth trusting.
01 · Showing the system thinking, not frozen. The 1–2 seconds while the AI queries the knowledge base is the highest-risk moment in the interaction employees don't know if the system is working or broken. A purposeful loading animation that communicates active search (not a generic spinner) is the difference between patience and abandonment.
02 · Revealing the reasoning path before the answer arrives. Showing source documents being "found" before the answer card appears makes the answer feel earned and traceable not arbitrary. This single motion sequence is the largest contributor to employees trusting an AI-generated answer on first use.
03 · Making the old experience vs. new experience vivid. The companion motion explainer contrasts the old fragmented portal journey with the new conversational flow communicating the product's value proposition to stakeholders and employees in a way a static presentation cannot.
08 · Results & Impact
Intended Impact
PeopleAssist AI was designed to move four specific outcomes tied directly to the Objectives in Section 00. The platform was shipped and is live with enterprise users. The metrics below represent the designed intent targets framed as intended impact per confidentiality standard.
Qualitative Outcomes
Field employees and mobile-first workers gained a functional self-service channel for the first time the mobile experience addressed a previously invisible use case.
Source attribution directly addressed the compliance anxiety that had made previous AI implementations in internal tools feel risky. Employees could verify before acting.
The KB gap feedback loop transformed the HR admin role from reactive to proactive identifying and closing knowledge gaps before they generate more tickets.
Designing escalation as a first-class path removed the ambiguity employees felt about when to raise a ticket vs. keep searching.
09 · Reflections
Hardest Trade-off Prototype-first vs. design-system-first
The project began with prototype designs to validate the conversational concept quickly with HR Directors and VP stakeholders. That was the right call for moving fast and securing approval. The cost came in phase two because the prototype components hadn't been built against a design system, the visual language had drifted in ways that required significant reconciliation work. If I were doing this again, I'd establish the core design tokens in week one even before the first prototype so that the prototype builds toward the final system rather than away from it.
What I'd research differently -The survey approach gave us broad signal efficiently, but it missed the nuance of watching employees actually use the old portal in context. One session of task-based observation would have given us the specificity in Finding 01 earlier and with more convincing evidence for stakeholder presentations.
What this project changed - PeopleAssist AI changed how I think about the first three interactions in any AI-assisted product. Those first moments are trust-formation interactions not task-completion interactions. Everything about the UI in that window is doing more psychological work than any other moment in the product lifecycle.
"Designing an AI self-service product isn't really about building a smarter search engine. It's about engineering the exact moment when an employee decides whether the answer in front of them is worth acting on and building everything backwards from that moment of trust."
— Charan Raj · PeopleAssist AI · 2024
Concept Case Study · Anonymised from real work · Client details available privately on request
