Stanlib Libby
Libby is an AI powered internal servicing platform for STANLIB call centre agents and advisors. It brings client lookup, case management, instruction tracking and a knowledge assistant into one workspace, so an agent can find an answer, act on it and leave a clear audit trail without leaving the call.
Role
Product Design Lead & Owner
Year
2026
Platforms
Internal Web App
Industry
Asset Management
Task
I owned the product design of Libby end to end: research, information architecture, user flows, the design system, high fidelity UI and interactive prototypes, through to launch material for the internal rollout.
The brief was to replace a scattered set of admin systems, shared documents and tribal knowledge with one calm, trustworthy servicing tool. Agents needed to answer investor queries faster, advisors needed confidence that every instruction was tracked, and the business needed every AI assisted action to be explainable and auditable.
Libby works in two modes. Execute helps an agent carry out servicing tasks. Advise helps them understand products, processes and rules. Two assistant identities, Libby (LB) and the LearnNow (LN), make it clear at a glance who is answering and where the answer came from.
Context
2002
STANLIB founded in Johannesburg, now part of the Standard Bank Group via Liberty.
R804bn
Assets under management in 2025, serving retail, advisor and institutional clients.
8am to 5pm
Weekday contact centre hours, when every minute an agent spends searching is a minute a client waits.
2 roles
Call centre agents and advisors, each with different jobs, permissions and levels of product depth.
The Challenge
Servicing investments is detail heavy and regulated. The problem was never a lack of information. It was that the information lived in too many places and agents had no quick way to trust it.
Fragmented
Too many systems per call
A single query could mean switching between admin screens, shared drives and colleagues on chat, all while keeping the client on the line.
Inconsistent
Answers depended on who you asked
Process knowledge on fees, tax, transactions and digital channels sat with senior staff, so new agents escalated often and answers varied.
Untraceable
AI had to earn trust
In a regulated environment an AI suggestion is only useful if the agent can see its source, confirm it, and prove afterwards what was done and why.
My Role
As Product Design Lead and owner of the UX and UI, I was the single design voice on Libby from the first flow diagrams to the Phase 2 release. I worked directly with the product, engineering and servicing teams and made the design calls across every screen.
Agent and advisor interviews, call observation, pain point mapping
Research
DTCG token architecture in Token Studio, components, patterns
Design system
Execute and Advise model, assistant identities, feature priorities
Strategy
Every screen, plus interactive HTML prototypes for testing
UI & prototypes
Phase 2A and 2B user flows, architecture diagrams
IA & flows
LearnNow copy,
Launch
Design Process
UX Research
Research focused on the real rhythm of a servicing call: verify the client, find the account, understand the request, act, and log it. I observed agents on live queries and interviewed advisors about how they track instructions they submit on behalf of clients.
The pattern was clear. Agents were confident with the systems they used daily but lost time on rarer requests, where the answer lived in a document they could not find quickly. Advisors had no single view of where an instruction sat once submitted.
➊
Knowledge retrieval was the slowest step on complex calls
➋
New agents escalated process questions that seniors could answer in seconds
➌
Instruction status required follow up calls and emails
➍
Any AI answer without a visible source would not be used
"I know the answer is in a document somewhere. The client just doesn't want to wait while I find it."
Call centre agent
"Once I submit an instruction I lose sight of it. I only find out there's a problem when my client calls me."
Advisor
"If it tells me something, I need to see where it got it from. I'm the one accountable."
Senior agent
User Flows
I mapped the servicing journey in two phases. Phase 2A covered the core agent loop: verify, search, ask, act and log. Phase 2B extended it to advisor instruction tracking, escalation and team oversight.
Every flow ends in the audit trail. That rule shaped the whole product: if a path could not be logged, it was redesigned until it could.
Two Modes, Two Voices
The central design decision was to separate doing from knowing. A single toggle at the top of the workspace tells the agent, and the AI, which job is under way.
For carrying out servicing work on a live case. Libby prefills actions and the agent always confirms before anything is submitted.
Execute
-
Prefilled instructions with review step
-
Case context carried across screens
-
Every action written to the audit trail
For understanding products, rules and processes. The Knowledge Bot answers from approved content and shows its sources.
Advise
-
Cited answers from the knowledge base
-
LearnNow guided decision paths
-
No actions taken on client accounts
LB
Libby, the servicing assistant that acts in Execute
LN
LearnNow, the reference voice that answers in Advise
Key Features
Case Search
A stepped search with a query builder and custom date range picker, so agents can find any case from partial details.
6 iterations
Track Instructions
A dense, scannable table of client instructions with status at a glance, filters and a full history per row.
Built with realistic IFA data
CLO Dashboard
A team view of workload, turnaround and exceptions for the people responsible for service quality.
Oversight
LearnNow
A chat interface that answers from approved STANLIB content, with source chips on every reply.
Advise mode
Actions Tab
A redesign that groups servicing actions by task, with clear confirmation and logging on each.
Redesigned
Profile & Settings
A profile dropdown with fully interactive drawers and modals for preferences, access and help.
Drawers and modals
Case Search
Case Search went through six iterations. Early versions used a single search box, which failed when agents only had fragments: a surname, a rough date, a fund.
The final design is a stepped flow. Agents pick what they know, build a query with plain language rows, and narrow by date with a custom calendar that handles ranges properly. Active filters sit as removable chips so the agent never loses track of what they asked for.
Track Instructions
Advisors told me they lost sight of instructions once submitted. I designed the table around the question they actually ask: what needs my attention today? Status is encoded in colour and label, and rows that need action sort to the top.
LearnNow
LearnNow is the guided learning layer inside Advise. Instead of a wall of policy documents, agents follow a three level decision tree from topic, to scenario, to a specific answer with its source.
I designed and prototyped the full chat interface across several versions, then built a suite of interactive prototypes covering eight topic areas, each validated so no path led to a dead end. I also wrote and iterated the LearnNow marketing copy for the internal launch.
Design System
To keep a growing product consistent I built a token based design system in Token Studio using the DTCG format. Primitive colours come from the STANLIB palette. Semantic tokens reference those primitives and never raw hex values, so the brand can be retuned in one place.
Semantic tokens follow a five level structure across four categories. The selected purpose is kept separate from brand so it can be themed on its own, and focus rings live only under border.focused to avoid clashing with interactive states.
Background
Surfaces and fills
Text
Content and labels
Border
Edges and focus
Icon
Glyphs and status
{
"color": {
"darkblue": {
"50": {
"$type": "color",
"$value": "#F4F5F6"
},
},
"xs": {
"$type": "lineHeights",
"$value": "15px"
},
},
"wider": {
"$type": "letterSpacing",
"$value": "1%"
},
},
"spacing": {
"0": {
"$type": "spacing",
"$value": "0px"
},
},
"base": {
"$type": "borderRadius",
"$value": "6px"
},
statureblue
darkblue
cmykteal
danube
oceangreen
greywhite
Prototype
Discover More
