at a time.
Nothing later can clarify what was misunderstood earlier.
05 Selected work / Human Factors Institute
Standard Bank · South Africa
Telephone banking · Audio-only interaction
Standard
Bank.
Language is
architecture.
For customers without smartphones or reliable data, telephone banking was not a fallback. It was the bank. With no screen to scan and no control to recognise, every word had to carry structure, state, and trust.
Information architectureFrom bank structure to human intent
Before · the bank’s modelWhat product
do you have?
do you have?
- Savings products
- Credit products
- Loan products
- Investment products
→Card sort
Mental model map
Mental model map
After · the customer’s intentWhat do you
want to do?
want to do?
- Check my balance
- Pay a bill
- Transfer money
- Report a problem
01
The defining constraint
You cannot
scan sound.
A visual interface lets people pause, compare, recognise controls, and look ahead. An IVR disappears the moment it is spoken. Customers had to hold the system in memory while deciding what to do next.
at a time.
not recognition.
Menu structure has to be learned through sound alone.
costs confidence.
For many customers there was no app or branch experience to recover with.
In an audio-only interface, phrasing is not the copy layer. Phrasing is the interaction model.
02
The finding
The wrong
word wins.
People did not patiently parse every prompt. They anchored to the first familiar, salient word and acted. One sentence exposed the failure with unusual clarity.
Please enter your telephone banking PIN.
10 digitsCustomers entered their phone number.
↳Same words.
Different order.
Different order.
Please enter the PIN for your telephone banking.
PINCustomers entered the intended credential.
Word order→
changed task completion without changing the vocabulary.
03
Testing an interface that did not exist
Prototype
the invisible.
Before implementation, we ran the IVR as theatre: a participant with a handset in one room, a researcher reading and branching through a Word script in another, and a camera capturing every hesitation.
Room A · ParticipantHears a real call.
Presses real keys.
Presses real keys.
Audio→DTMF←
Room B · “The system”
the branching logic.
IF 1 → BALANCEIF 2 → PAYMENTIF ERROR → REPHRASE
Researcher performsthe branching logic.
Observed failure
“Telephone” pulled attention away from “PIN.” The architecture was sound; the sentence was routing people incorrectly.
04
Designing beyond the happy path
Confusion should
change the system.
Silence and invalid input were treated as information. Recovery became progressively clearer, then deliberately handed control to a person instead of trapping the customer in a loop.
“Enter your sixteen-digit card number.”
“Find the number on the back of your card.”
Repeat the shortest valid instruction.
“Let me connect you to someone who can help.”
Distinct tones became acoustic landmarks between levels and states.
Consistent keys created affordances customers could learn once.
The agent was a designed destination, not an admission of failure.
05
Why this work still matters
An IVR was already
an agent interface.
It maintained conversational state, interpreted input, selected the next action, recovered from uncertainty, and handed control to a human when it reached its limit.
The capability ceiling has changed. The core responsibility has not: make intent legible, make state understandable, and never hide a loss of control behind fluent output.
Two rooms, one Word document, and a telephone handset taught me that conversational systems are designed in the moments when language changes what a person does.
Paired Human Factors case study
← Previous · The Minimalist