An AI system that produces good outputs but is presented poorly is not a good product. The interface through which users interact with AI determines whether they trust it, use it correctly, act on its outputs, and return. For African markets, this challenge is compounded by channel diversity — the same underlying AI capability may need to be surfaced through a smartphone app, a USSD menu, a WhatsApp chatbot, a voice call, and an agent portal simultaneously.
This session covers AI-specific UX principles that do not apply to traditional software, the specific design constraints of channels dominant in African markets, and how to design for failure, uncertainty, and human escalation — the three UX challenges most specific to AI products.
Why AI UX is different from traditional software UX
Traditional software UX is designed around deterministic behaviour: a button does the same thing every time, a field accepts a defined set of inputs, a process has a fixed sequence of steps. Users learn these patterns quickly and develop trust through consistency.
AI introduces probabilistic behaviour: the same input can produce different outputs, outputs can be wrong even when they look confident, and the system can fail in ways that are not legible to the user. These properties require a fundamentally different UX approach.
The three AI-specific UX challenges are: communicating uncertainty (how do you show a user that an AI recommendation might be wrong without undermining their confidence in the product?), handling failure gracefully (what happens when the AI gets it wrong, and how does the user know?), and managing appropriate reliance (how do you prevent users from over-trusting the AI, while also preventing them from ignoring its recommendations?).
A useful framing is the concept of appropriate reliance calibration. Research on human-AI teaming consistently shows two failure modes: over-reliance (users accept AI recommendations without sufficient critical evaluation, including when the AI is wrong) and under-reliance (users ignore AI recommendations even when they are correct, because they do not trust the system). Good AI UX design calibrates reliance to match actual system reliability: high reliance when the system is reliable, scepticism when confidence is low.
For a loan pre-screening assistant, appropriate reliance calibration means: when the extraction confidence is high, present the recommendation clearly and reduce friction for the loan officer to accept it. When confidence is low, present the uncertainty explicitly, show the raw data alongside the interpretation, and make human review the default path rather than an escalation.
Designing for USSD and low-bandwidth channels
USSD (Unstructured Supplementary Service Data) is a session-based protocol that works on all mobile phones, including basic feature phones, and does not require an internet connection — only a voice connection. In East and West Africa, USSD is the dominant channel for mobile financial services, particularly for users outside the middle class. Any AI product that ignores USSD is ignoring a large fraction of the addressable market in most African markets.
USSD imposes severe design constraints: sessions are time-limited (typically 180 seconds total), menus are text-only with no formatting, responses are limited to approximately 160 characters, and the interaction is sequential — you cannot jump back to a previous step. These constraints make it impossible to simply port a smartphone AI experience to USSD.
Designing AI for USSD requires a different frame: AI in the backend, simple UI in the front. The AI processing — intent classification, context retrieval, decision support — happens server-side, and only the output of that processing reaches the USSD session. The USSD session presents pre-formatted options and short response fragments, not open-ended AI responses.
A practical pattern for USSD AI: use AI to personalise the menu options shown to a specific user (based on their history and context), to interpret an ambiguous selection (when the user selects 'Other' and types a free-text reason, classify it before routing), and to generate the brief follow-up messages after the session (a 160-character SMS summary of the session outcome, or a callback reason summary for a loan officer). The AI is invisible to the user but improves every step of the interaction.
SMS and WhatsApp offer more flexibility than USSD — longer messages, richer formatting in WhatsApp, asynchronous communication — but still require careful adaptation from smartphone app experiences. WhatsApp Business API with AI integration is increasingly common in African fintech for loan status updates, repayment reminders, and customer service queries. Key design principle: AI-generated messages in WhatsApp should be short, conversational, and written in the register of the customer — not the register of an English corporate communication.
Voice UX for African markets
Voice AI — conversational agents that interact with customers through phone calls — is among the most promising AI applications in African markets because it eliminates the literacy barrier entirely. A customer who cannot read or write English can still interact naturally with a Swahili voice agent. This represents a fundamental accessibility advance over any text-based interface.
Building voice AI for African markets requires addressing several challenges that are not present in English-language voice AI development:
Accent and dialect variation. African languages have significant regional variation. Nairobi Swahili sounds different from Mombasa Swahili; Lagos Yoruba is not the same as Ibadan Yoruba. A speech recognition system trained primarily on standard or textbook versions of these languages will have higher error rates on regional varieties — and the users with the strongest regional accents are often the least well-served by existing financial services, making accuracy particularly important.
Code-switching. Many African speakers naturally switch between two or more languages within a single conversation — mixing English with Swahili, French with Lingala, Arabic with Somali. A voice AI system that cannot handle code-switching will frequently misinterpret inputs from natural speakers.
Background noise. A significant proportion of calls to financial services come from noisy environments — markets, transport, outdoor spaces. Speech recognition systems trained primarily on clean audio will degrade significantly in noisy conditions. This should be tested explicitly during system evaluation, not just assumed to be acceptable.
Designing the dialogue. Voice AI dialogue design requires a different skill set from chatbot or text AI design. Conversations do not have back-button or correction affordances. Misunderstandings must be resolved within the flow, not by starting over. The system needs to confirm critical information (loan amounts, repayment dates) through read-back and explicit confirmation, not assume the user heard correctly. Confirmation steps should be designed to catch misrecognition without creating excessive friction for the common case.
Designing for uncertainty and progressive disclosure
AI systems produce outputs with varying degrees of reliability. Good UX design communicates this variance — not hiding uncertainty behind false confidence — while avoiding creating anxiety about outcomes that are actually reliable.
Progressive disclosure is the principle that the default view should show the most important information clearly, with detail available on demand. Applied to AI outputs: show the high-level recommendation or result prominently, make the underlying reasoning and confidence available one level deeper, and make the raw data or model details available for the user who wants to go further.
For a loan pre-screening recommendation presented to a loan officer: the primary view shows the recommendation (proceed / review further / decline) with a one-line reason. A secondary view shows the extracted income figures, the confidence interval, and the key supporting signals. A tertiary view shows the raw statement data used for extraction. Most loan officers will operate at level one most of the time; experienced officers and complex cases will use levels two and three.
Confidence as a design element. Rather than hiding model confidence from users, expose it in ways that are actionable. A colour-coded confidence indicator (green/amber/red) is more usable than a 0.73 probability score. A label like 'Strong signal' / 'Moderate signal' / 'Weak signal, manual review recommended' communicates the same information in a register that a loan officer can act on.
Uncertainty in customer-facing contexts requires more care. Customers have a lower tolerance for technical uncertainty language than internal users. Instead of 'the model is 65% confident you qualify', prefer 'based on what you've shared, you may qualify — complete your application and we'll review within 24 hours'. The goal is honesty without anxiety: the customer should not be given a false certainty, but should also not be left in a state of confusion that prevents them from completing the application.
Human-in-the-loop design
The phrase 'human-in-the-loop' is often used as a compliance or safety measure — a way of saying 'a human reviews AI decisions'. This framing is too narrow. Human-in-the-loop is a design discipline that determines which decisions AI makes autonomously, which decisions AI supports but humans make, and which decisions are always made by humans regardless of AI capability.
The framework for making these determinations has two dimensions: decision stakes (how much harm results from an incorrect decision?) and AI reliability (how often is the AI correct on this type of decision?). The combination of these two dimensions produces four quadrants:
High stakes, high reliability: AI makes a recommendation with high confidence; human reviews and can override but defaults to the recommendation. Appropriate for: standard credit approvals within established parameters.
High stakes, low reliability: Human always makes the decision; AI provides supporting information. Appropriate for: first-time applications from new market segments, complex income structures, disputed defaults.
Low stakes, high reliability: AI makes the decision autonomously. Appropriate for: eligibility pre-screening, automated document classification, fraud risk flagging.
Low stakes, low reliability: Consider whether the AI feature adds value at all, or whether a simpler system would serve better.
The design of the human escalation path is as important as the AI path. How does the system detect that an application should be escalated? How is it routed? What information does the human reviewer receive? How does the human's decision get fed back into the system? Poor escalation design produces bottlenecks that undermine the efficiency gains from AI automation — the AI flags 30% of applications for review, but the review queue is staffed for 5%, creating a backlog worse than the manual system it replaced.
Self-check questions
Before moving to the next session, make sure you can answer these without re-reading. If you cannot, go back to the relevant section.
- 1What is appropriate reliance calibration and why does it matter for AI UX design?
- 2Describe three USSD constraints that affect AI product design, and explain a practical pattern for delivering AI value within those constraints.
- 3Using the two-axis framework (decision stakes × AI reliability), classify five different decisions in a lending product into the appropriate human-AI collaboration pattern.
- 4What is code-switching and why is it a specific challenge for voice AI in African markets?
Session deliverable
AI UX Flow Diagram
Design a UX flow for an AI-powered feature across at least two channels, including the error states, uncertainty communication, and human escalation paths. The deliverable should be specific enough to hand to a designer for wireframing.
Show worked example ↓
Loan eligibility pre-check — automated income assessment
Show worked example ↓
USSD (*334#), WhatsApp Business chatbot, Loan officer portal (web dashboard)
Show worked example ↓
USSD: Customer dials *334#, selects 'Check loan eligibility'. Prompted to upload M-Pesa statement (receives SMS with upload link). AI processes statement (~60s), returns eligibility indicator via SMS: 'You may qualify for up to KES 15,000. Reply YES to proceed or visit [link]'. WhatsApp: Same trigger, but customer can send statement directly in chat. AI responds with richer message: estimated limit + 3 bullet supporting signals + CTA button. Officer portal: High-confidence applications appear in 'Ready for review' queue with pre-filled income assessment. Officer sees recommendation + confidence badge (green) + one-click approve.
Show worked example ↓
USSD: If confidence < 0.7, SMS reads: 'We need more information to check your eligibility. Our team will call you within 24 hours.' No limit shown — avoids anchoring on unreliable estimate. WhatsApp: Low confidence triggers: 'We're reviewing your details and will update you within 24 hours. You can also visit your nearest agent.' Officer portal: Application appears in 'Needs review' queue with amber confidence badge. Income estimate shown with wide confidence range (e.g. KES 8,000–22,000). Raw statement data accessible one click deep.
Show worked example ↓
USSD: API failure or processing error → customer receives: 'We're experiencing a delay. Please try again in 30 minutes or visit your nearest agent.' No error code shown to customer. WhatsApp: Same message. Officer portal: Application flagged as 'Processing error' with error code and timestamp. Auto-routed to manual queue. Error logged to monitoring dashboard with alert if error rate > 2% in any 5-minute window.
Show worked example ↓
Triggers: (1) confidence < 0.7, (2) income source = 'mixed business', (3) anomaly flags present, (4) statement < 2 months, (5) requested amount > 2× estimated income. Human reviewer receives: statement summary (not raw PDF by default), AI extraction with confidence intervals, anomaly flags with plain-language explanation, customer history (previous loans, repayment record). Decision recorded in loan management system with reason code. All escalated decisions feed back to model evaluation dataset with a 2-week delay.