AI Risk Scoring: From Hospitals to SA Online Shops

Most companies think “AI” means flashy chatbots and auto-generated product copy. The more valuable use case is quieter: risk scoring—using your existing data to spot problems early and trigger the right follow-up.

A South African-linked example from healthcare makes the point better than any e-commerce pitch deck. AI engineer Sylvester Tafirenyika, co-founder of RoyalTech AI Labs, has built a patented system that reads hospital discharge notes and predicts who’s likely to be readmitted within 30 days. That’s a life-and-death environment, and the lesson for e-commerce and digital services is refreshingly practical: accuracy is useless if it doesn’t fit into real workflows, protect privacy, and produce clear next actions.

If you run an online store or a digital service in South Africa, you’re fighting a different version of the same problem: customers “bounce back” with repeat support tickets, refunds, churn, and missed deliveries. The playbook is surprisingly similar.

Risk scoring is the real AI workhorse

Risk scoring is AI that helps teams act earlier, not just analyze later. In hospitals, it means identifying patients likely to return soon after discharge. In e-commerce, it means spotting orders likely to be returned, customers likely to churn, or tickets likely to escalate.

The healthcare readmission problem is well known: avoidable readmissions are expensive, stressful, and often preventable with timely follow-up. Tafirenyika’s approach focuses on a practical target—the 30-day readmission window—and uses a model trained for medical text (BioClinicalBERT) to pull signal from discharge summaries.

Here’s the direct translation to South Africa’s digital economy:

• Readmission risk → Return/refund risk
• Follow-up reminders → Proactive customer comms (WhatsApp/SMS/email)
• Discharge notes (free text) → Support tickets, reviews, delivery notes, call logs
• Clinical workflows → Ops and support workflows (Zendesk/Freshdesk, OMS, courier portals)

A strong AI initiative is rarely “build a model.” It’s build a loop: detect risk → trigger action → measure outcome → improve.

The underused goldmine: unstructured text

One detail from the hospital story should make every e-commerce operator sit up: the best information often lives in free text.

Clinicians write nuance into discharge notes that structured fields miss. The same is true for online retailers and digital services:

• Agent notes explaining why a customer is upset
• Product reviews that mention sizing, quality, or packaging issues
• Delivery exception notes (“customer not home”, “unsafe area”, “address unclear”)
• Chat transcripts where intent changes mid-conversation

Most companies over-invest in dashboards built on clean tables, while ignoring the messy text that explains why things happen. Natural language processing (NLP) is where AI earns its keep—because it converts chaos into patterns you can act on.

What healthcare gets right (and e-commerce often gets wrong)

The best AI systems reduce complexity for frontline teams. Tafirenyika’s stance is blunt and correct: accuracy alone is not enough. Tools have to be practical, scalable, and easy to adopt.

In South African e-commerce and digital services, most AI failures fall into one of these traps:

1. Model-first thinking: A data science project that looks great in a notebook but never reaches the warehouse floor or the support desk.
2. No operational decision: The model produces a score, but nobody knows what to do with it.
3. No privacy design: Customer data is copied into too many tools and spreadsheets, increasing POPIA exposure.
4. No measurement: Teams can’t prove the model reduced costs or improved customer experience.

Healthcare teams don’t have the luxury of hand-wavy success metrics. If a tool increases admin burden or creates confusion, it gets sidelined. That discipline is exactly what e-commerce needs.

Privacy-by-design isn’t optional in SA

One of the most practical design choices in the hospital system is keeping sensitive data from being sent to external servers by operating in a browser-based environment.

E-commerce isn’t dealing with clinical notes, but it is dealing with:

• Identity numbers (sometimes captured for credit checks or delivery)
• Home addresses and phone numbers
• Payment signals and fraud markers
• Potentially sensitive purchase history (health, religion, personal items)

Under POPIA, “we’ll sort compliance later” is a bad plan. If you’re implementing AI in digital services in South Africa, aim for:

• Data minimisation (only use what you need)
• Role-based access (support agents shouldn’t see everything)
• Audit trails (who accessed what, when)
• Clear retention rules (don’t keep transcripts forever)

A simple rule I’ve found useful: If your AI needs a full customer record to be useful, you probably haven’t defined the decision clearly enough.

Turning “readmission prevention” into an e-commerce playbook

The operational shape of the solution matters more than the algorithm. The hospital readmission predictor isn’t just a score; it includes structured profiles, reminders, timestamped notes, and dashboards.

That same structure maps cleanly to online retail and digital services.

Step 1: Pick a single “30-day problem”

Hospitals use 30 days because it’s measurable and meaningful. You need your equivalent.

Good “30-day problems” for SA e-commerce/digital services:

• 30-day churn risk for subscription services (streaming, insurance, fintech)
• 14-day return risk for apparel and electronics
• 7-day delivery failure risk for high-volume couriers
• 48-hour escalation risk for support tickets

Keep it narrow at first. Broad AI initiatives die from scope creep.

Step 2: Define what “follow-up” means in your world

Hospitals can schedule check-ins, meds reconciliation, or home-care support. In e-commerce, follow-up is your retention engine.

Examples of follow-ups triggered by a risk score:

• A proactive message: “We noticed delivery may be tricky—confirm your address pin?”
• A support callback for high-value customers with negative sentiment
• A sizing recommendation or exchange flow before a return request happens
• A fraud-safe step-up verification before shipping
• A warranty/installation nudge for complex products (reduces “it’s broken” returns)

The score is only valuable if it reliably triggers an action that a team can execute.

Step 3: Use text + structured data together

Tafirenyika’s work highlights the value of combining structured history with unstructured notes. That hybrid approach is usually stronger than either alone.

For an online store, a practical feature set might include:

• Structured: order value, delivery region, product category, prior return rate, payment type
• Unstructured: sentiment from chat, review themes (“too small”, “poor quality”), agent notes (“angry about delays”)

If you’re starting small, don’t chase perfection. Start with one or two text sources (support tickets are often the highest value).

Step 4: Build the workflow, not just the model

Hospitals succeed with AI when it sits inside the work clinicians already do. Your AI should live where your teams already work:

• In the helpdesk tool (ticket sidebar risk flag)
• In the order management system (pick/pack “hold for verification”)
• In the CRM (task list for proactive outreach)

A good workflow design includes:

• A reason code (“high risk because: repeated delivery exceptions + negative sentiment”)
• A recommended action (“send address confirmation + offer pickup point”)
• A timer (act within 2 hours, 24 hours, etc.)

If the output is just a number, your agents will ignore it.

“People also ask” (and what I’d answer)

Does AI risk scoring require big data?

No. It requires consistent data and a clear decision. Many SA businesses get value from a few thousand historical orders or tickets if labels are trustworthy (returns, churn, escalations).

Should we buy a tool or build in-house?

Buy when the workflow is standard; build when your edge is operational. If your returns problem is driven by unique catalog quirks, local delivery realities, or complex support categories, custom work can outperform generic tooling.

How do we prove ROI?

Tie the model to one measurable outcome and track it weekly. Examples: return rate reduction, fewer repeat contacts, lower refund costs, improved on-time delivery, churn reduction.

The South African angle: local constraints create better AI

South Africa’s e-commerce and digital services environment forces practicality. Load shedding, courier variability, address complexity, WhatsApp-first communication, and uneven data quality mean you can’t rely on “perfect pipelines.”

That’s why the hospital example matters. The discipline of building AI that works under operational constraints—something Tafirenyika sharpened during his time working in South Africa—maps directly to what local online businesses need.

If you’re building AI for customer engagement in South Africa, optimise for:

• Clarity over complexity (teams need to trust the output)
• Privacy and minimal data movement (POPIA reality)
• Workflow fit (tools should reduce clicks, not add them)
• Actionability (every score should have a next step)

What to do next if you want AI that actually ships

Start with a single outcome and treat it like an operational product, not a science project. Choose one “readmission-style” metric (returns, churn, escalations), pull in the text data you’ve been ignoring, and build a workflow that tells your team what to do—then measure whether it worked.

This post sits in our series on how AI is powering e-commerce and digital services in South Africa, and the theme is consistent: the winners aren’t the companies with the fanciest models. They’re the ones that turn data into timely follow-up.

If hospitals can use AI to spot risk hidden in messy notes and act before a patient ends up back in a ward, online retailers can do the same before a customer ends up filing a refund, posting a one-star review, or leaving for a competitor. What “30-day return” are you still treating as unavoidable?