Skip to content
AI & Technology EngineeringTrue Access2026

Building an AI System That Answered Questions at 2am So the Team Didn't Have To

A three-person team was spending 60% of their capacity answering the same 12 questions about SASSA, COIDA, and disability rights. We built an AI assistant that handles 91% of enquiries — without human escalation.

91%

enquiries handled without human escalation

47 sec

average response time (was 2.3 days)

3 hrs/day

caseworker time freed per person

500+

individuals assisted in first 90 days

The Challenge

True Access supports over 500 individuals and families navigating South Africa's disability grant system — SASSA applications, COIDA claims, school placement appeals, assistive device applications. The processes are complex, the official language is bureaucratic, and the consequences of getting it wrong are serious. Their small team was handling enquiries via WhatsApp at all hours. The same 12 questions were coming in constantly. The team was burning out on work that was important but repetitive.

Six Weeks of Knowledge Extraction Before a Line of Code

We read every SASSA circular and grant guideline we could find. Every COIDA documentation requirement. Every school placement appeals procedure for the Eastern Cape. But more importantly, we spent six weeks interviewing the True Access team — because the real answers to the real questions are almost never what the official documents say.

The knowledge base we built has 847 indexed items across six categories. It captures not just what the official process is, but what it actually looks like on the ground: which offices in which districts process applications fastest, which supporting documents are technically optional but practically required, which appeal procedures have real teeth and which are bureaucratic dead ends.

The WhatsApp AI Assistant

We built the assistant on Claude Haiku, integrated via the WhatsApp Business API. Every conversation starts with a triage menu — four options that route to the right knowledge cluster immediately. The assistant speaks plainly, follows up when answers need clarification, and knows when to say "this needs a human to look at it."

That last part matters most. The assistant handles 91% of enquiries autonomously. When a query falls outside its scope — a complex case, an appeal that needs documentation review, anything that requires judgment the system shouldn't make — it flags for human follow-up and creates a case record in the intake system automatically.

Integration Without Disruption

True Access didn't want to learn new software. The team has three people and no capacity for a system migration. We integrated the case creation workflow directly into their existing Google Sheets case management system via n8n webhook. New cases from WhatsApp conversations appear automatically in the right tab, pre-filled with name, query type, date, conversation summary, and escalation reason.

The team reviews escalations in the same spreadsheet they've used for three years. Nothing changed except what arrives in it — and what doesn't need to arrive at all. In the first 90 days after deployment, 500+ individuals were assisted. Average response time went from 2.3 days to 47 seconds.

People were messaging us at 2am because they were scared and they needed answers. Now those answers exist, instantly, for anyone who needs them.

Director, True Access

Disability Rights Advocate

Start a Project

Ready to build something like this?

Tell us about your project. We'll scope it, price it, and propose an approach — within 24 hours.

True Access — Mirembe Muse Case Study | Mirembe Muse