evidoria

← Back to browse

Good practice Imported

The Policy Innovation Lab's WhatsApp AI Chatbot for Citizen-Generated Service Delivery Data

South Africa · Stellenbosch · See the South Africa profile · See the Stellenbosch profile

Evidence: Descriptive / self-reported Top 83% 40/100 · Ask Evidence Copilot about this practice

Stellenbosch University's Policy Innovation Lab built a WhatsApp chatbot using ChatGPT so citizens can report service failures like broken water pipes, mapping de-identified reports by district for officials.

The Policy Innovation Lab's WhatsApp AI Chatbot for Citizen-Generated Service Delivery Data

Details

Maturity
Pilot
Promoter
The Policy Innovation Lab, Stellenbosch University
Period
2024–ongoing
Keywords
civic tech, service delivery, data science, WhatsApp chatbot

Context

The Policy Innovation Lab at Stellenbosch University, which works with South Africa's Presidency on responsible AI for policymaking, developed a WhatsApp-based chatbot that lets citizens report public service delivery failures — such as broken water pipes — in natural language.

Activities

The tool uses OpenAI's ChatGPT as its underlying language model. After a citizen reports an issue, the system runs a de-identification process that strips phone numbers and coarsens time and location details before the report is analysed. Topic analysis then generates an interactive map letting officials sort reports and compare service-delivery problems across districts.

Results

OpenUp, an independent South African civic-technology organisation, applied its own risks-and-harms framework and flagged concerns the Lab documented openly: the underlying model can reflect and amplify inequalities affecting under-represented-language speakers, generative models can produce confident but false ('hallucinated') answers, routing citizen data through WhatsApp and a third-party model leaves government with limited control over data storage/processing, and the model's decision logic is a hard-to-audit 'black box'. A proportionality question was also raised about environmental costs versus benefits.

Conclusions

No user-count, report-volume or service-resolution figures have been published, so the case is best read as a carefully self-audited pilot rather than a proven-at-scale service.

Implementation
Indicative cost
Low (< €50k) — Not disclosed; built on a third-party LLM API (ChatGPT) rather than bespoke model infrastructure, suggesting a low-cost pilot.
Time to results
Short (< 1 year) — Ongoing since 2024, described as a pilot with no reported user, volume or resolution data as of the source's publication.
Staffing & skills
The Policy Innovation Lab, Stellenbosch University (developer), Partnership with South Africa's Presidency on responsible AI

Conditions for success

  • Independent third-party risk audit (OpenUp) applied before/alongside deployment
  • De-identification pipeline stripping phone numbers and coarsening location/time data
  • Low-barrier citizen channel (WhatsApp) already in wide use

Common failure modes

  • Reliance on a third-party commercial LLM (ChatGPT) reduces government control over how citizen data is stored and processed
  • Underlying model can amplify inequalities for under-represented-language speakers
  • Generative hallucination risk in citizen-facing responses
  • Decision logic remains a 'black box' that is hard to hold accountable
  • No government department has taken confirmed formal ownership or ongoing funding of the tool

Where it fits

Governance type
university-government partnership
Scale
local pilot
Income level
upper-middle-income

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful