Case Study - Replacing Paper Forms with a Conversational Government

Overview

When a county needs to verify that trees have actually been planted, or a ministry wants to know what's happening with a project in the field, the answer usually comes back as a paper form — eventually. Field reporting in government runs on the wrong stack: paper forms get hand-transcribed; narrative reports arrive in inconsistent shapes; structured questions get answered in unstructured paragraphs; and by the time the data reaches an analyst, the moment it could have informed has passed. Meanwhile the channel that the field workforce actually uses several times a day — WhatsApp — has no first-class place in the workflow. Ngao was built to turn that familiarity into an operational advantage for public-sector data collection.

Ngao has four moving parts that together close the loop from a citizen's WhatsApp message to a signed-off report on an executive's desk. An admin research dashboard lets administrators design research pipelines, build the schema of what they want to know, define an organisational chart that determines who reports up to whom, and import contacts from spreadsheets — with unauthorised senders blocked at the gateway. A community-facing WhatsApp agent prompts users per research goal, handles text and voice notes, applies content filters, and routes drafted reports up the accountability chain. A backend data-collection agent decides which incoming messages are relevant to which research goal, fits responses to the schema, and orchestrates the WhatsApp agent to come back with the right follow-up question — with computations on the data running through deterministic code rather than the model, so numbers can be trusted. A research agent runs structured retrieval over accumulated data to generate reports on demand and surfaces patterns across pipelines.

Research schemas are defined once and enforced end-to-end, which guarantees that whatever the AI produces fits the structure the administrator asked for. Where a research goal requires arithmetic — totals, survival rates, supplier breakdowns — the system runs deterministic calculations rather than asking the model to do the math, because hallucinated numbers in a government dashboard are worse than no numbers at all. Messaging is invite-only, the agent is multimodal (text, audio transcribed automatically, images, media), and translation sits as its own layer on both the incoming and outgoing path. The first pilots covered tree-planting verification (count, geotag, photo evidence), citizen baraza feedback (open-ended post-meeting channels), and ministry project reporting — each one a deliberately different shape of data so the platform's flexibility could be stress-tested under real field conditions.

What we did