AI Crop Advisory Chatbot Pilot
The Ministry of Agriculture is piloting an AI chatbot to give crop advice to 15,000 farmers in three districts. It will answer questions on pest identification, disease control and fertilizer use in Sinhala, Tamil and English over a low-bandwidth messaging channel. The six-month pilot has LKR 12 million and support from 45 extension officers. The chatbot uses a cloud language model and stores phone numbers, location and crop details. Early tests show confident but wrong advice on pesticide dosage, and there is no agronomist review before answers are sent. Farmers sometimes cannot explain their problem in text. The vendor wants to scale nationally after three months and is asking for permission to use conversation data to improve its model. The ministry must decide what to launch, what to withhold, and what safeguards are non-negotiable.
1. Identify the highest risks in the pilot and rank the safeguards that must be in place before any farmer receives advice.
Read the worked answer
The highest risk is physical harm and legal liability from confident but wrong pesticide advice. Language misunderstanding and hallucinated disease identification are next, followed by privacy and vendor reuse of conversation data. Before any farmer receives advice, I would require: a curated, agronomist-approved knowledge base; no pesticide dosage answer without verification; confidence thresholds that trigger escalation to a human; clear statements of the chatbot's limits; a complaint and incident process; consent and data minimisation; and a contract that blocks model training on conversation data unless separately approved.
The ranking matters because a general disclaimer does not neutralise a safety risk. Advice on dosage should be withheld or routed to an extension officer until the content and review process are proven.
2. Propose a target architecture and operating model for the pilot that fits low bandwidth and 45 extension officers.
Read the worked answer
A suitable pilot architecture is a messaging front end over SMS or a low-bandwidth app, a language detector and triage layer, and a retrieval component that answers only from approved crop, pest and fertilizer content. The model should not answer outside that corpus. Low-confidence or safety-critical queries should create a case for one of the 45 extension officers, with a callback or message. Every interaction should be logged with a study ID, not unnecessary personal data.
The operating model needs a content owner, a rota of reviewing agronomists, multilingual quality checkers and a weekly review of wrong answers. Store only phone number, district and crop where needed, with retention limits. This keeps humans accountable for high-risk advice while using automation for routine questions.
3. Define evaluation criteria for deciding whether to scale after six months, including safety, accuracy, adoption and cost.
Read the worked answer
Pre-register go/no-go criteria before the pilot. Safety: zero serious incidents and full escalation of dosage queries. Accuracy: independent tests by language and crop, with an agreed minimum and no unsafe answer. Adoption: active farmers, repeat use, completion of conversations and reach across districts. Service: escalation rate and response time by extension officers. Cost: cost per useful interaction and support load. Compliance: consent, retention and vendor data-use audit. Equity: use by women, older farmers and low-literacy users.
Scale only if safety and accuracy thresholds are met and the extension-officer workload is sustainable. A national launch after three months should be refused because it would precede evidence; a phased district expansion after six months with monitoring is more defensible.