Safety Testing for Farm Advisory Chatbots

• CTO, Agritech Company
• Head of Agronomy
• Product Manager, Farmer Advisory Bot
• Programme Lead, Digital Agriculture Initiative

Agritech companies and public programmes are putting LLM chatbots and voice bots in front of farmers to answer crop questions. A fluent but wrong answer about a pesticide, a dose or a disease can damage a crop and endanger health, and farmers are likely to trust an answer that sounds confident in their own language.

We build and run a safety test programme for your farm advisory bot before launch and after every major change. With your agronomists, we define safe behaviour: which topics must be referred, which products and doses are allowed, and how uncertainty should be expressed. We then create multi-turn test conversations in English, Hindi and your regional languages, covering pesticide and dose questions, look-alike symptoms, out-of-season and out-of-region crops, scheme and loan queries, attempts to make the bot ignore its rules, and requests for other farmers' data. Answers are scored automatically against the rules, then agronomists review every risky or failed case. You receive PASS, RISK and FAIL findings with transcripts and specific fixes.

• Safety policy written with your agronomists: referral topics, allowed products and dose rules
• Multi-turn test conversations in English, Hindi and regional languages
• Checks for invented doses, unregistered products and unsafe mixing advice
• Diagnosis behaviour tests with vague descriptions and poor-quality photos
• Rule-bypass, misinformation and data-leakage test sets
• PASS, RISK and FAIL findings with transcripts, root causes and fixes

1. 01Agree the safety policyWe work with your agronomists to define referral topics, allowed products, dose rules and how uncertainty must be stated.
2. 02Build the test setsWe create multi-turn conversations and photo cases across crops, regions, languages and risk categories.
3. 03Run the testsTest conversations run against your bot through its API, a WhatsApp test number or a voice test line.
4. 04Score automaticallyRule checks and an evaluator model score each answer and mark anything risky or failed.
5. 05Agronomist reviewAgronomists review every RISK and FAIL case and confirm the finding and its severity.
6. 06Report and retestYou receive findings with fixes, and we rerun the suite after your changes to confirm they hold.

1. Policy and test design
   • Safety policy and referral rules
   • Approved product and dose list
   • Multilingual test case library
2. Test execution
   • API, WhatsApp and voice test runners
   • Multi-turn conversation driver
   • Photo test cases
3. Scoring
   • Rule-based dose and product checks
   • Evaluator model with agronomist-calibrated rubric
   • Cross-language consistency comparison
4. Review
   • Agronomist review screen for risky and failed cases
   • Severity assignment
5. Reporting
   • Findings report with transcripts
   • Regression suite for retests

Agronomists confirm every RISK and FAIL finding and set its severity, because only they can judge whether a crop answer is unsafe. Your product team decides which fixes to make and when to release.

• Your chatbot API or LLM gateway
• WhatsApp Business API test number
• Voice bot test line
• Issue tracker such as Jira or Linear
• CI pipeline for regression runs

• Testing only in a test environment or on test numbers, never with real farmers
• No real farmer data in test prompts
• Findings shared only with named contacts
• Severity confirmed by agronomists, not by the evaluator model alone
• Regression suite versioned with each bot release

• Test categories map to OWASP Top 10 for LLM Applications risks such as prompt injection, sensitive information disclosure and misinformation.
• Findings help show that pesticide guidance stays within uses registered under the Insecticides Act 1968, as defined by your agronomists.

• Agriculture & Agritech

  Crop Disease Detection from Farmer Photos

  Suggest likely diseases, pests and nutrient problems from farmer photos sent on WhatsApp or an app, reply in the farmer's language, and route uncertain or dose questions to an agronomist.

• Agriculture & Agritech

  Farmer Advisory Voice Helpline in Hindi and Regional Languages

  A voice helpline that answers farmers' routine questions on sowing, irrigation, weather and schemes in Hindi and regional languages, and hands crop-protection and dose questions to agronomists.

• Healthcare & Hospitals

  Safety and Patient-Data Leakage Testing for Healthcare Chatbots

  Test patient-facing chatbots and voice bots for unsafe medical advice, missed emergency escalation, invented information and cross-patient data leakage, with clinician-reviewed evidence.

• Telecom & Contact Centres

  Voice Agent Red-Teaming and Jailbreak Testing

  Adversarial phone calls against your voice agent in Hindi, Hinglish and English, testing verification bypass, data leakage, prompt injection and off-script promises, with recorded evidence.