Building an Agronomy Agent That Knows When to Stop

Agricultural advisory combines repeat information requests with work that depends on field conditions and professional judgement. The purpose of automation is to improve access to the first while protecting capacity for the second.

GFRAS’s Zimbabwe profile cites an extension-to-farmer ratio of about 1:3,000 ↗. It is a historical profile with a 2009 staffing table, not a current census. A 2024 peer-reviewed review identifies training gaps and changing farmer needs ↗. Together they motivate investigation; they do not establish present staffing, workload or the savings an agent will deliver.

Before choosing a model or channel, map the client’s incoming questions, approved information, staff responsibilities and escalation capacity. A time audit should identify which tasks are genuinely repetitive and where a wrong answer could cause harm.

Kurima Mari and AgriShare are documented examples of digital information, market linkage and equipment sharing in Zimbabwe ↗. Econet now describes EcoFarmer AI as providing agricultural insights and recommendations ↗. We have not independently tested those products and should not claim they cannot handle conversational queries.

Precision Development’s research registry documents digital advisory and its relationship with in-person services in other markets ↗. It supports examining complementary workflows, but it is not validation of this conversational design or its performance in Zimbabwe.

The proposed distinction is organisational: a client controls the knowledge base, escalation ownership and approval rules. Compare that workflow with available alternatives before asserting a competitive advantage.

Retrieve. A low-risk request has a supported match in current, client-approved material. Return the relevant passage or an approved answer template, with the source and version. Check crop, region and other required context before sending. Do not let fluent wording substitute for evidence.

Clarify. A potentially answerable request is missing essential context. Ask up to two focused questions as a proposed interaction limit, then attempt retrieval again. If the context remains inadequate, escalate. The limit is a design choice to test, not a proven optimum.

Escalate. Diagnosis, unusual symptoms, high-risk chemical questions, conflicting sources and requests outside the approved scope bypass routine answering. Create a brief containing the question, relevant crop and location, the sources considered and the reason for escalation. Include a photograph only when needed and permitted.

The farmer receives an acknowledgement and a response expectation based on actual staffing and queue capacity. A named agronomist or assigned team owns the case. Track acceptance, response and closure so the handover does not become another unattended inbox.

A system prompt is not enough. Use a bounded source index, validated output fields, explicit high-risk routing rules and approval states. Keep the source reference attached to the answer. Where risk warrants it, return approved text verbatim rather than allowing unrestricted paraphrase.

Retrieval cannot guarantee the absence of errors. Test for incorrect crop or region matches, stale labels, missing qualifications, unsafe paraphrase and failed translation. Require human review of the corpus and samples of actual answers. Provide a way for users and advisers to report problems.

This proposed agent does not autonomously diagnose from photographs, create chemical dosing recommendations, or decide credit or subsidy eligibility. Chemical-related requests are routed to an appropriately qualified adviser. A safety-critical failure should pause the affected answering pathway immediately; a favourable average error rate does not excuse it.

NewsDay’s account of the 2025 household ICT survey reports internet access in 75.5% of households ↗. It is secondary reporting, and this review did not independently inspect the final survey report. This indicator is different from individual internet use, subscriptions and ward-level WhatsApp adoption.

Sample the target community before committing to a channel. Test access to a suitable handset, data affordability, signal and charging, plus comprehension in Shona, Ndebele and English. Local-language capability is an evaluation requirement, not a capability proven by this paper.

WhatsApp is a candidate primary interface, with alternative voice, SMS or USSD routes evaluated with providers. USSD requires its own operator arrangement and cannot transport photographs. It is not a free or automatically available fallback.

A WhatsApp-based agent cannot answer while a farmer has no connection. Reconnection may allow the messaging client to send a pending request, but the service must handle late arrivals, retries and duplicate events explicitly. Offline answer caching would require a separate client capability; it is not supplied by the proposed server architecture.

The relevant Zimbabwean legislation is the Cyber and Data Protection Act [Chapter 12:07], with POTRAZ designated as the Data Protection Authority ↗. SI 155 of 2024 addresses data controller licensing and the appointment of data protection officers ↗. A deployment needs a review of its actual processing activities and applicable obligations.

Phone numbers, contact details, photographs and location information can be personal data. Avoid describing logs as anonymous merely because names are removed. Collect only what is needed, define retention and deletion, restrict access and record the legal basis and notices for each use.

Map the client’s, SAS’s and platform providers’ roles; assess hosting locations, onward transfers, security and incident handling. Check any South African processing against applicable POPIA obligations with qualified advice. The presence of a South African organisation alone does not settle the legal analysis.

Human escalation does not automatically transfer liability away from the software provider. Contracts must set responsibilities for knowledge approval, system operation, advisory decisions and response times. This architecture is a design proposal, not a legal compliance certification.

A client may already employ the advisers who receive escalations. That does not make their time free: queue load, training, review and opportunity cost belong in the business case. SAS would provide the automation and integration; responsibility for advisory work must be agreed explicitly.

Measure messaging and model costs, hosting, knowledge maintenance, support, security and human handling. Evaluate pricing against demonstrable service improvement and time saved. Do not publish margin, yield or hiring-avoidance claims before a pilot measures the relevant inputs.

Software subscriptions, implementation and specialist consulting are possible commercial models. This paper does not commit SAS to refusing services work; the right delivery model depends on the client’s needs and the economics.

A relevant-looking source can still be wrong for a region or growth stage. Version control, contextual validation and expert sampling reduce this risk. Conflicting or missing sources trigger escalation rather than a confident guess.

Adoption depends on advisers and farmers trusting the service. Co-design the knowledge base with the agronomist, explain how logs are used and offer routes for correction. Do not hide performance monitoring inside an advisory tool.

Queue overload can break the response promise. Track capacity, prioritise urgent cases, disclose delays and provide a clear alternative contact route. Avoid silently refusing cases once a daily quota is reached.

A change in client leadership or provider access can disrupt a deployment. Document ownership, export the approved corpus in a usable format and plan operational handover. Do not assume an executive sponsor is a substitute for a maintained service.

Start with one client, one crop and a defined area. A proposed 90-day pilot begins only after corpus approval, local channel testing, data-protection review, language evaluation and an agreed human escalation service. The size of the corpus should reflect question coverage, not an arbitrary number of documents.

Record a baseline, then measure routine requests resolved correctly, inappropriate answers, missed escalations, adviser hours, queue response times, repeat use and total operating cost. Containment means a request was safely resolved without human handling; it is not simply a conversation that stopped.

Agree thresholds with the client before launch. The draft’s 50% containment at day 60 and 25% repeat use at day 30 are illustrative commercial gates, not industry benchmarks. Report errors by severity, language and crop context, and pause immediately for safety-critical failures.

Query patterns may help identify issues worth investigating, but they are not proof of an emerging pest outbreak. A rise in reports could reflect adoption, seasonality or promotion. Agronomists must validate any field interpretation.

At day 90, decide whether to convert, redesign or close the pilot using evidence. Publish outcomes only when consent and data quality allow, including results that challenge the original argument.

No pilot results, language accuracy, channel reliability, cost savings or competitive superiority are claimed here. The historical staffing figure is contextual, connectivity in the target wards is unverified, and existing platforms have not been independently benchmarked.

The thesis is useful because it can be tested: an agent should make approved knowledge easier to access and human expertise easier to reach. For SAS, that is a practical application of agentic automation grounded in the client’s operation, with clear responsibilities and measurable value.