The voice assistant does not know whether a newly announced purchasing rule applies, so it stops instead of improvising: a practical comparison of constrained selection, generative answers and human escalation

The assistant should stop, say that the new purchasing rule is not covered by its reviewed information, and send the question to a named human reviewer. That response protects the farmer from a confident guess while creating a clear route to a verified answer.

Why a new purchasing rule creates a hard boundary

A farmer may ask whether a newly announced rule changes who can buy cocoa, what documents are required, or when payment should arrive. The wording might sound close to an existing question in the assistant’s content library, but “close” is not enough.

This matters as young cocoa farmers demand clarity ahead of the 2026/2027 purchasing season. A voice assistant cannot infer the scope of an announcement from an older policy answer. A wrong interpretation could affect when a farmer sells, which buyer they accept, or what paperwork they prepare.

Before answering, the system needs three things:

• The exact rule or announcement from an approved source.
• A reviewed explanation of who it applies to and when it takes effect.
• A person responsible for questions the approved explanation does not cover.

Without those prerequisites, the safe response is brief: “I cannot confirm whether this new rule applies to your situation. I will send your question to the extension officer.”

Constrained selection keeps answers inside reviewed boundaries

Constrained selection gives the assistant a fixed library of reviewed answers. The language model listens to the question, identifies the relevant content block, and returns that block by its ID. It does not write purchasing guidance from scratch.

Suppose the library contains approved explanations for standard purchasing documents, weighing procedures, and payment questions. A farmer then asks about a rule announced that morning. The selector may find related blocks, but none should claim to interpret the new rule. The correct outcome is “no approved match” followed by escalation.

This approach offers strong control. Reviewers can see every answer that might reach a farmer, correct translations before use, and withhold uncertain material. It also limits coverage. New policies require a person to verify the source, update the relevant block, review the spoken Asante Twi, and approve the change.

For information that can affect income or contractual decisions, that delay is often the responsible tradeoff.

Generative answers cover more questions but introduce more risk

A generative system can compose a fresh answer from a prompt, a model’s prior knowledge, retrieved documents, or all three. It may explain an unfamiliar rule in fluent language within seconds.

Fluency can hide uncertainty. The model might combine an old purchasing process with the newly announced rule, assume the announcement applies nationwide, or omit an exception. A natural voice can make that answer sound more authoritative than the evidence allows.

Document retrieval improves the situation only when the source is current, authentic, complete, and correctly interpreted. A retrieved announcement may still leave practical questions unanswered. Does the rule begin immediately? Does it cover every district? Does it apply to existing arrangements? Who resolves a conflict at the buying point?

Generative answers can help staff draft an explanation for review. They should not become the final authority when the source leaves room for interpretation.

The same principle appears in Kwaku’s conflicting purchasing answers: verification has to happen before a spoken answer influences a decision.

Human escalation needs an owner and a response path

“Ask a person” sounds safe, but it fails when nobody owns the queue. The assistant must send the question to a named extension officer or designated partner representative, preserve the farmer’s wording, and record when the request was received.

A useful escalation contains:

• The original voice message or an approved transcript.
• The content blocks considered and why none was sufficient.
• The specific uncertainty, such as the effective date or affected buyer category.
• A target response time and the responsible reviewer.

For AgriVoice’s planned cocoa-farmer pilot, the operating target is a median human response time below one working day. That target does not make urgent uncertainty disappear, but it gives the farmer a realistic expectation and makes neglected questions visible.

Human escalation also produces valuable evidence. If ten farmers ask the same question, the institution can prepare one reviewed answer, translate it into spoken Asante Twi, and add it to the constrained library. The queue becomes a record of missing information rather than a place where difficult questions vanish. The escalation queue guide explains why named responsibility matters.

Choose the response pattern by consequence and evidence

Use constrained selection when incorrect advice could affect money, safety, legal obligations, or access to services. It works best when the answer set is limited, reviewed content exists, and updates can be approved quickly.

Use generative answers for lower-consequence explanation, drafting, or navigation where uncertainty can be stated clearly and users can verify the result. Current documents and source citations remain essential.

Use human escalation when the question depends on a new announcement, local interpretation, missing facts, or an exception the approved content does not cover. Escalation should also trigger when speech recognition is uncertain or two content blocks conflict.

A practical system can combine all three. Constrained selection handles reviewed questions. Generation helps an authorised reviewer prepare a response. A human decides what the new rule means before that explanation enters the answer library.

Run one purchasing-rule test before farmer exposure

Create a test question based only on the newly announced rule, then run it through the complete voice workflow. Confirm that the assistant avoids selecting an older, merely similar answer; states that it cannot verify applicability; sends the case to the named officer; and records the response time.

Next, add a deliberately ambiguous version and an ASR-corrupted transcript. If either produces a confident answer, keep the workflow away from farmers until the selection and escalation rules are repaired.