Improving a voice assistant with reviewed prompt candidates

A voice assistant does not become better merely because it handles more calls. Improvement requires an identified failure, a proposed change and evidence that the change helps without creating a new problem.

Burki's learning services create suggestions and draft prompt candidates. Candidate generation is not automatic production approval. Availability and usage eligibility depend on the enabled configuration; do not infer that every call triggers optimization.

Start with one observed failure

For example, a repair assistant asks for a callback time twice after the caller has already supplied it. Preserve the selected version and the relevant conversation evidence. Determine whether the problem came from conflicting instructions, missing context, a workflow transition or an unavailable action. A prompt change is useful only if the prompt caused the issue.

Propose a narrow change

Write the intended behavior and the exact instruction change. Keep business facts and permissions separate from conversational style. Do not allow a customer statement to become a new business policy merely because it appeared in a transcript.

A candidate should explain which failure it addresses and what evidence supports it. Review generated text for new promises, deleted safeguards and unnecessary expansion of scope.

Evaluate against successes as well as failures

Use the failing scenario, a nearby correction case and scenarios the current version already handles correctly. Keep evaluation inputs separate from examples used to propose the change. Record automated assessments as assessments, not customer feedback or proof of caller-heard audio.

For a voice behavior change, run the relevant audio acceptance separately. A text test cannot show that an interruption worked or a telephone handoff retained the caller.

Approve and observe

Burki's prompt promotion flow includes human approval. Confirm the version selected for the actual runtime before relying on a release. A learning-service status alone does not establish that all active voice sessions use the candidate.

After publishing an accepted change, inspect the targeted failure and adjacent outcomes. Keep a known previous version and a recovery plan. Do not claim a rising conversion rate or shrinking call duration was caused by the prompt without checking other changes.

See the learning release checklist for the evidence to keep with a change.