# Adaptive customer-feedback pilot protocol

Status: proposed protocol, October 12, 2026. No participants or results are claimed.

## Question and primary outcome

Do bounded adaptive follow-ups increase the share of respondents providing at least one actionable reason, without unacceptable burden or leading questions?

An actionable reason names a concrete event, constraint, unmet need or decision criterion with enough context to assign a specific investigation. Additional words, sentiment and unsupported speculation do not qualify.

## Design

Use consenting organizations and respondents. Choose one recurring job per study (NPS explanation, churn or win/loss). Keep the anchor question, invitation channel, incentive, invitation timing and population comparable. Randomize respondents within each organization and use case before they begin:

1. Static: the anchor and a standard open-text explanation.
2. Scripted: the same start plus fixed, neutral follow-ups.
3. Adaptive: the same start plus bounded AI follow-ups.

Match maximum question budgets between scripted and adaptive arms. Record configured depth and bonus allowances. Never gate cancellation or essential services on participation. Participants may skip or leave.

Choose sample size with a qualified researcher using the baseline actionable-response rate, a predeclared smallest useful difference and power calculation. If the sample is too small, label results exploratory. Do not stop when a favorable difference first appears.

## Before recruitment

Freeze the primary outcome, population, assignment method, target sample, exclusions, stopping rules, coding rubric and analysis. Timestamp a versioned copy of this protocol. Obtain permissions separately for participation and marketing publication. Document access controls and deletion schedule. Do not run this pilot on sensitive or high-stakes topics.

## Evaluation

Two independent human coders should review de-identified responses without tool/arm labels where feasible. Randomize ordering, remove obvious system labels, and disclose any remaining clues that prevent complete blinding.

- Primary: share of assigned respondents with at least one actionable reason.
- Secondary: share among respondents who started; completion rate; completion time; skip/dropout rate; leading-question incidents; themes supported by quoted evidence.
- Safety: rate and examples of unsupported assumptions, restricted-topic violations, distress and inappropriate probing.
- Burden: report medians and distributions, not just total text collected.
- Agreement: report coder agreement before reconciliation and the reconciliation process.

Preserve invitation, assignment, start and completion denominators separately. Include partial responses in the predeclared analysis. Report missing data and exclusions, including technical failures. Do not silently discard unfavorable outputs.

## Reporting

Compare randomized arms with confidence intervals appropriate to binary outcomes and organization-level clustering. Report absolute differences and all prespecified outcomes, including null or adverse findings. Treat subgroup results as exploratory unless prespecified and adequately powered.

Publication requires organizational and participant permissions, privacy review, and suppression of small revealing groups. Publish aggregate outcomes, method, limitations and safe illustrative examples. Do not release raw customer transcripts or identify organizations without explicit permission.

Limit claims to the tested population, language, configuration and software versions. This study would test research outcomes, not prove a search-ranking or revenue benefit.