Cases & Technical Insights

CASE STUDY · SPECIALIST TRAVEL AGENCY

Personalized Customer Proposals in Record Time

Specialized travel advisors spend hours in deep consultations uncovering intricate vacation itineraries and customer constraints. In a May 2026 pilot we measured the time on a proposal in the agency’s own CRM, from the consultant opening the proposal to closing it: about 12 minutes with the solution against about 40 without.

Engineering Highlights

1. Cleaned transcription: In a test on two hours of audio from six scripts, the cleaned transcription had about 50% lower Word Error Rate (WER) than the supplier’s standard transcription.

2. Automated cleaning and entity extraction: Raw transcripts are polluted with conversational filler and background noise. Our pipeline scrubs conversational noise and extracts normalized parameters: destinations, budget constraints, travel weeks, children’s ages, and specific accessibility requirements.

3. Email draft in the voice of the consultant and the agency: The system writes a structured proposal draft in both the consultant’s and the agency’s voice, so the consultant can review and edit it before it is sent.

4. QA check against the customer call: The system holds the prepared material up against the customer call and proposes mitigations that resolve potential discrepancies between the two, before the consultant signs off.

How the figures were measured

What

Time spent on a proposal in the agency’s own CRM, from the consultant opening the proposal to closing it (timestamps).

Period

A pilot in May 2026.

Sample

5 consultants, 5 proposals each with the solution (25), against 5 proposals each from the same consultants without AI (25). The proposals were randomly selected.

Result

About 12 minutes against about 40. The 40 is the BizOps colleague’s figure; the measured baseline matched it within 13 seconds. The saving comes from the research phase, where an AI research assistant that knows the agency’s specialisation and the destination country gives real-time destination and experience insight.

Transcription

Word Error Rate about 50 percent lower, measured on two hours of audio from six scripts read by different consultants: the supplier’s standard transcription against Caliper’s cleaned transcription.

Check

The transcription passes three gates in n8n. Only the essential fields are sent as JSON by REST call, with no personal data. A pass or fail comes back in about 12 seconds on average (it depends on the transcript length and the summary quality). The summary is scored with a BANT scorecard.

Applies to

One Danish specialist travel agency offering bespoke trips to Asia or Latin America. The sample is small, and this is a measured result, not a promise about your business.