Benchmarks & sources / Specialist product studies

Assistant Benchmark: consumer travel

Five services are observed through booking, seats, changes and cancellation. Miso includes a human travel team.

David Pawlan / Assistant Benchmark · Updated 2026-10-23 · reviewed 2026-10-29

Scope and method

Five services × eight scored dimensions; dimensions can share a trip. Independent trip/repeat count is not established. Protocol travel-v1 is Draft.

Anchored 1–10 reviewed scores: booking speed receives 40%, the other seven dimensions share 60%. Speed score starts at flight selection. Human assistance is recorded separately.

How PAEval uses it

A travel track of the existing AB family, not an additional independent evaluator. Miso remains a source-labelled travel service, not a new profiled core PA.

Evidence family: assistant-benchmark. Shared tasks and tracks are not independent evidence votes.

Published travel ratings /10

Publisher weighted overall copied unchanged; do not average it with the general AB score.

DimensionMisoInstinctGrok BotDotsMuse
Overall9.38.07.97.16.9
Onboarding107777
Booking7710710
Seats10109103
Traveler profile101010107
Cancellation779710
Rebooking1037103
Booking speed109757
Proactiveness87788

Intervention and follow-through

Observed instances; not guarantees of current support or future fulfillment.

ServiceChangesCancellation / refund follow-up
MisoCancelled and rebooked; human backupHuman team confirmed seat refund
InstinctNot completed; support needed$35 seat-refund request failed
Grok BotUser clicked Change flightUser completed steps including cancellation
DotsCompleted$60.01 credit unresolved
MuseNot completedFull refund confirmed

Limits and interpretation

  • Different trips, fares, starting profiles and intervention levels; no controlled success-rate estimate.
  • Cancellation does not establish refund settlement. General-board travel cells and this report are overlapping evidence.
  • Miso had a saved traveler profile; its human backup is part of service delivery.

Sources and reuse

PAEval summarizes existing published evidence and runs no new tests. Public readability does not establish permission to redistribute full datasets, conversations or images.

Related product records

Instinct · Grok Bot · OpenAI dots · Meta Muse

All benchmark sources → · How to read outcomes and intervention →