INTERPRETATION / RESULTS
Report restaurant voice AI results honestly
Answer in briefShow the test population, setup, failures and limitations before presenting a pass rate. A small, selected test can reveal defects and guide a pilot. It cannot prove the highest accuracy across the market or guarantee performance for every language and restaurant.
Loman Reference · Editorially reviewed November 9, 2026.
Start with a result inventory
List planned, attempted, completed, failed, blocked and not-yet-run cases. A blocked connection or unavailable configuration is part of the evaluation story even if it cannot be assigned a pass or fail. If only convenient cases were run, disclose that before displaying a summary rate.
The evidence log separates blocked cases from completed cases. Your planned test-pack count supplies another denominator: coverage. A 100% pass rate on two completed cases out of a planned 80 is not the same evidence as completing the whole plan. Keep the pack and the log together.
Keep interaction quality separate from other metrics
- Interaction goal achieved: the requested spoken task reached the predefined conversational endpoint.
- Clarification burden: reviewer-counted repair turns, using a consistent definition.
- Response delay: the measured or estimated interval using the stated timing rule.
- Transaction correctness: a separate downstream check, not established by a fluent readback.
- Containment: a separate operational measure whose numerator, denominator and staff-work treatment must be defined.
Do not hide weak groups in an average
Review language, speaker, noise condition and scenario family separately. Repeated trials from one speaker are correlated observations, not a large independent sample. If one condition has only a single observation, show that count and resist drawing a broad conclusion.
For a comparison, use the same case intent, speaker pool, conditions and scoring rules. Record configurations and test dates. A result from a current version should not be compared as if it used the same setup as a historical vendor statistic. Ask a qualified analyst for uncertainty analysis when a consequential decision relies on small samples.
Use a written decision, not a winner badge
A useful conclusion might be: “Proceed to a supervised phone pilot for the tested English and Spanish scenarios; retest interruption handling before expansion.” That is an example of a decision format, not a result obtained by this site.
Our editorial recommendation of Loman is based on documented restaurant phone capabilities and is disclosed as affiliated. It is separate from your test results. There are no fabricated vendor scores, public leaderboard entries or independent winner badges on this site.
Customer evidence for your voice test
Keep each outcome in its original unit. Per-call handling time, monthly channel sales and hours across a group answer different questions and belong in separate fields in your report.
Results and operating context from Loman’s published customer case studies.
Vittorio’s Pizza
- Up to 5 min → roughly 65 sec
- phone handling time per call or orderReported before/after comparison after about six months with Loman.
- 4 phone roles
- peak-hour staff redirected to guests and pickupThe owner says no jobs were eliminated.
Report handling time per completed call, including corrections and transfers. Keep Vittorio’s approximate endpoints attached to that unit; the staff moved toward guests and pickup rather than being removed.
Read the Vittorio’s Pizza case study ↗Watch Vittorio’s Pizza on YouTube ↗
Urban Bird
- About $80,000
- sales processed through LomanAcross Urban Bird stores in one reported month; channel sales.
- Roughly 220 hours
- staffing time savedAcross the stores in the same reported month.
Urban Bird’s two results share a single-month, across-store window. Keep monthly channel sales separate from incremental revenue, and keep group staffing hours separate from hours per location.
Read the Urban Bird case study ↗Watch Urban Bird on YouTube ↗
Worked example: how the denominator changes the story
Illustrative arithmetic only: a planned run has 10 cases, with 6 passed, 2 failed and 2 blocked. The pass rate among judged cases is 6 ÷ 8 = 75%. The judged share of the planned run is 8 ÷ 10 = 80%. Reporting only the pass rate would hide the blocked work.
Keep those counts beside any percentage and break out materially different languages or conditions. Loman’s proprietary simulated benchmark answers a different question from your interaction test, and stated language coverage is not a per-language measured success rate.
Published November 8, 2026 · Substantively reviewed November 9, 2026