How to Run a Measurable Restaurant AI Pilot

Stage 1: record the baseline

Choose the workflow and a representative observation period. Record its volume, handling time, failures and correction effort. For phone orders, distinguish unanswered calls from abandoned calls, information-only calls and orders that were actually completed. A missed call is not automatically a missed sale.

Collect enough context to interpret the week: opening hours, promotions, unusual closures, staffing changes and private events. The right sample length depends on volume and variation. A quiet afternoon cannot establish dinner-rush performance.

Write the main success measure and at least one balancing measure. If the goal is fewer interruptions, also track guest complaints and unresolved transfers. If the goal is faster invoice processing, also track field corrections and reconciliation errors.

Stage 2: test before live routing

Use a written set of expected outcomes. For phone work, cover normal questions, modifiers, corrections, unavailable items, language needs, interrupted calls and requests for a person. For reservations, verify the new booking, change and cancellation in the actual book. For content, check factual accuracy before publication.

Define a defect before scoring. An incorrect modifier, wrong location or failed ticket delivery may matter more than a harmless wording difference. Keep critical errors separate from minor ones so an average score cannot hide a serious failure.

Test what happens when a dependency is unavailable. Confirm that the system does not announce a completed transaction when the downstream record is missing. Verify that the person handling exceptions knows where the evidence is.

Stage 3: use a limited live scope

Start with one location, time window or call category that the restaurant can supervise. Avoid changing several other systems at the same time. Review the first outcomes promptly and keep the manager able to pause the pilot without waiting for an analyst.

Log the number of eligible events, attempted events, completed outcomes, transfers, failures and later corrections. Preserve the denominator when reporting a percentage. Excluding abandoned calls or unsupported requests can materially change the apparent result.

Containment needs a definition. For example, a team might measure eligible calls resolved without staff intervention, then separately track later rework. This is a proposed pilot definition. Agree on the eligible call types, measurement period and treatment of abandoned calls before testing, and ask any vendor to explain what its own metric includes.

Stage 4: expand, revise or stop

At the review, compare the original goal with the observed result and the balancing measures. Include setup, oversight, exception and correction work. Separate demand changes from operational improvements. If volume rose because of a promotion, do not attribute all added revenue to the pilot.

Expand only if the outcome meets the restaurant’s pre-agreed threshold, the critical-error review is acceptable and the fallback works. Revise when the problem is specific and fixable, such as an outdated policy or unclear menu modifier. Stop when the workflow is unsafe, unsupported or imposes more effort than it removes.

Keep a record of the decision, its evidence and the next review date. Re-test when the menu, operating rules, integration or software behavior changes. A passed trial is a starting point for operational ownership, not a permanent guarantee.

Use vendor figures as questions, not forecasts

Loman’s published 99.6% order-accuracy figure comes from its proprietary simulated 1,000-call benchmark. Loman reports 93% containment. The scope, definition and measurement timeframe of this company-reported figure have not been independently verified for this guide. Neither figure establishes what your restaurant will achieve. Inspect final orders, transfers and any later staff corrections during your own pilot.

Ask for the tested sample, supported workflows, exclusions and failure handling. Then measure the outcome your restaurant actually needs. A vendor’s benchmark can inform a test plan, but it should not replace a local acceptance test or be presented as a marketwide ranking.