How should guest memory be measured?
Measure whether correctly attributed, current knowledge reaches the responsible team in time for service, and then measure the action and guest response separately. Counting stored notes answers a different question. A useful measurement system preserves the denominator, the observation window and the source of each result.
This is AFG’s proposed measurement dictionary, version 1.0. It defines evaluation measures rather than reporting benchmark results or describing GuestMemoryOS’s proprietary contribution score. Operators can apply the definitions to their existing workflow or to a proposed trial.
NIST’s voluntary AI Risk Management Framework supports considering trustworthiness throughout AI design, use and evaluation. It does not prescribe these hospitality metrics or certify this method. [1]
Define the unit before calculating a rate
A note, a guest, a memory and a service action are different units. One note may describe several facts. Several notes may repeat one request. An action may address more than one memory. Do not divide totals from unrelated categories and label the result a conversion rate.
For each measure, record the property or vessel, period, eligible population, sampling method, exclusions, missing records and reviewer. Preserve raw counts beside percentages. Use “not available” when no eligible cases exist; a zero denominator is not a 0% result. Establish inclusion rules before reviewing outcomes.
AFG’s 21 September operating snapshot reports 5,375 staff notes, 1,453 opportunities and 114 recorded service actions, among other activity. Those categories may overlap, and the public totals do not establish linked eligible events. Dividing 114 by 1,453 would not establish an action-completion rate. [2]
1. Reviewed attribution rate
Definition: sampled person-specific records with attribution verified against the agreed reference evidence, divided by all sampled person-specific records.
The reviewer needs a reference: for example, an attributed guest statement or a service record that identifies the intended individual. Separate verified correct, verified incorrect and unresolved cases. Keep unresolved cases in the headline denominator so missing evidence cannot improve the apparent result. This is a review outcome, not a universal claim of identity accuracy.
Party-level arrangements belong in a separate category. A shared airport transfer does not require inventing an individual preference for each passenger.
2. Timely handover availability
Definition: eligible memory–service-decision pairs available to the authorised receiving team before the decision deadline, divided by all eligible pairs identified under the review rules.
A pair means one relevant memory and one specified decision, such as preparing tomorrow’s breakfast. Establish eligibility through independent review of the scheduled service and current request, rather than counting only items the software chose to surface. Record late, missing and unverifiable delivery separately; unknown timing does not pass.
Availability is a delivery measure. Add a distinct receipt or comprehension check if you need to know whether the employee saw and understood the brief. The handover paper supplies that check.
3. On-time completion of assigned actions
Definition: distinct assigned actions with verified completion by their agreed deadline, divided by all assigned actions whose deadline falls in the reporting window.
Link completion evidence to a unique action. Keep suggestions outside the denominator until they are assigned. Report cancellations, duplicate assignments and justified decisions not to proceed as separate outcomes under rules agreed in advance. If an adjusted rate excludes valid cancellations, show the original denominator as well. A completion tick proves only what the verification process actually checks.
4. Guest-response coverage and response distribution
Coverage: completed actions with a linked direct guest response within the agreed follow-up window, divided by completed actions eligible for that follow-up.
Report positive, neutral, negative and unclear responses among responses actually received, alongside the no-response count. Staff interpretation belongs in a separate field. Do not convert silence into approval or use an employee’s prediction as the guest’s rating.
To limit selective recording, define eligible service types and invite feedback consistently within normal hospitality. A linked compliment helps explain an interaction; it does not establish that the software caused the compliment.
5. Timely correction propagation
Definition: eligible corrections reflected in all identified affected briefs or pending actions before their next decision deadline, divided by all eligible corrections reviewed.
Define an eligible correction as a verified change with at least one affected future service decision. Review the affected destinations and their timestamps. A correction passes only if every identified destination meets the deadline. Report time to the last update as a secondary measure, including overdue and unverified cases.
This checks the practical consequence of the memory lifecycle. Updating the central profile alone is not sufficient evidence that the kitchen or housekeeping team received the change.
6. Staff effort and repeated requests
Report capture, clarification, manager review, briefing and follow-up time separately, then total them for the same observation period. Show minutes per occupied room-night or another agreed operating unit, plus staff role and workload context. Use medians and a spread of observed times where feasible; a single average may hide difficult cases.
For repeated requests, count documented repetitions of previously recorded requests in a defined set of observed service encounters. Report the count per 100 observed encounters and describe how encounters were selected. A rise after introducing better logging may reflect improved detection, so compare collection methods as well as numbers.
A worked example with explicit denominators
A reviewer identifies 40 eligible memory–decision pairs. Of these, 30 briefs arrive before the decision, six arrive late and four have no verifiable delivery record. Timely availability is 30 ÷ 40 = 75%. The four unknown cases remain visible. This says nothing by itself about whether guests benefited.
Separately, 20 assigned actions fall due and 15 have verified on-time completion: 75%. If all 15 are eligible for follow-up and nine have a direct response, response coverage is 60%. If eight responses are positive, report “eight positive responses among nine received; six eligible actions without a response”. Do not describe that as an 89% satisfaction rate for the whole trial.
Report the population and sampling uncertainty before comparing teams. Multiple records from one stay are not independent guests. The pilot protocol addresses comparison design; the evidence review explains what AFG’s current figures support.
Sources and evidence notes
- [1] NIST. AI Risk Management Framework.
- Official overview of the voluntary framework. General AI evaluation context, not certification. Source reviewed 25 September 2026.
- [2] AFG Holdings. GuestMemoryOS live service, reviews and testing.
- Company-reported evidence. Activity window: 14 July–21 September 2026. The review-record check remains dated 21 September 2026. Source reviewed 25 September 2026.
Cite this paper
AFG Holdings Ltd. (2026, September 25). How to Measure Guest Memory Quality: Definitions and Worked Examples. AFG Guest Memory Reference Series, paper 3, version 1.0. Permanent article URL
Use the relevant section link when citing a specific definition or measure. Read the editorial and correction policy.
Extend this evidence and method.
Original analysis of documented airline service models, connected to AFG’s guest-memory evidence and operating methods.
AI needs a method.
Guest memory needs evidence.
Read how AFG’s reported experience of repeated failures during live development shaped the GuestMemoryOS approach—and why identity, context, corrections and service follow-through matter.
Why AI alone is not guest memory →