Customer Operations
A Bounded Research Method for Cannabis Dispensary Wait Times
Why arrival, service, queue, and exception clocks should not be collapsed into one customer-wait metric.
Key takeaways
- A queue metric needs defined start and end events
- Median alone hides long waits
- Accessibility and service complexity affect transfer
Research question
Which wait-time measure best describes a dispensary customer’s experience? The unit is one visit or order episode with a defined arrival and completion event. The population is all observed episodes during the named period and channel. Queue entry, identity verification, consultation, payment, and pickup may have different clocks; a single “wait” number can conceal the part that needs attention.
Queueing research, accessibility guidance, Virginia oversight, and NIST measurement principles support explicit definitions and consistent timestamps. They do not provide a universal cannabis wait-time target because layout, service model, appointment rules, and customer needs differ.
Evidence and findings
Record arrival or check-in, queue entry, service start, payment start, completion, channel, and exception status. State whether timestamps are observed, system-generated, or self-reported. Use one time zone and preserve corrections. If a customer leaves before service, treat abandonment as its own outcome rather than assigning a zero service time.
Report count, median, and a tail measure such as the 90th percentile. The mean can be distorted by rare long visits; the median can hide them. Segment by channel and visit complexity only when sample sizes support interpretation. Small samples should show raw counts and be labeled exploratory.
| Clock | Question | Risk of confusion |
|---|---|---|
| Arrival to service | How long before staff engagement? | Includes queue and check-in |
| Service to completion | How long to finish? | Depends on transaction complexity |
| Total episode | What did the customer experience? | Requires complete timestamps |
Interpretation and transfer
A longer median may reflect more complex transactions, not inefficient service. A shorter average can coexist with a severe tail. Compare like days and channels, and note staffing, promotions, system outages, and accessibility accommodations. Do not use public wait-time research to infer one store’s cause without local evidence.
Customer experience data should be minimized and separated from unnecessary identity details. Accessibility needs may change the appropriate service time; that is not a defect. A fair measure evaluates whether the process accommodates the defined population, not whether every visit has identical duration.
Designing a comparable observation window
The observation window should begin before the customer reaches a staff member. If the system starts timing at payment, the queue disappears from the measure. If a person joins a line, leaves, and returns, decide whether that is one episode with abandonment or two episodes. Define the rule before looking at results. Retain an episode identifier that is not unnecessarily tied to identity, and document whether appointments, online orders, caregivers, accessibility accommodations, and returns are included.
Arrival is often the least reliable timestamp. A door counter may count companions, a ticketing system may start only after a form is completed, and a customer survey may be submitted after memory has compressed the visit. Label each clock as observed, system-generated, or reported. Estimate the share of episodes missing each event. Complete cases can be summarized separately, but excluding incomplete episodes without reporting them creates survivorship bias. Abandonment is an outcome that can be analyzed by channel and time of day, not a zero-minute success.
Queue measures need operational context. A long service interval may reflect an age-verification question, a product consultation, a split payment, or a system outage. Those are not interchangeable delays. Add a small set of mutually exclusive reason codes and permit an unknown value. Overly detailed codes encourage inconsistent labeling; overly broad codes make the result unhelpful. The goal is to explain which process segment changed, not to assign blame to the person serving the customer.
Summaries should expose the shape of the data. Report episode count, median, upper percentile or maximum, and the number still open. The 90th percentile is unstable in a very small sample, so show the raw distribution or use a clearly labeled exploratory summary. Means are useful when the cost of every minute is the question, but they should be accompanied by tail information. A store can improve its median while leaving a small group with extreme waits untouched.
Comparisons need matching strata. Compare weekdays with weekdays, appointment visits with appointment visits, and similar product or service complexity. Record staffing, promotional activity, weather disruption, software outages, and local events. If a redesigned queue produces a shorter wait, ask whether arrival volume, staffing, or the customer mix also changed. Without a concurrent comparison or stable repeated periods, describe the result as an association rather than an effect of the redesign.
Accessibility changes the interpretation of “efficient.” A customer who needs communication support, a seating break, an alternative format, or a different route may require a different service sequence. That is not evidence of poor performance. The fair question is whether the process provided the needed accommodation and whether the person could complete the visit, not whether every episode matches a single duration. Store only the accommodation information needed for the review and avoid inferring a disability from a long interval.
A useful report separates operational action from public benchmarking. It might show that payment starts are delayed on a particular shift, that abandoned episodes cluster before opening, or that a system outage creates a long tail. It should not claim that a public survey establishes a Virginia dispensary target. W3C and accessibility sources inform inclusive measurement; NIST and labor statistics provide context, not a local causal estimate. The bounded conclusion is that explicit clocks and outcomes can reveal where the experience changes, but no single wait number explains the customer journey.
A practical comparison boundary
Choose the observation cutoff before collecting post-visit surveys or delayed system corrections. A later completion timestamp can change the episode’s classification. Preserve the initial state, correction source, and missing-clock reason. This lets a team improve the data without pretending that a reconstructed duration was observed in real time.
What a reviewer should request
Request the event definitions, episode population, timestamp source, abandoned-visit treatment, channel segmentation, and staffing or outage notes for the same period. A median from completed visits cannot show the experience of people who left before service. Preserve incomplete episodes and explain any imputation; a reconstructed duration should never be presented as a directly observed clock.
Why tail cases deserve a named review
The longest episodes often contain the information needed for improvement, even when they contribute little to the median. Review a sample of tail cases by process segment and disposition, while protecting customer identity. A tail review can reveal outages, complex transactions, or access barriers that a single average cannot identify. It should be reported as an explanatory sample, not as a population rate.
A final evidence check
Before reporting a wait-time result, ask whether arrival, service, completion, abandonment, channel, and exception definitions are stable. Ask whether missing clocks and accessibility-related service differences remain visible. If not, limit the conclusion to the observed subset and do not present a median as a universal customer experience.
Reading uncertainty correctly
An incomplete episode should not be assigned the fastest or slowest plausible value. Retain the missing-clock status and analyze it as a data-quality outcome. If imputation is necessary for a planning model, state the assumptions and show the observed results separately. That boundary prevents a convenient estimate from becoming a claim about actual customer experience.
Bounded conclusion
Wait-time research becomes actionable when clocks, outcomes, segments, and tails are explicit. It can identify a queue or process question. It cannot produce a universal target or explain a difference without context.
Sources
Reviewed by the DispensaryVA editorial team on 2026-08-14.