Journal / Kirill Naumov
AI in Retail Scheduling: Forecast the Work, Then Let Managers Make the Rota
How retailers can use AI to anticipate store workload without turning a forecast into an opaque or unfair staff schedule.
By Kirill Naumov ·
A store schedule is not just a demand curve with names attached. It has to cover the work customers actually create: deliveries, replenishment, service questions, returns, breaks, training and the unexpected rush that arrives just after someone says “it looks quiet”. AI may help estimate when and where work is likely. It should not quietly decide what a person’s week looks like.
Forecast tasks, not just footfall
Start with the operational question. Is the team short at fitting rooms on busy evenings? Does click-and-collect create a predictable picking peak? Do deliveries collide with customer-service demand? A useful forecast combines appropriate signals—such as historic transactions, planned promotions, delivery windows and local opening hours—with a clear definition of the task to be covered. More data is not automatically better; stale or irrelevant data can make an elaborate forecast less useful than a manager’s simple checklist.
Keep the first use narrow. Compare a small set of stores or shifts with the existing planning method. Include quiet periods and unusual days, not only the weeks when the model looks clever. A forecast is an estimate with uncertainty, not a promise that demand will arrive on schedule.
Give the manager a reason, not a mysterious number
If a system recommends more coverage between 5 and 7 p.m., show the operational reason: a promotion, delivery window, repeat service queue or a pattern in comparable weeks. Show uncertainty too. If there is not enough reliable history, say so. A manager should be able to challenge the recommendation, add local context and record why the final schedule differs.
The recommendation should not present a precise-looking staffing figure without explaining its assumptions. A short list of evidence—plus a range—often supports a better decision than a single number with decimal places and a confident tone.
Build fairness into the workflow
Scheduling affects income, caring responsibilities, commuting and recovery time. Before using AI, decide which employee information is genuinely necessary, who can see it and how long it is kept. Do not infer availability, reliability or commitment from sensitive personal signals or incomplete attendance records. Give colleagues a clear route to correct bad data and ask for a human review.
The [International Labour Organization’s 2025 working paper on AI in human resource management](https://www.ilo.org/publications/ai-human-resource-management-limits-empiricism) warns that flawed objectives, biased data and opaque systems can undermine AI uses in areas including scheduling and performance management. That is a reason to define the intended benefit and the human process before a tool is connected to a rota—not an argument that every scheduling aid is inherently harmful.
NIST’s [AI Risk Management Framework](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) likewise treats governance, context, measurement and ongoing management as connected responsibilities. Applied to scheduling, that means a named owner, documented limits, checks before rollout, a way to report problems and regular review after the system enters real operations.
Measure the shift people actually worked
Track more than forecast accuracy. Look at uncovered service periods, last-minute changes, overtime, schedule stability, customer waiting and employee-reported fairness. Compare similar stores and shifts, account for promotions and local events, and examine who repeatedly receives the least predictable hours. Accuracy can improve while the lived experience gets worse; both matter.
Set a pause rule before the pilot. If the tool systematically underestimates a type of work, creates repeated short-notice changes or produces a pattern managers cannot explain, stop and investigate. The right next move may be better task data, a different planning rule or no AI at all.
A practical first test
Pick one recurring workload, one group of stores and a short, representative test window. Run the AI forecast beside the current approach; do not let it publish schedules automatically. Have managers record which recommendations they use or change and why. Ask colleagues whether coverage and predictability improved. Expand only when the forecast helps teams serve customers without shifting hidden costs onto workers or managers.