Most front desk schedules come from one number: forecast occupancy. The manager sees a 90% Saturday, remembers that 90% Saturdays feel busy, and puts three people on the evening shift. It is the wrong input. Occupancy counts guests already in their rooms; the desk's workload comes from people walking through the door, and the two move apart.
A front desk staffing forecast built on arrivals changes the schedule on roughly a third of the days in a season, usually by one position for three or four hours. The service effect is the bigger one, because waits at a counter rise slowly, then all at once.
In the American Hotel & Lodging Association's February 2026 survey of 246 hoteliers, more than half said their properties were somewhat or severely understaffed and 65% named labor costs as a financial pressure. Hiring out of it is hard: BLS August 2026 JOLTS data puts the job openings rate in accommodation and food services at 4.7% against 4.3% across all industries. What is left is putting the staff you have where they are needed.
Table of Contents
- Why Occupancy Is the Wrong Input for a Front Desk Staffing Forecast
- The Four Inputs Your PMS Already Has
- From Arrivals to Agents: The Math, Worked
- What It Saves, and What Being Short Costs
- Where the Forecast Breaks, and What Not to Automate
- A 90-Day Plan to Put the Forecast in Production
- Frequently Asked Questions
- Ready to Build a Front Desk Staffing Forecast You Can Defend?
Why Occupancy Is the Wrong Input for a Front Desk Staffing Forecast
Occupancy tells you how many rooms are sold; arrivals tell you how many transactions the desk processes. The ratio between them is average length of stay.
A 140-room hotel at 94% has 132 occupied rooms. At an average stay of 1.9 nights, arrivals run about 69 a day. At 4.2 nights, the same 132 rooms produce about 31, less than half the workload at identical occupancy. That is why resort and extended-stay properties are most often overstaffed midweek and caught short on changeover day.
Grain compounds it: guests experience the forty minutes when eleven people arrive and two agents are working, not a daily average.
The costlier error is averaging away the clumps. The staffing advice circulating online totals expected transaction minutes and divides by the peak window, which assumes arrivals are evenly spaced. They are not, and the clumping creates the queue. The research literature handles this with square-root staffing: positions equal average workload plus a safety factor times the square root of that workload, as in Zan, Hasenbein and Morton on staffing large service systems. That term is what front-office guides leave out, and it is usually worth a position at peak.
The Four Inputs Your PMS Already Has
All four already sit in your property management system.
- Daily arrivals, actual and on-the-books. Not occupancy. Pull 24 months of actual arrivals by date, plus on-the-books arrivals for future dates. The gap between the two at 7, 14 and 30 days out is your pickup curve.
- Check-in timestamps. Export every arrival's check-in time over 90 days and compute each hour's share, separately for weekdays, weekends and your changeover day. That is your arrival curve, and it is stable at a given property.
- Average handle time. Time 40 check-ins across agents and dayparts, greeting to guest walking away. Most properties guess three minutes and measure four and a half.
- Non-arrival desk work. For two weeks, have agents tally minutes per hour on anything that is not a check-in: phones, departures, packages, guest requests. This is the input that gets omitted, then blamed when the schedule fails.
No check-in timestamps? Door lock systems keep first key-encode times. Process first, technology second.
From Arrivals to Agents: The Math, Worked
Three steps: turn arrivals and handle time into offered load, add the square-root safety term, round up. Worked below on a hypothetical 140-room select-service hotel.
Assumptions, stated: a summer changeover Saturday at 94% occupancy, 132 occupied rooms, 64 arrivals on the PMS arrivals report, 4.5-minute handle time, 22 minutes of non-arrival work per hour, and an arrival curve of 8% before 3 p.m., 24% from 3 to 4, 27% from 4 to 5, 19% from 5 to 6, 13% from 6 to 8 and 9% after 8.
Offered load. The 4 to 5 p.m. hour takes 27% of 64 arrivals, or 17.3. At 4.5 minutes each that is 77.9 minutes of work inside a 60-minute hour, or 1.30 positions. Add 0.37 for non-arrival work: 1.67 positions.
Safety term. The square root of 1.67 is 1.29. A coefficient of 1.0 gives 2.96, rounding to three agents. A lean 0.5 gives 2.32, still three; a service-first 1.5 rounds to four.
What each option buys. Model that hour as a simple queue at 1.67 positions of load and a 4.5-minute service time. Two agents run at 84% utilization, 76% of guests wait, and the average wait across all arrivals is about ten and a half minutes. Three agents run at 56% utilization, 30% wait at all, and the average wait is about one minute. One position is the difference between a ten-minute line and a one-minute line, which is why "only one short" is never only one short.
Across the evening the schedule writes itself: three agents from 3 to 6 p.m., two after. Now run an in-season Wednesday, when occupancy is still 88% but weekly guests are in-house and arrivals are 24, not 64. Peak load falls to 0.86 and positions to 1.79, rounding to two, while the occupancy method says three. One position across four hours, most weeks of the season, is the money. One caution: the model assumes random arrivals and a single line.
What It Saves, and What Being Short Costs
Price both sides first. BLS reports average hourly earnings of $22.07 for nonsupervisory accommodation employees as of August 2026, in its industry data for NAICS 721. Front desk roles sit at or below that; with taxes and benefits, $26 to $29 an hour is defensible. The example uses $27.
The saving. Say the forecast removes one evening position for four hours on 110 in-season midweek days and three hours on 40 light shoulder days, adding one back for three hours on the 25 heaviest changeover afternoons. That is 485 coverage hours a year, about $13,100, or roughly a quarter of one position. Not a headcount cut; the same people on the hours that matter. Converting coverage hours into budgeted positions, divide by 1,880 productive hours, not 2,080.
Being short never shows in the labor line, which is why it goes unmanaged. It shows in the arrival that sets the tone for the stay, the upsell nobody offered, and review scores. HotStats reported in December 2025 that US non-union hotels ran a 33.5% labor cost ratio against a 36.5% gross operating profit margin, so an avoidable labor hour is close to a dollar of profit, and so is a bad arrival.
Self-service changes the arrival rate itself. A Mews survey of 2,000 US travelers found 70% likely to check in by app or kiosk (June 2025). Run that through the model, though: pulling 30% of arrivals out of the peak hour drops load only to 1.28, still three agents. Deflection shortens the wait; it does not remove the position.
Where the Forecast Breaks, and What Not to Automate
Six failure modes, each with a check.
Forecasting the wrong variable. A model trained on occupancy tracks occupancy and misses the desk. Check: the target is arrivals by day, and by hour.
Daily grain. A forecast accurate on the day but silent on the hour cannot write a schedule. Check: every output is an hourly position count.
A frozen arrival curve. Curves shift when a flight schedule changes or your OTA mix moves. Check: recompute quarterly.
Group blocks and same-day pickup. A 40-room block arriving together is not a random process, and same-day bookings land after the schedule is posted. Check: carry group arrivals as a separate line, and track what books inside 48 hours.
Publishing too late to matter. A forecast that improves on Thursday for a Saturday posted two weeks ago is a report, not a schedule. Check: forecast to your posting deadline and cover the gap with a flex position.
Grading the model against itself. Check: compute the naive forecast's error first — last year's same weekday — as mean absolute percentage error on daily arrivals. Any model must beat that baseline within a season or be switched off.
What not to automate: accessibility and ADA requests, which need a human reading the reservation before arrival; group blocks; service recovery for a guest already unhappy; early-arrival and upgrade calls, which are revenue judgments. The pattern is the one our overview of AI for hospitality operations sets out: the model says how many people to put where and when, and a person decides what happens with the guest at the desk.
A 90-Day Plan to Put the Forecast in Production
Measure for 30 days, build and shadow for 30, schedule from it for 30. Only the first phase lets you prove it worked.
Days 1 to 30: build the baseline. Export 24 months of daily arrivals and 90 days of check-in timestamps. Compute the arrival curve by day type. Time 40 check-ins and run the non-arrival tally. Then compute your naive forecast error. Buy nothing.
Days 31 to 60: build it and shadow it. Start with a spreadsheet, not a platform. Forecast arrivals from on-the-books plus your pickup curve, apply the arrival curve, convert to offered load, add the square-root term, round up. Each week, produce the forecast schedule beside the one you posted and log the differences. Our piece on seasonal demand forecasting covers the pickup-curve mechanics. Kill criterion: if after four weeks the model matches what the manager would have written, your constraint is elsewhere.
Days 61 to 90: schedule from it and keep score. Post from the model, with the manager holding an override that gets logged with a reason. Those overrides are the quarter's most valuable output: they show what the model does not know. Track four measures weekly: forecast error, hourly variance between forecast and posted schedule, desk hours paid, and a service proxy such as arrival-experience review mentions. If error beat the naive baseline and hours fell without the proxy moving, put the model into the budget process alongside the revenue side of the business.
CoStar and Tourism Economics put 2026 US occupancy at 63.1%, with gross operating profit per available room up 4%, in their August 2026 forecast upgrade. Demand is not the binding constraint at most properties. Cost and coverage are.
Frequently Asked Questions
How much historical data do I need before this works?
Ninety days of check-in timestamps is enough for a usable arrival curve, because intraday patterns are stable at a given property. Daily arrival forecasting needs more: 24 months separates seasonality from day of week and catches holidays. With 12 months, build the curve now and forecast from last year's weekday.Does this work for a 40-room property with one person on the desk?
The math applies but the decision changes, because you cannot schedule 1.4 agents. The outputs get narrower: which hours justify a second position at all, whether a four-hour evening shift beats a full one, and whether self-service is cheaper than a hire.Do I need an AI tool, or will a spreadsheet do?
A spreadsheet does everything described here, and building it that way first shows which input drives the answer. Machine learning earns its place later, on the harder parts: predicting pickup between booking and arrival, and spotting that the curve has shifted. Do not buy a model before you can state your error.What service level should I target for check-in waits?
There is no industry standard, so set one and hold it. A common structure is 80% of arrivals waiting under two minutes, with the peak hour allowed a looser standard than the rest of the day. Which number you pick matters less than picking one: the safety coefficient then follows.Will this cost people their jobs?
Done properly it moves hours rather than removing them. The net change in the worked example is about 485 coverage hours a year, roughly a quarter of one position, and most properties absorb that by shifting hours to peaks they were already failing to cover. More than half of properties in the AHLA's February 2026 survey were understaffed already.Ready to Build a Front Desk Staffing Forecast You Can Defend?
A model built from your own arrivals, handle time and arrival curve beats any benchmark, and it turns a recurring argument about the evening shift into a decision you can check in ninety days. Process first, technology second: the measurement phase usually finds more than the model does. ANtreliX is a strategy-first AI and business intelligence consultancy in North Myrtle Beach, South Carolina, serving hospitality operators on the Grand Strand and nationwide. Talk to ANtreliX, or see our case studies.