Industries / Finance and business services / Call centers and BPO
Staff scheduling
Staff scheduling, or workforce management (WFM), is forecasting how many calls and messages will arrive in each half-hour, working out how many agents are needed to answer them within a target time, building shifts and breaks that cover that need within labor rules, and adjusting them during the day.
7
Milestones achieved, from 1917 to 2005.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1917
A formula for waiting calls
Agner Krarup Erlang joined the Copenhagen Telephone Company in 1908 to head its new laboratory, showed in 1909 that calls placed at random follow a Poisson distribution (the pattern of independent random arrivals), and in 1917 published formulas for lost calls and for waiting time that telephone companies in many countries, including the British Post Office, soon used. The waiting-time model, now called Erlang C, is "the model most often used in call center analysis" (Garnett, Mandelbaum and Reiman, 2002), and in 1946 an international telephone committee named the unit of traffic the erlang.
Credit
A. K. Erlang, Copenhagen Telephone Company
1954
Shifts as a covering problem
The Port of New York Authority had more than 250 toll collectors, set by judgment, with drivers waiting anywhere from 2 to 50 seconds. Leslie Edie used probability theory to set how many booths to open each hour, and George Dantzig of RAND showed that linear programming could choose the collectors' work patterns to cover those hourly needs at least cost within meal-relief limits. Call centers still plan in these two steps: a queueing formula gives the need for each half-hour, then a "covering" model picks shifts that meet it.
LessonBreaks in the middle of shifts make the covering problem hard, so in practice it is solved with heuristics (methods that find good answers, not proven-best ones), as Gans, Koole and Mandelbaum noted in 2003.
Credit
Leslie C. Edie (Port of New York Authority); George B. Dantzig (RAND Corporation)
1981
The square-root staffing rule
Shlomo Halfin and Ward Whitt of Bell Laboratories studied queues in which calls and agents grow together while the chance that every agent is busy stays fixed, work that underpins the rule of thumb N = R + β√R: staff the offered load R (arrival rate times average call length) plus a safety margin that grows with its square root, with β set by the service target. Erlang had described a square-root relationship by 1924.
LessonThe spare capacity needed grows more slowly than the load, so large centers need proportionally fewer spare agents.
Credit
Shlomo Halfin and Ward Whitt, Bell Laboratories
1988
Rostering moves into software
IEX Corporation was founded in Richardson, Texas, in 1988; by 2006 it said its TotalView system was installed in over 3,000 contact centers, and that year its parent, Tekelec, sold it to NICE Systems for $200 million. A 2003 academic review found that building rosters "is very often supported by WFM software, whose core function is to forecast arrival rates and average service times" and then solve the staffing and scheduling models.
Credit
IEX Corporation (later NICE); TCS Management Group, another early vendor
2002
Counting callers who hang up
Erlang C assumes every caller waits as long as it takes. Ofer Garnett and Avishai Mandelbaum of the Technion and Martin Reiman of Bell Labs analyzed "Erlang A", which gives each caller a random patience, building on Conny Palm's 1943 work; in their example of 50 agents, 48 one-minute calls a minute and a 2-minute average patience, Erlang C predicts an average wait of 20.8 seconds, while with abandonment 3.1% of callers hang up and the average wait is 3.6 seconds. Free tools now take patience as an input, including Ger Koole's Erlang X calculator.
LessonTools that ignore abandonment over-staff, while service levels counted only on answered calls look better than they are and lead to under-staffing.
Credit
Ofer Garnett, Avishai Mandelbaum, Martin Reiman
2003
Labor law shapes the roster
The EU Working Time Directive 2003/88/EC, which codified rules first adopted in 1993, gives every worker at least 11 consecutive hours of rest in each 24 hours, a break when the working day is longer than six hours, an average of at most 48 hours a week, and at most 8 hours in 24 on average for night workers. Other countries price hours instead: the Philippine Labor Code (1974) adds at least 10% to pay for each hour worked between 10 pm and 6 am and at least 25% for overtime.
LessonThe same half-hourly demand gives a different cheapest roster in each country, so a schedule cannot be copied between sites without re-checking the rules.
Credit
European Parliament and Council of the EU; Labor Code of the Philippines (Presidential Decree 442)
2005
Real call data, in public
Lawrence Brown and six co-authors analyzed every 1999 call to a small Israeli bank's call center: of about 1.2 million calls, about 450,000 asked for an agent and about 20% of those hung up before being served. Arrivals fit a time-varying Poisson process, as Erlang assumed, but service times were close to lognormal (most calls short, with a long tail), not the exponential shape the standard models assume; and the manager had found agents hanging up on customers to get extra rest, under incentives that rewarded short talk time. The Technion's SEELab offers this dataset free for academic use, with a US bank's.
LessonA target on short calls can change how agents behave and corrupt the handle-time data that staffing is built on.
Credit
Brown, Gans, Mandelbaum, Sakov, Shen, Zeltyn and Zhao; Technion SEELab
Watch for
Software agents adjust staffing
Verint says its Workforce Intelligence adds "agentic intraday supervisor capabilities with real-time controls to adjust staffing and task assignments as conditions change", and Assembled describes AI agents that flag coverage gaps, fix adherence misses and publish overtime while planners set the rules. Neither published accuracy, service-level or cost figures.
Achieved whena customer or analyst publishes measured service level, abandonment and paid hours for centers where the software changed schedules without a person approving each change, against a comparison period or site.
Claimed by
Verint, June 2026; Assembled, May 2026
One forecaster for every series
Researchers at Amazon Web Services released Chronos-2, a pretrained model that forecasts without task-specific training, and report state-of-the-art results on three benchmarks, including GIFT-Eval; their case studies are energy and retail, and GIFT-Eval's 23 datasets include no contact center series.
Achieved whenan independent, reproducible test on public contact center arrival series at 15- or 30-minute intervals, over at least a full year including holidays, shows it matching or beating seasonal statistical baselines on error and on the staffing the forecasts imply.
Claimed by
Amazon Web Services, October 2025
Open
01
A benchmark for call arrivals
An open test set of contact center arrival counts at 15- and 30-minute intervals with fixed training and test periods, and a scoreboard that rates each forecast by ordinary error and by the staffing error it causes: the number of agents short or over in each interval when the forecast is run through Erlang A.
The briefCredit · unclaimed02
Rosters under real labor law
A library of call center shift and break scheduling problems, each pairing a week of half-hourly staffing needs with one country's rules written as constraints a program can check, with a validator that names every broken rule by article and baseline solutions.
The briefCredit · unclaimed03
Staffing behind a bot
A staffing calculator, checked by simulation, for a queue where an AI agent takes every contact first and passes a share to people, whose handle times differ and whose callers arrive having already spent some patience with the bot; compare its agent counts with the common shortcut of cutting volume by the bot's resolution rate and running Erlang C.
The briefCredit · unclaimed
Open 01 · the brief
A benchmark for call arrivals
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open test set of contact center arrival counts at 15- and 30-minute intervals with fixed training and test periods, and a scoreboard that rates each forecast by ordinary error and by the staffing error it causes: the number of agents short or over in each interval when the forecast is run through Erlang A.
Why it is still open
GIFT-Eval (Salesforce, 2024), a main benchmark for forecasting models, has 23 datasets in seven domains and none is a contact center series; an independent May 2026 evaluation of TimesFM and Chronos used traffic, energy, exchange-rate and M4 data; and the nearest call center study found (2023) tested neural networks on one company's private call volumes, so no one else can rerun it.
Built with what exists
NYC Open Data's 311 Call Center Inquiry dataset (105.7 million agent-handled calls timestamped to the second, 2010 to 2026, through the city's open-data API); SEELab's USBank call records (2001–2003); Chronos-2, Google's TimesFM and Salesforce's Moirai; Holt-Winters and seasonal ARIMA as baselines; GIFT-Eval's data format; the open-source pyworkforce package's Erlang A to turn forecasts into agent counts.
First step
Pull the NYC 311 inquiry records for 2015–2019 (the city says counts after March 2020 are incomplete), add them into 30-minute counts, and score a seasonal-naive forecast against Chronos-2 on 2019.
Done looks like
A public repository and scoreboard rebuilt from one script, covering at least three arrival series at 15 and 30 minutes, day-ahead and week-ahead, scoring at least three pretrained models and three classical baselines on error (MASE, WAPE) and on agent-intervals short.
Open 02 · the brief
Rosters under real labor law
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A library of call center shift and break scheduling problems, each pairing a week of half-hourly staffing needs with one country's rules written as constraints a program can check, with a validator that names every broken rule by article and baseline solutions.
Why it is still open
The public sets use generic rules, not named laws: Curtois's schedulingbenchmarks.org offers nurse rostering and multi-activity instances with constraints such as minimum rest between shifts, TU Wien's shift-design and break-scheduling instances are randomly generated, and a Mendeley call center set gives demand only. pyworkforce warns that "workforce policies vary by operation, country, and labor agreement" and leaves users to encode them.
Built with what exists
Mendeley Data's call center staff requirement set (Lalita and Murthy, 2021, CC BY 4.0, 336 half-hourly requirements per week) and requirement curves made from NYC 311 arrivals with Erlang A; EU Directive 2003/88/EC (Articles 3–6 and 8) and the Philippine Labor Code; Google OR-Tools CP-SAT through pyworkforce for baselines; the schedulingbenchmarks.org XML format so its RosterViewer can check solutions.
First step
Encode the EU 11-hour rest rule, the EU break-after-six-hours rule and the Philippine 10 pm to 6 am premium as constraints on one Mendeley demand week, and solve it with pyworkforce.
Done looks like
At least 30 instances (10 demand weeks times EU, Philippine and no-rules versions), each with a baseline schedule, its cost and a lower bound, and for each rule set a published "cost of compliance": extra paid hours against the no-rules version.
Open 03 · the brief
Staffing behind a bot
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A staffing calculator, checked by simulation, for a queue where an AI agent takes every contact first and passes a share to people, whose handle times differ and whose callers arrive having already spent some patience with the bot; compare its agent counts with the common shortcut of cutting volume by the bot's resolution rate and running Erlang C.
Why it is still open
The free tools in common use model human agents only: Ger Koole's calculators cover Erlang C, Erlang X, skill simulators and inbound/outbound blending, and pyworkforce covers Erlang C, A and B and multi-skill staffing. The closest research, Montanari, Scarsini and Perchet (July 2026), proposes a rule for routing between a chatbot and human queues and tests it only on synthetic cases, with no staffing method; vendors market a "hybrid workforce of human and AI agents" (Verint, 2026) without publishing one.
Built with what exists
SEELab's USBank records, which log each call's voice-response leg and its agent legs, a pre-AI case of an automated first stage handing over to people; pyworkforce's Erlang A as the baseline and Koole's simulator for cross-checks; a discrete-event simulation in Python with SimPy.
First step
For one month of USBank data, compute for each half-hour the share of calls that end in the voice-response unit and the handle times of calls that reach an agent.
Done looks like
Published tables of agents required under the new model, the shortcut and simulation for automated-resolution shares from 0% to 80% in 10-point steps and three handle-time ratios, with the new model within one agent of the simulation in at least 95% of cells, and a back-test on 100 USBank half-hours in which it predicts observed abandonment within 2 percentage points in at least 80% of them.