Workload & staffing demand forecasting (call volume, patient volume)

Every service operation that staffs people against incoming demand — a contact centre, an emergency department, a hospital ward, a retail store, a parcel depot, a Bürgeramt — has to decide weeks in advance how many people with which skills must be present in every 15- or 30-minute slot. Too few means queues, abandoned calls, crowded waiting rooms and burnt-out staff; too many means idle labour cost. AI forecasting predicts the workload and its handling effort per interval from history, calendar and external drivers, gives uncertainty bands, and converts the forecast into required headcount that the scheduling engine (OPT-003) then turns into rosters.

Problem & value

• Who has the problem: contact centres and shared-service desks (inbound calls, chats, emails, tickets), hospitals (ED arrivals and occupancy, admissions, ward census, OR load), care services, retail and hospitality (footfall, transactions), logistics (inbound parcels, picking volume), public administration (citizen-service appointments, benefit applications, 112/116117 call load).
• Today: Excel with "same week last year × growth factor", built-in moving averages in the ACD/WFM tool, or fixed staffing ratios. Workload planners spend much time manually adjusting for holidays, campaigns and one-off events, and intraday surprises are handled by overtime.
• Why AI helps: gradient-boosted and deep models, and since 2025 zero-shot time-series foundation models, capture multiple seasonalities (intraday × weekday × annual), holiday effects and external drivers far better than seasonal naive rules, and produce quantiles so that staffing can be set to a risk level rather than the mean.
• Evidence: in a large emergency department, N-BEATS and LightGBM forecasting occupancy 24 h ahead outperformed statistical benchmarks by 11 % and 9 % respectively, and DeepAR predicted next-day crowding with an AUC of 0.76 [1]. WFM vendors report customer outcomes such as a 20 % reduction in over- and under-scheduling (Quinyx) [6] and a 32 % service-level improvement within five weeks (Peopleware customer case) [5] — vendor-reported figures, not independent benchmarks.

Variants

• Contact centre / customer service: volume + AHT per queue, channel and skill at 15/30-min intervals; Erlang C/A or simulation to required agents; multi-skill and blended channels (voice, chat, email, messaging) where AI agents/voicebots (CONV-007) absorb part of the volume — forecasting must now predict the residual human workload after bot containment.
• Hospitals & care: ED arrivals and occupancy, admissions, ward census and nursing workload; drivers include weekday, holidays, temperature/heat waves, respiratory-virus surveillance. Bed and OR planning itself is HLTH-008; nurse rostering is OPT-003. In Germany, nurse staffing is constrained by Pflegepersonaluntergrenzen and PPR 2.0, so the forecast feeds a rule-based target.
• Emergency services & control rooms: call-load forecasting for 112/116117 dispatch centres (aggregate, not per-call triage — that is CONV-021; ambulance deployment is OPT-029).
• Retail, hospitality, logistics: labour-demand forecasting from sales/footfall (VIS-008), delivery waves and parcel volumes; strongly coupled to PRED-001 demand forecasting.
• Public sector: Bürgeramt/job-centre appointment and application volumes, tax-office seasonal peaks, citizen hotlines (115).
• Strategic workforce planning (long horizon): 1–5-year headcount and skills supply/demand — workload forecast × productivity × attrition (PRED-012) × hiring lead times, run as scenarios. Skills inference and the skills inventory belong to KNOW-012; the AI core here is the long-horizon demand forecast plus scenario simulation, so it is treated as a variant of this use case.

How it works

1. history (ACD/CRM/HIS/POS logs, per interval)
2. cleaning (outages, misrouted calls, abandonment & redial de-duplication, data gaps)
3. feature store (calendar, holidays, campaigns, weather, epidemiology, web traffic)
4. forecasting (volume + handle time/effort, quantiles; hierarchical: total
5. queue
6. interval)
7. reconciliation & planner overrides (known events) ← human-in-the-loop
8. staffing conversion (Erlang C/A, multi-skill simulation, ratio rules, shrinkage)
9. capacity plan / scheduling engine (OPT-003)
10. published rosters
11. intraday: streaming actuals
12. re-forecast every 15 min
13. alerts, VTO/overtime proposals

• Forecast engine: model ensembles — MSTL/ETS/TBATS for strong multi-seasonality, LightGBM with calendar and driver features, deep models (N-HiTS, TFT) for large portfolios, and foundation models (Chronos-2, TimesFM, TiRex) for zero-shot and cold-start queues. Volume and handle time are forecast separately because they react to different drivers.
• Staffing conversion: queueing models (Erlang C assumes waiting callers, Erlang A adds abandonment) or discrete-event simulation for multi-skill routing; then shrinkage (breaks, training, sickness) on top [8].
• Human-in-the-loop: workforce planners review forecasts, add known events (mailings, outages, marketing pushes), and approve the published capacity plan; rosters remain subject to works-council agreements. LLM assistants in recent WFM suites explain forecast changes and answer "what if" queries, but do not replace the planner's sign-off.

Data preparation

Which data: historical volumes (calls, tickets, patients, orders, footfall) at hourly or daily granularity, plus calendar effects, campaigns, weather, shift rosters, absences and service levels achieved.

Where it lives in an SMB: telephony or ticketing systems, point-of-sale or hospital information systems, ERP, time-tracking and shift tools, and Excel rosters on file shares.

Typical gaps: changing definitions of a "contact", gaps and outages in logs, volumes censored by capacity (abandoned calls or turned-away patients), campaign dates kept only in people's heads, rosters in spreadsheets, and personal data in shift records.

Steps: 1. Export two to three years of volumes from source systems (PREP-005). 2. Restructure rosters and Excel plans into tidy tables (PREP-011). 3. Align definitions of volume, handling time and service level (PREP-012). 4. Build an event calendar of holidays, campaigns and closures. 5. Aggregate to anonymous counts per interval before modelling so no personal data is needed (PREP-022). 6. Keep a hold-out period for backtesting (PREP-018).

Effort share: around 30-40 percent. Bought tools often connect directly to telephony and time-tracking systems, which shortens steps 1 and 2. See GUIDE-003.

Implementation paths

Which path for whom

• SMB (A): Workforce-management and shift-planning tools with built-in forecasting are mature, EU-hosted and affordable for standard demand patterns.
• Enterprise (B): Large operations have multiple demand drivers and own planning systems, so forecasts via cloud ML APIs integrated into existing workforce tools fit best.
• Developer (C): Open-source time-series libraries (Prophet, statsforecast, Chronos) run on HW1 and are easy to embed.
• Public (C): Hospitals and authorities handle patient or citizen data and procurement rules, favouring a self-hosted or sovereign-cloud solution.

Path A — Buy (recommended for most organisations)

• Contact-centre WFM: NICE CXone WFM, Verint WFM (Calabrio now part of Verint), Genesys Cloud WFM, Amazon Connect forecasting & capacity planning (short-term up to 18 weeks, long-term up to 64 weeks, intraday every 15 min, minimum data volumes per queue apply) [4], Assembled, Peopleware (formerly injixo; Peopleware GmbH, German) [5], Five9/Talkdesk built-in WFM.
• Retail / healthcare / general WFM: ATOSS (Germany) with demand-based duty planning for retail, healthcare and call centres [7], Quinyx (Sweden) [6], Planday (Denmark), Legion, UKG, Workday; hospital-specific rostering suites (e.g. Polypoint in DACH, Dedalus modules).
• Choose A when: the volume source is a standard ACD/CCaaS or POS system with a connector, and forecasting must sit next to scheduling. Check EU hosting, DPA, works-council-ready configuration (no individual performance monitoring by default) and whether forecasts are exportable.

Path B — Build on cloud APIs

• Managed forecasting: Amazon SageMaker Canvas / AutoGluon-TimeSeries on AWS (Frankfurt), Google BigQuery ML AI.FORECAST with TimesFM (EU multi-region) [3], Azure Machine Learning AutoML forecasting (West Europe/Germany West Central), Nixtla TimeGPT API.
• Integration: pull interval data from ACD/CCaaS reporting APIs, hospital information systems (HL7/FHIR encounter feeds) or POS; write required-staff curves back into the WFM tool via CSV/API.
• Choose B when: you have a data team and specific drivers (e.g. billing runs, epidemiology, campaign data) that packaged WFM forecasting ignores, or when a single forecast must feed several scheduling systems.

Path C — Self-host (open source)

• Libraries: Nixtla StatsForecast (AutoETS, MSTL, TBATS; fast for many series) [9], MLForecast (LightGBM), NeuralForecast (N-HiTS, TFT, DeepAR; Apache-2.0) [10], AutoGluon-TimeSeries, Darts, sktime.
• Foundation models (open weights): Amazon Chronos-2 (120M, zero-shot with past/future covariates, Oct 2025) [2], Google TimesFM 2.5/3.0 (3.0 adds native dynamic covariates, Aug 2026) [3], TiRex (NXAI, Austria; xLSTM, 35M) [11], Salesforce Moirai.
• Staffing: pyworkforce (Erlang B/C/A, multi-skill staffing, OR-Tools shift scheduling; MIT, beta) [8]; SimPy for contact-centre simulation; Timefold/OR-Tools for rostering (OPT-003).
• Hardware: HW1 — a CPU server handles thousands of interval series with statistical/GBM models; foundation-model inference for typical portfolios also runs on CPU, a small GPU only speeds up backtesting. Ops: daily batch jobs, monitoring of forecast error, data-pipeline maintenance.

Path D — Custom training

Worth it only for large operators (multi-site contact centres, hospital groups, national postal or benefit agencies) with years of interval history: fine-tune a foundation model or train a global deep model across all queues/sites, typically 2+ years of interval data per series plus driver history. More often, the gains come from better driver data (campaign calendars, outage feeds, epidemiology) and planner-override learning rather than from a bespoke architecture.

State of the art (2026)

• Foundation models went mainstream: Chronos-2 (Oct 2025) reports best pretrained-model results on fev-bench, GIFT-Eval and Chronos Benchmark II, with the largest gains on tasks with exogenous features [2]; TimesFM 3.0 (Aug 2026) adds native past/future covariates and is rolled out in BigQuery ML [3]; TiRex claims top GIFT-Eval/fev results with a small xLSTM model [11]. Covariate support matters here because holidays, campaigns and weather drive staffing demand.
• Probabilistic staffing: planning against quantiles (e.g. P80 volume) instead of the mean, and propagating uncertainty into the staffing curve, is replacing single-point Erlang inputs.
• Agentic WFM: vendors (NICE, Verint, Peopleware, Assembled) add LLM copilots that explain forecasts, simulate scenarios and propose intraday actions; value depends on the quality of the underlying forecast.
• Bots change the target: with AI voice/chat agents containing a growing share of contacts, the forecast must model bot containment and the harder residual contacts with longer handle times.
• Limits: structural breaks (pandemic, outages, new products, regulatory deadlines) and rare events are not predictable from history; small queues are dominated by noise; Erlang C overstates staffing when abandonment is high (use Erlang A or simulation); ED studies show solid next-day occupancy forecasts but limited skill for extreme crowding peaks [1].

Shared resources & combinations

• Same forecasting stack (time-series + tabular-ml, HW1): PRED-001 demand forecasting, PRED-007 energy load, PRED-016 cash flow, PRED-019 traffic & mobility demand — one feature store, backtesting framework and foundation-model service can serve all of them.
• Downstream optimisation: OPT-003 workforce scheduling & shift planning (the natural pair), OPT-035 predictive customer-to-agent routing, OPT-029 emergency resource deployment, HLTH-008 bed/OR/staff planning in hospitals.
• Upstream drivers: PRED-017 weather impact, PRED-022 no-show prediction (effective appointment load), VIS-008 footfall counting, LANG-012 ticket classification (workload by skill).
• Strategic variant: PRED-012 attrition prediction and KNOW-012 skills mapping feed long-horizon headcount and skills-gap scenarios.
• Contact-centre AI: CONV-007 agent assist and voicebots change AHT and residual volume — forecast them jointly.

Effort, cost & KPIs

• Path A: weeks to configure forecasting inside a WFM suite; licence typically per agent/employee per month; the main effort is clean history and event calendars.
• Path B/C: 1–3 months for a backtested pipeline with drivers and WFM integration; compute cost is negligible, staff cost dominates.
• Path D: 3–6 months with a data-science team; only justified at large scale.
• KPIs: interval-level forecast error (WAPE/MASE per queue and interval, quantile coverage), service level and average speed of answer, abandonment rate, ED waiting times / crowding hours, over-/under-staffing hours, overtime and agency-staff cost, planner time spent on overrides.

People & change

Involve early: management (service levels, budget), the operations or team lead who owns the roster, IT (interfaces to telephony, ticketing, time tracking), HR (working-time rules), and the works council. Include end users who will see changed shift proposals; they know the real patterns. Add the DPO when individual shift or absence data is used.

Works council: depends. A pure forecast on anonymous volumes triggers little, but the moment it feeds shift plans, tracks individual handling times or logs user activity, co-determination applies (§87(1) no. 6 BetrVG, plus no. 2 and 3 on working hours). The council must be informed when AI-supported planning is introduced (§90(1) no. 3); an AI expert may be engaged (§80(3)).

Data protection: a DPIA is usually not needed for aggregated counts, but check it if individual performance data enters the model. AVV with any SaaS provider (GUIDE-005).

AI literacy: Art. 4 AI Act requires measures that support literacy; teach planners how to read forecast intervals (GUIDE-006).

Pilot and communication: run the forecast beside the current plan for one team for a few weeks (GUIDE-009).

Concerns: staff fear opaque scheduling and lost shift preferences. Address this with transparent rules, published forecast accuracy and planner override rights, so humans keep the final roster (GUIDE-004, GUIDE-008).

Risks, regulation & compliance

• EU AI Act: forecasting aggregate workload is minimal risk. Annex III point 4(b) makes systems high-risk when they "allocate tasks based on individual behaviour or personal traits" or "monitor and evaluate the performance and behaviour" of workers [12] — so keep the forecast aggregate and do not mix in individual productivity scoring. Point 5(d) (classifying emergency calls, dispatch, patient triage) is high-risk, but aggregate call-load forecasting for a control room is not triage. AI literacy (Art. 4) applies.
• GDPR: input data is mostly aggregated counts; raw ACD/HIS logs and individual AHT contain personal data (staff and customers/patients) — aggregate early, pseudonymise, apply purpose limitation; health data (Art. 9) for patient-volume models stays in hospital infrastructure.
• Co-determination (Germany): WFM systems that can monitor behaviour or performance trigger works-council co-determination under §87(1) Nr. 6 BetrVG, and working-time/shift distribution under §87(1) Nr. 2 [13]; conclude a Betriebsvereinbarung before go-live.
• Sector rules: hospital nurse staffing minimums (Pflegepersonaluntergrenzen, PPR 2.0) and Working Time Act limits override any forecast-optimal staffing; DORA operational-resilience expectations apply to financial-sector service desks only indirectly.
• Risks: chronic under-forecasting used to cut staff, feedback loops (understaffing → abandonment → lower recorded demand), ignoring redials. Mitigate with demand-not-handled correction, quantile staffing and transparent override logs.

Getting started

1. Export 2+ years of interval data (offered, handled, abandoned contacts and AHT; or arrivals/census) and build an event calendar of past campaigns, outages and holidays.
2. Establish a baseline (seasonal naive, current WFM forecast) and measure WAPE per queue and interval.
3. Backtest StatsForecast/LightGBM and a foundation model (Chronos-2 or TimesFM) with calendar covariates on a rolling origin; compare against the baseline.
4. Convert forecasts into required staff with Erlang A or simulation, and compare with actual service levels to validate the full chain.
5. Integrate into the WFM/scheduling tool with planner overrides, agree a works-council agreement, and monitor forecast error weekly.

Sources

1. Tuominen et al. — Forecasting Emergency Department Crowding with Advanced Machine Learning Models and Multivariable Input (arXiv:2308.16544) — https://arxiv.org/abs/2308.16544
2. Amazon Science — Chronos / Chronos-2 forecasting models — https://github.com/amazon-science/chronos-forecasting
3. Google Research — TimesFM (2.5 / 3.0) — https://github.com/google-research/timesfm
4. AWS — Forecasting in Amazon Connect — https://docs.aws.amazon.com/connect/latest/adminguide/forecasting.html
5. Peopleware (formerly injixo) — AI-native WFM suite for contact centres — https://www.peopleware.com/
6. Quinyx — workforce management with AI demand forecasting — https://www.quinyx.com/
7. ATOSS — workforce management and demand-based duty planning — https://www.atoss.com/en/
8. pyworkforce — Erlang C/A/B staffing, shift scheduling and rostering in Python — https://github.com/rodrigo-arenas/pyworkforce
9. Nixtla — StatsForecast — https://nixtlaverse.nixtla.io/statsforecast/index.html
10. Nixtla — NeuralForecast — https://github.com/Nixtla/neuralforecast
11. NXAI — TiRex time-series foundation model — https://github.com/NX-AI/tirex
12. EU AI Act — Annex III high-risk use cases — https://artificialintelligenceact.eu/annex/3/
13. Betriebsverfassungsgesetz §87 (co-determination) — https://www.gesetze-im-internet.de/betrvg/__87.html
14. Verint — Workforce Management (incl. Calabrio ONE) — https://www.verint.com/