Call center forecasting: Predicting volume with accuracy

Monday morning arrivals overshoot the released plan just as Finance locks the labor budget and Operations must decide whether to move skills, authorize overtime, or accept abandonment. Call center forecasting sits at that pressure point, translating historical demand into the interval-level plan that determines whether customers wait, agents burn out, or wages sit idle.

Treat each forecast as a versioned operational input. Record extraction time, queue definitions, routing state, event calendar, model version, and planner overrides so misses trace cleanly to data, assumptions, or model behavior. That audit trail gives Finance and Operations one demand record to approve, challenge, and defend, protecting service levels without obscuring who changed the plan or why.

What is call center forecasting?

Call center forecasting is the discipline of predicting future contact volume, handle time, and arrival patterns by queue and interval so workforce planners can staff each period to expected demand. A usable forecast is a versioned table of expected contacts and handle time, accompanied by its assumptions, event calendar, and accuracy results.

Workforce planners use that output to test whether a proposed roster covers the arrival pattern, not just a daily total. Two shifts with identical daily volumes can require very different headcounts if one concentrates arrivals into a narrow window, and the forecast must expose that difference before publishing schedules.

Building a forecast that holds up under Finance and Operations scrutiny starts with disciplined data preparation, deliberate method selection, and a clear-eyed view of what accuracy is worth.

Why forecast accuracy decides cost and service levels

An interval-level forecast typically models contact volume, average handle time (AHT), and the daily arrival pattern that drives demand. Small percentage errors compound across every interval, and cross-vertical call volume rose 16.1% year-on-year between 2024 and 2025 according to Natterbox, so every point of error now applies to a larger base.

The consequences of forecast error fall into four familiar categories:

• Idle labor cost: Overstated demand puts excess wages on the labor line without adding service value.

• Overtime and abandonment: Understated demand forces overtime, extends queues, and raises abandonment.

• Customer experience decline: Long waits and abandoned contacts pull down Net Promoter Score (NPS) and long-term retention.

• Planning credibility: Persistent misses erode Finance's trust in the numbers Workforce Management brings to the budget conversation.

Set the target internally against your own interval-level history, and review it alongside the workforce optimization (WFO) metrics your contact center WFO stack already reports.

How to build a volume forecast

Build a reproducible training set before fitting a model. Preserve the raw extract, record transformations and exclusions, and keep interval and event definitions consistent across telephony, routing, and workforce records.

1. Clean the historical data

Missing intervals from telephony outages or routing changes teach the model a demand dip that never happened, and outlier days such as a product recall or a billing error teach it a peak that will not recur on schedule. Keep those days and tag them as events, because the response pattern is worth retaining. Tag abandonment spikes linked to AI-agent latency or outages as events. For legitimately sparse intervals, distinguish a true zero from missing data so the model does not interpret a collection failure as an absence of demand.

Then select representative history that covers a full seasonal cycle, including the annual peak the model must predict.

2. Choose the forecast interval

Forecast at 15- or 30-minute intervals. Workforce teams often staff rosters against the worst half hour, and daily averaging hides exactly that peak. Two days with identical totals can require very different headcounts if one concentrates its arrivals into a narrow morning window.

Pair interval-level volume with interval-level AHT, because handle time moves across the day too, and a staffing calculation built on the daily average of both inputs inherits both errors at once.

3. Encode seasonality and known events

Models can learn recurring seasonality from history, including weekly and month-end rhythms plus annual peaks. Encode one-off events explicitly as model inputs. A marketing campaign or a product launch belongs on a shared calendar that Marketing and Operations both maintain.

Forrester predicts that at least three major brands will see single-day call spikes 100 times above normal on six occasions in 2026. For predictable seasonal peaks, plan how AI agents will capture and summarize requests during surges.

Choosing a forecasting method

Choose among methods by the length and stability of the available history, the frequency of sparse intervals, the availability of event variables, and the level of explainability planners and Finance require. Document each method's assumptions and the conditions that should trigger replacement or recalibration.

1. Erlang-based staffing conversion

Erlang-based queuing models translate a volume forecast into staffing requirements, and they remain a common method for that conversion. These models assume effectively single-skilled human agents handle homogeneous calls.

Multi-skill, multi-channel enterprise operations violate those assumptions every day. In multi-skill routing, an aggregate Erlang result can show sufficient total headcount. That aggregate result can hide shortages in the busiest queues. Use Erlang-based conversion only when queue assumptions hold; otherwise, model skills and queues separately to prevent hidden coverage gaps.

2. Time-series forecasting

Time-series models such as Holt-Winters and autoregressive integrated moving average (ARIMA) fit stable seasonal patterns cheaply by extrapolating history. A structural change such as a new channel or a deflection program can invalidate their predictions. Use these models for stable queues, and recalibrate or replace them after structural changes to keep schedules tied to current demand.

3. Machine learning forecasting

Choose machine learning when your team needs to model high-granularity demand across many external variables. Long short-term memory (LSTM) networks and other neural models can handle that complexity.

Dirty inputs and missing event encoding cap model accuracy and can produce inaccurate staffing requirements with high confidence. Most enterprises pair machine learning to predict volume with queuing math to convert the prediction into staffing, preserving high-granularity demand signals while producing staffing requirements planners can test.

Measure interval error and test model performance

The International Customer Management Institute (ICMI) reports that abandonment (85%), AHT (84%), and quality (77%) sit among the most measured contact center metrics, while forecast accuracy does not appear among the five most measured metrics.

Finance needs an error metric that shows both the size and direction of a miss. Select the measure according to the operational question it must answer and the queue volumes it will compare:

• Mean absolute percentage error (MAPE): averages percentage misses by interval but can overstate error when low volume makes a handful of calls look catastrophic.

• Weighted absolute percentage error (WAPE): weights errors by volume so large queues dominate, and raw team rankings can unfairly penalize low-volume queues.

• Mean absolute error (MAE): averages missed contacts but ignores scale, so the same result means different things on small and large queues.

• Forecast bias: shows whether forecasts run persistently high or low, exposing one-directional errors that inflate budgets or abandonment.

Report interval-level WAPE with forecast bias so Finance can see both the size and direction of the miss.

Validate on a holdout set before a model ships. For rolling-origin backtesting, move the cutoff forward through successive periods and evaluate each forecast only against observations that had not yet occurred when the team produced it. This shows whether performance holds across different weeks, peaks, and event conditions, not only at one convenient split.

After launch, compare forecasts with actual demand and test capacity scenarios. Validate interval-level fit; a correct daily number can completely conceal the uneven intraday arrivals that escalation clustering produces in the phone queue. Put the accuracy figure in the weekly operations review and the quarterly capacity meeting with Finance, with a named owner.

Set the right re-forecasting frequency

WAPE and bias checks only help when teams re-forecast fast enough to adjust the schedule. A Society of Workforce Planning Professionals survey found 40% of respondents reforecast monthly, 38% weekly, and 5% daily in spring 2025. Monthly and weekly re-forecasting often move more slowly than demand.

Align cadence to how quickly the operation can act on a revised forecast. Weekly cycles suit stable queues; daily or intraday cycles suit queues exposed to campaigns, outages, or shifting AI containment. Name an owner authorized to trigger an off-cycle re-forecast when arrivals, handle time, or containment deviate from plan, and confirm the operational levers, such as breaks, skill moves, overtime, or AI capacity, can still change the outcome before the affected intervals pass.

Forecasting when AI agents handle part of the volume

McKinsey partner Eric Buesing projects that AI could reduce current phone volumes by 50% within five years. That shift is an opportunity to sharpen forecasts, because AI-handled and human-handled work each carry their own volume pattern, handle time, and recontact behavior. Create separate forecasts for each stream and use the tips below to keep capacity plans accurate as automation scales.

Forecast the residual human mix as its own stream

AI agents typically resolve repetitive, short contacts first, so the calls reaching human agents skew toward complex, emotionally loaded cases. Treat that shift as useful signal. Tag and forecast the residual human mix as a distinct stream, then align skills and routing so complex queues retain enough coverage during peak windows. A dedicated residual forecast also gives Workforce Management a cleaner baseline to compare against as AI containment rates evolve month over month.

Model AHT separately for each stream

Handle time diverges once AI absorbs the short, simple contacts, and a blended AHT quietly under-staffs the hardest hours. Build separate AHT profiles for AI-handled and human-handled work so staffing reflects the actual workload each stream produces. Splitting AHT also makes coaching investments easier to justify, because Finance can see the handle-time trend on complex cases isolated from the automation gains showing up in the AI stream, and Operations can target training where it moves the number.

Add recontact volume to the human forecast

Contacts that fail in self-service return as calls, and the headline deflection figure excludes that recontact volume. Measure recontact by interval, attribute it to the originating AI flow, and add it to the human volume forecast so deflection does not understate required capacity.

Tracking recontact at interval grain also surfaces automation problems early: a rising recontact rate on a specific intent points Product and Operations to the flow that needs a fix before the human queue absorbs the miss.

Plan containment as a scenario range

Containment shifts as use cases expand and models change, so a single-point assumption turns the human-stream forecast into a guess. Plan at least three containment scenarios (conservative, expected, and aggressive) and validate each against measured containment actuals before capacity planning or budgeting. For each scenario, retain the AI volume, human volume, recontact, and AHT inputs separately so Finance and Operations can identify which assumption changed and move deliberately to the next scenario when triggers fire.

Build a defensible call center forecasting plan

A forecast earns trust not by being perfect, but by exposing the assumptions behind every number so Finance, Operations, and Workforce Management argue from the same evidence. Reconcile telephony, routing, workforce, and automation records before Finance approves hiring or overtime, and record each containment assumption, its confidence range, and the operational trigger that will move the plan to another scenario.

Parloa supplies governed AI Agent performance data across Build, Optimize, and Observe, with support for 140+ languages and deployments in as little as a few weeks, giving planners the containment, recontact, and AHT signals a defensible forecast requires.

Book a demo to connect your forecast thresholds to measured AI performance before customers and human agents absorb the next miss.

FAQs about call center forecasting

How often should you re-forecast?

Assign a named Workforce Management owner authority to trigger an intraday re-forecast when actual arrivals, handle time, routing, or containment depart from the released plan. The trigger must occur early enough for Operations to change breaks, move skills, offer overtime, or adjust AI capacity before the affected intervals pass.

How do AI agents change volume forecasting?

They change the unit of operational measurement. Track entry volume, successful containment, escalation, and recontact by interval, then reconcile those records with the human queue that received the work. Give Operations ownership of the handoff definitions and Finance visibility into the assumptions so a change in automation performance is not mistaken for a change in underlying customer demand.

Is Erlang C still useful?

Yes, when the queue used for staffing reasonably matches its assumptions. Before using the result, check whether skill sharing, channel blending, or AI-driven changes in contact mix make an aggregate headcount look sufficient while an individual queue remains exposed.