WFM guideCold-start forecasting

Forecasting a new contact centre operation

Q: How do you forecast contact volume for a new operation with no historical data?

When there is no historical contact data, you cannot use any extrapolation-based forecasting method (time series, day-of-week decomposition, seasonal models) because they all require history. Instead, build the forecast bottom-up from demand drivers: (1) Identify the population that generates contacts — number of customers, accounts, orders, policies, devices, or whatever unit drives demand. (2) Estimate the contact rate — how many contacts each unit generates per period. This is the hardest number; source it from a comparable existing operation, an industry benchmark, or the vendor/product team's expectation. (3) Multiply: population × contact rate = total contact volume for the period. (4) Distribute the total across the operating week using an intraday and day-of-week profile borrowed from a comparable operation (new operations rarely have a unique arrival shape — they usually resemble similar operations in the same sector and geography). (5) Apply the standard staffing model (Erlang C for voice, etc.) to the distributed volume to get the interval-level requirement. The output is necessarily uncertain — the contact rate assumption could be off by 30–50%. So you staff defensively (see the defensive staffing approach) and recalibrate the forecast weekly for the first 8 weeks as real arrival data accumulates.

Every standard forecasting method assumes you have history to extrapolate from. A brand-new operation has none. Cold-start forecasting is a different discipline: build demand bottom-up from drivers, borrow proxy data, staff defensively for the uncertainty, and recalibrate aggressively as real data arrives.

Why standard forecasting methods cannot be used

Time-series forecasting, day-of-week decomposition, seasonal indices, and trend extrapolation all share one requirement: historical data to learn from. A new operation, new queue, new product line, or freshly migrated customer base has no history — so there is nothing to extrapolate. Applying these methods to a cold start produces either an error (no data) or a false-confidence forecast built on a handful of early data points that are not yet representative. Cold-start forecasting replaces extrapolation with a bottom-up demand model plus proxy data from comparable operations.

The bottom-up demand model

Population

The unit that generates contacts — customers, accounts, orders, policies, devices, members. Source from the commercial / product team's launch expectation. This is usually the most reliable input.

Contact rate

How many contacts each unit generates per period (e.g. 0.3 contacts per customer per month). The hardest and most uncertain input — source from a comparable operation, an industry benchmark, or the product team. Model a range, not a point.

Total volume

Population × contact rate = total contacts per period. If 50,000 customers each generate 0.3 contacts/month, that's 15,000 contacts/month ≈ 3,460/week.

Distribute

Spread the total across the operating week using an intraday + day-of-week profile borrowed from a comparable operation. New operations rarely have a unique arrival shape — they resemble similar operations in the same sector.

Apply the staffing model

Run Erlang C (voice), the concurrency model (chat), or the backlog model (email) on the distributed interval volume to produce the interval-level agent requirement.

Four proxy data sources — in order of preference

A comparable existing operation (same company)

What it provides

Contact rate per customer/unit, intraday arrival profile, day-of-week pattern, AHT, and contact reason mix from an operation the company already runs that serves a similar customer base or product.

Reliability

Highest available for a cold start. Same company means similar systems, processes, and customer expectations.

Caveat

Adjust for known differences: a new product may generate more 'how do I' contacts early in its life; a different customer segment may have a different contact propensity. Do not assume the new operation is identical to the comparator.

Industry benchmark data

What it provides

Typical contact rates per customer, AHT ranges, and channel mix for the sector (e.g. utilities generate ~X contacts per customer per year; retail generates ~Y per order).

Reliability

Moderate. Useful as a sanity check and when no internal comparator exists, but benchmarks span a wide range and may not match your specific proposition.

Caveat

Benchmarks are averages across very different operations. Use the range, not a single point. If a benchmark says 2–6 contacts per customer per year, model all three of low (2), mid (4), and high (6) scenarios.

The product / commercial team's volume expectation

What it provides

The expected customer/order/account volume (the population that drives contacts) and the launch ramp curve (how quickly the customer base grows after go-live).

Reliability

Variable. The population estimate is usually more reliable than the contact-rate estimate. Commercial teams know how many customers they expect; they rarely know how often those customers will call.

Caveat

Commercial launch forecasts are often optimistic. Build the staffing forecast on the commercial team's expected case, but model a low case too — under-launch is as common as over-launch, and overstaffing a slow launch is expensive.

Migration / transfer data (if customers are moving from elsewhere)

What it provides

If the new operation is taking over an existing customer base (TUPE transfer, insourcing, acquisition), the prior provider's contact volume — if available — is the best possible proxy.

Reliability

High if the data is genuinely available and recent. This is real demand from the actual customers who will contact the new operation.

Caveat

Migration itself generates a temporary volume spike (customers confused by the change, new processes, new contact details). Model a migration-period uplift of 20–50% above the steady-state forecast for the first 4–8 weeks.

Staffing defensively under cold-start uncertainty

Staff to a scenario band, not a point estimate

Build low / mid / high volume scenarios (e.g. contact rate at benchmark-low, benchmark-mid, benchmark-high). Staff to somewhere between the mid and high case for go-live — understaffing a launch causes long queues, abandonment, and reputational damage at the worst possible moment (when the operation is being judged). Overstaffing is recoverable; a failed launch is not.

Front-load flexibility, not fixed headcount

Rather than committing to a high permanent headcount, build launch capacity from flexible sources where possible: overtime availability, a temporary cohort on fixed-term contracts, or agents borrowed from a comparable queue. As the real volume reveals itself, flexible capacity can be scaled down without redundancy cost.

Plan the ramp, not just the steady state

A new operation does not hit steady-state volume on day one. Customer numbers grow, awareness of the contact channel grows, and early-life contact reasons (onboarding, 'how do I') differ from mature-life reasons. Model the first 8–12 weeks as a ramp with rising volume and elevated AHT, not as immediate steady state.

Build the recalibration loop before go-live

Decide in advance how the forecast will be updated as real data arrives: who reviews actuals vs. forecast, how often (weekly for the first 8 weeks), and what triggers a staffing change. A cold-start forecast is a hypothesis — the value is in how fast you correct it, not in how accurate the initial guess was.

The first 8 weeks: rapid recalibration

The cold-start forecast is a hypothesis, not a plan

The value of a cold-start forecast is not its initial accuracy — which is necessarily low — but how quickly it is corrected. Build the recalibration loop before go-live:

→Weeks 1–4: review actual vs. forecast volume daily. The arrival profile (intraday + DOW shape) usually stabilises faster than the volume level — lock the shape early, keep adjusting the level.
→Weeks 4–8: move to weekly review. Recompute the contact rate from real data and re-run the bottom-up model with the corrected rate.
→Week 8+: enough history exists (8+ weeks) to begin using standard extrapolation methods. Transition from cold-start to normal forecasting.
→Throughout: track AHT separately — new operations almost always start with elevated AHT (agents are new, processes are immature) that falls as the operation matures.

New operation forecasting questions

How do you forecast contact volume for a new operation with no historical data?

Build it bottom-up from demand drivers, because no extrapolation method works without history. (1) Identify the population that generates contacts (customers, accounts, orders). (2) Estimate the contact rate per unit — the hardest number; borrow it from a comparable operation or an industry benchmark and model a range. (3) Multiply population × contact rate for total volume. (4) Distribute across the week using an intraday/DOW profile borrowed from a comparable operation. (5) Apply the standard staffing model to the distributed volume. Because the contact rate could be off by 30–50%, staff defensively (scenario band, flexible capacity, planned ramp) and recalibrate the forecast weekly for the first 8 weeks as real data arrives.