Demand decomposition by call reason

Forecasting · ~6 minute read

The single number that hides everything

Most contact centres forecast a single number per interval — total contact volume. The number is convenient, but it hides the structure that drives the operation. A forecast of 500 calls in an hour means something very different if it’s 450 billing queries and 50 complaints, versus 100 billing queries and 400 retention conversations. The AHT differs, the agents who can handle it differ, the seasonality differs, the cost-to-serve differs, and the management response when it goes wrong differs. Demand decomposition is the practice of forecasting each contact reason separately, recombining them into a total, and using the structure of the demand — not just its size — to drive better decisions. This article walks through what decomposition is, how to set up the call-reason taxonomy, the forecasting benefit, the operational benefit, and the practical pitfalls.

What decomposition is and isn’t

Decomposition is forecasting at a more granular level than the headline volume and aggregating up. The grain that consistently works in a contact centre is the call reason — billing query, claims, technical support, complaint, retention attempt, account change, and so on. Each reason has its own arrival pattern, seasonality, AHT, and operational handling. Decomposing the forecast by reason rather than just by skill or channel reveals structure that aggregated forecasts hide.

Decomposition isn’t the same as multi-skill scheduling, although they sometimes overlap. Multi-skill scheduling assigns agents to multiple skill queues; decomposition is about the planner’s view of demand, regardless of how the routing then sends it to agents. The two complement each other, and operations that do both well produce both better forecasts and better routing.

Setting up the call-reason taxonomy

The taxonomy — the set of categories the operation uses to label contacts — is the foundation. A taxonomy with 4 categories is too coarse to be useful; a taxonomy with 80 categories is too granular to manage. The right size for most operations is between 8 and 20 categories, organised hierarchically so granular categories can be rolled up to broader ones for higher-level reporting.

The taxonomy needs three properties to be useful: collectively exhaustive (every contact fits somewhere), mutually exclusive (no contact fits in more than one primary category), and operationally meaningful (categories the operation actually treats differently). Categories that meet the first two but fail the third (“billing-positive” vs “billing-negative”) are noise; categories that meet the third but fail the first two (overlapping or incomplete) create misclassification that distorts the data.

How the data gets captured

Three approaches consistently work. Agent-tagged wrap codes. The agent selects a wrap code at the end of the call from a structured list. Cheap to implement, but agents under time pressure choose the easiest option rather than the right one, and the data quality is mediocre. IVR-driven categorisation. The customer selects from menu options; the system records the path. Better for some contact types, useless when the customer’s actual issue differs from the menu path they took. Speech analytics. The platform listens to every contact and categorises it. Slowest to set up, highest data quality, requires investment. See speech analytics for planners for the wider treatment.

Most operations end up combining the three: speech analytics for the high-volume routine categories, agent wrap codes for the long tail, IVR data for the categories that map naturally. The combination delivers more than any one source alone.

The forecasting benefit

Decomposed forecasting is usually more accurate than aggregated forecasting because each category has cleaner structure. Billing volumes are driven by billing cycles, which are predictable. Complaints are driven by service incidents, which are spike-prone but distinct. Retention contacts are driven by competitive activity and seasonality. Each category has its own statistical model that fits it well; the aggregate of those models usually beats a single model fitted to the total.

The benefit is largest where the category mix is changing. An operation whose billing share is shrinking and whose retention share is growing will have its aggregate AHT drift over time, and a single-number forecast will absorb that drift as noise. A decomposed forecast captures the mix shift explicitly and surfaces the AHT and staffing implications.

The operational benefit

Decomposition pays back beyond the headline accuracy. Strategic conversations get better. “Complaints are up 15% year-on-year” is actionable in a way that “volume is up 2%” is not. Driver analysis becomes possible. A category that’s growing or shrinking can be investigated; an aggregate that’s drifting is harder to diagnose. Self-service and deflection investment can be targeted. The categories ripe for deflection (high-volume, low-complexity, repeat-prone) become visible. Real-time interventions can be smarter. A volume spike in one category triggers different responses than a spike across the board. AHT changes can be diagnosed. When average handle time drifts up, decomposition tells you whether it’s a mix shift (more complex categories) or genuine within-category change.

Practical setup

A workable decomposition setup follows four steps. The first is to design the taxonomy with operations, training, QA, and the planning team in the room. The second is to instrument the capture — configure wrap codes, the speech analytics platform, and the IVR for the chosen categories. The third is to build per-category forecasts, starting with the largest categories where the data is densest and the accuracy gain is largest. The fourth is to maintain the taxonomy — review quarterly, retire categories that no longer apply, split categories that have grown too coarse. The taxonomy is a living artefact, not a one-time design.

Common mistakes

Three patterns recur. Too many categories. A 50-category taxonomy produces small-volume categories that are hard to forecast accurately and difficult to manage. Most operations are better with fewer, broader categories. Trusting wrap codes alone. Wrap-code data is well-known to be imperfect; treating it as ground truth misleads. Triangulate with speech analytics or sample QA. Decomposing without operationalising. A decomposed forecast that nobody uses is wasted effort. The decomposition should drive the schedule, the routing, and the reporting; if it doesn’t, the additional complexity has no return.

Conclusion

Demand decomposition is one of the most valuable upgrades an operation can make to its forecasting practice. The setup cost is real — taxonomy design, data capture infrastructure, more complex forecasting — and the payback is substantial: better forecasts, better operational diagnostics, better strategic conversations, and better-targeted investment in self-service and deflection. The operations that take it seriously develop a view of demand that is structural rather than aggregate, and the planning function’s influence grows correspondingly. The discipline is in the taxonomy and the data capture, not in the modelling sophistication.

Pair this with speech analytics for planners, forecasting method: hierarchical reconciliation (which is the technique for recombining the decomposed forecasts cleanly), and using AI for contact centre forecasting.