Supply Elasticity in Workforce Planning

From WFM Labs

Supply elasticity is the extent to which staffed, proficient capacity can be changed inside the horizon over which demand is predicted. It is a property of the workforce and its routing rules, not of the planning function, and it determines how much of any forecast improvement can actually be converted into an operational outcome. Where elasticity is low, the marginal value of forecast accuracy is bounded: a surge can be predicted precisely and still not be met. This page sets out the ceiling, the two directions from which it can be raised — distributional demand modelling and dynamic supply reallocation — and the harder limit that both eventually reach, which is the composition of the workforce itself.

The framing is most consequential in complex servicing environments such as travel, financial services and technical support, where time to proficiency is measured in months rather than weeks.

The forecast value ceiling

Demand in these environments is volatile and event-driven. Arrival patterns are time-varying, stochastic, dependent across time periods and call types, and frequently driven by external events.[1] The standard institutional response is to invest in forecast accuracy.

That response is incomplete rather than wrong, and the incompleteness is expensive. A forecast is worth what the response mechanism can cash. An accuracy improvement inside a window in which supply cannot move produces no operational benefit at all — the error is merely measured more precisely.

Stated as a proposition: the marginal value of forecast accuracy is bounded by the elasticity of the supply it informs.

The supporting evidence in the operations literature is consistent but indirect, and the pattern of resolution is itself the tell: when operations research confronts arrival-rate uncertainty, it resolves toward flexibility mechanisms rather than toward better point prediction.

  • Bassamboo, Randhawa and Zeevi examine capacity sizing when the mean arrival rate is itself random — a randomness they attribute directly to forecasting error — and find that the classical square-root safety staffing prescription has to be revisited under those conditions.[2] Uncertainty in the rate does not merely make the staffing problem harder; it changes it qualitatively.
  • Koçağa, Armony and Ward state the practitioner's version precisely: staffing must be committed before the arrival rate is known, so the operation either pays for idle staff or loses callers to abandonment. Their remedy is not a better forecast but a flexibility mechanism — a threshold rule that dynamically outsources calls when the system becomes too crowded.[3]

Both analyses begin from forecast error and end at capacity flexibility.

The proposition is offered here as a framing rather than as a citation. It was not located as a named result in the service operations literature during targeted searching; the supporting evidence is the consistent direction of resolution in the two studies above rather than a direct finding. The individual components — arrival-rate uncertainty, learning curves, chaining, contract-induced quality effects — are each well established in their own right.

Sources of low elasticity

Four causes compound in complex servicing.

Time to proficiency is long. Learning by doing is a well-established driver of productivity in this kind of work. Kim, Krishnan and Argote, studying a computing call centre, find measurable learning effects expressed as reductions in average resolution time as experience accumulates, together with knowledge transfer within groups.[4] In travel servicing the effect is amplified, because competence requires GDS fluency, fare construction, policy knowledge and account-specific context. Industry benchmarking commonly places baseline proficiency at 60–90 days and experienced-level performance at six to eight months; these are trade-source benchmarks rather than peer-reviewed findings, and are directionally consistent with the academic learning-curve literature.

Tenure is short, so the ramp is a standing cost rather than a transient one. Arlotto, Chick and Gans formalise the consequence: hiring and retention of workers who learn is a problem in which the value of a worker is a function of accumulated experience, so turnover destroys capital that must be rebuilt at the same slow rate.[5] Where first-year attrition is high, a material share of the pool is permanently mid-ramp. This is an operating condition, not a transitional inefficiency.

Fragmentation destroys fungibility. Work divided into many small pools cannot absorb or donate capacity. Small pools also run structurally lower occupancy at the same service level, so the estate is simultaneously idle and unable to respond — an appearance of slack that cannot be spent. See Multi-Skill Pooling and the Double-Counting Trap.

There is no flexible labour pool at the required skill level. The part-time and contingent supply that retail and hospitality draw on does not exist, unmodified, for work with a six-month competence curve.

The combined effect is an operation that can predict a surge with precision and remain unable to act on it within the horizon of the prediction.

Raising the ceiling from the demand side

The ceiling argument is frequently misread as a case against investing in forecasting. It is better read as a case for changing the instrument rather than the degree of the investment.

A point forecast answers a question no flexibility mechanism can use. A threshold rule of the kind in Koçağa, Armony and Ward,[3] a reallocation policy, or a decision about how much contingent capacity to place on standby all consume a distribution — the probability that demand exceeds a level, and the cost asymmetry either side of it. Improving the central estimate while continuing to publish a single number leaves that mechanism unfed.

Three shifts follow, each already documented on this wiki:

The demand-side upgrade that pays is therefore a change of kind. It raises the ceiling not by reducing error but by producing the input a response mechanism can act on, and by pricing the response itself.

Raising the ceiling from the supply side

Dynamic reallocation

A staffing decision taken before the arrival rate reveals itself carries the full uncertainty; a routing or reallocation decision taken afterwards carries almost none. This is the structural reason the operations literature resolves toward real-time mechanisms.[3] Practically, it covers Skill-Based Routing, Real-Time Schedule Adjustment, overflow and threshold escalation, and the discretionary movement of activities such as training and offline work — see Intraday Management and Real-Time Threshold Alerts and Escalation Protocols.

Reallocation is powerful and cheap relative to its effect, but it is a redistribution instrument. It moves hours already bought; it does not create hours.

Designed fungibility and the chaining result

The flexibility literature is unusually encouraging, and its central result is counter-intuitive enough to be routinely missed in practice.

Jordan and Graves establish that limited flexibility, configured as a chain, captures nearly all the benefit of total flexibility — approximately 98% of the throughput of a fully flexible system, using resources capable of only two tasks.[6] The configuration matters more than the quantity: the benefit comes from connecting resources into a single chain, not from every resource being able to do everything.

Wallace and Whitt reproduce the result in the call centre setting. Where service time does not depend on call type or agent, two skills per agent in the right combinations performs almost as well as every agent holding every skill, and required staffing under limited cross-training is nearly the same as under full cross-training.[7] Bassamboo, Randhawa and Van Mieghem extend it to parallel queueing systems, showing that a tailored chaining configuration using dedicated and level-2 resources is asymptotically optimal, and that in some parameter regions the fully flexible resource is not worth using at all.[8]

Three independent results, across three decades and three settings, converge. The implication for a fragmented estate is direct: fungibility does not require universal cross-training, and the barrier is skill-map design rather than training budget. See Cross-Training and Skill Mix Strategy.

What outsourcing does and does not buy

Outsourcing is routinely justified on two grounds — lower unit cost, and speed to ramp. The first is real. The second requires a distinction that is rarely drawn: a vendor seat fills quickly, but a proficient vendor agent does not arrive quickly, because time to proficiency is a property of the work rather than of the employer. The learning curve applies identically on the other side of a contract.

The distinction that matters operationally is between two arrangements that share a name:

  • Co-sourcing as overflow delivers genuine elasticity, because contacts overflow to capacity that is already standing and already proficient. The mechanism in Koçağa, Armony and Ward is a real-time routing threshold — a decision made after the arrival rate reveals itself.[3]
  • Outsourcing as hiring inherits the buyer's ramp physics unchanged where the partner must recruit and train to meet a surge. It supplies a lower unit cost and no elasticity.

Where the second is procured and the first is expected, a predictable sequence follows: capacity is bought for speed to seat, judged months later on proficiency, and the supplier is held responsible for a shortfall that was purchased deliberately.

A second structural result is worth carrying into contract design. Ren and Zhou analyse the standard forms — piece-meal and pay-per-call-resolved — and show that although they can coordinate the staffing level, the resulting service quality is below the system optimum.[9] Quality shortfall in an outsourced estate is therefore in part a predictable property of the contract form rather than evidence about the supplier — a diagnosis that points at procurement rather than at performance management. See BPO and Vendor Management for WFM and Business Process Outsourcing.

The composition limit

Improving both sides raises the ceiling, but neither instrument escapes a constraint that sits beneath both: the composition of the workforce being planned.

A pool built exclusively of full-time, experienced agents has three structural properties that bound its response regardless of forecast quality or routing sophistication.

  • A fixed hours envelope. Contracted weekly hours are the supply. Inside the forecast horizon there are no marginal hours to buy except overtime, which is bounded by fatigue, cost and consent, and which draws on the same people the surge is already loading.
  • Coarse granularity. The smallest unit of supply is a full-time equivalent. Demand volatility in these environments arrives in shapes — a six-hour disruption peak, a two-week seasonal ridge — that a full-time shift lattice cannot match without buying idle time either side of it.
  • Ramp-gated growth. The only mechanism for adding capacity is to hire, and hiring is gated by the learning curve.[4][5] Growth is therefore available on a horizon of months, while the variance that matters arrives on a horizon of hours to weeks.

The consequence is that distributional forecasting and dynamic reallocation both eventually act on the same fixed quantity of hours. Chaining redistributes capacity; it does not create it. Reallocation moves hours; it does not add them. Each is worth doing, and each has a terminal value set by what is in the pool.

Where both instruments are already in place, workforce composition becomes the binding lever. That reframes staffing-model design from a cost or engagement topic into a capacity one.

Staffing models that add elasticity

The models below add response capability that a uniform full-time pool structurally lacks. Each carries preconditions and costs, and none is universally preferable.

Model Mechanism it adds Principal preconditions and costs
Part-time and micro-shift cores Finer granularity; coverage shaped to the demand curve rather than to a shift lattice Higher per-head overhead for training, systems and supervision; scheduling complexity rises
Annualised hours and seasonal contracts Moves contracted hours across the year toward known seasonal ridges Requires reliable seasonal signal and jurisdictional support; weak against unforecast disruption
Contingent and gig pools Standby capacity callable at short notice; converts fixed cost to variable Proficiency and quality control are the limiting factors; onboarding, security and access management costs recur
Internal marketplace across brands or segments Pools capacity that fragmentation has stranded; no new hiring required Requires a common skill taxonomy, routing that can cross organisational boundaries, and a settlement mechanism between units
Alumni and returner pools Capacity whose proficiency is already banked, attacking the ramp constraint directly Depends on exit relationships and re-certification; supply is finite and not controllable
Blended human and automated capacity Capacity with near-zero marginal ramp time Quality profile varies sharply by task type; see below
Self-scheduling and flexible shift models Voluntary supply released where employee preference and demand shape coincide Yields uplift only where preference and demand curves overlap; needs governance to prevent coverage gaps

The proficiency objection applies to most of these models and is the reason they are frequently rejected. The chaining result is what answers it. If a role can be decomposed so that a chained two-skill scope reaches competence in weeks rather than months, then part-time, contingent and returner supply becomes viable against that scope even where it is not viable against the full role. Fungibility design and staffing-model design are therefore complements: chaining defines the narrower, faster-to-proficiency role that alternative supply models can fill, and those models supply the marginal hours that chaining alone cannot create.

This joint dependency — that flexible staffing models are gated by scope design, and scope design is unrewarded without flexible supply to fill it — is an inference drawn from the chaining[6][7] and learning-curve[4][5] literatures rather than a result reported in either. It is stated here as an argument.

Where automation and AI fit

The category error

Under expense pressure the reflex is to deploy AI to remove supply and bank the saving. That treats AI as a cost lever, and it is the same category error as treating outsourcing as a flexibility lever, committed in the opposite direction. It forfeits the more valuable property.

Mechanism one: elastic capacity

Once built, automated capacity has near-zero marginal ramp time and can be applied to a surge within the horizon in which the surge is forecast — which no human pool subject to a six-month learning curve can do. This is the obvious mechanism and it is genuine. See AI Containment Rate and Its Workforce Implications.

Mechanism two: compression of time to proficiency

The second mechanism is less frequently counted and the evidence for it is stronger.

Brynjolfsson, Li and Raymond studied 5,179 customer support agents through the staggered rollout of a generative AI assistant. Access raised issues resolved per hour by 14% on average — 34% for novice and low-skilled workers, and negligibly for experienced and highly skilled ones. The authors state the mechanism explicitly: the system disseminates the practices of more able workers and moves newer workers down the experience curve. They further observe improved customer sentiment and increased employee retention.[10]

The finding replicates outside customer service. Noy and Zhang, in a preregistered experiment with 453 college-educated professionals on occupation-specific writing tasks, found time taken fell 40% and output quality rose 18%, with the effect compressing the productivity distribution by benefiting lower-ability workers most.[11]

Two independent studies, different domains, the same shape: AI compresses the skill distribution by accelerating the inexperienced. If time to proficiency is the binding constraint on elasticity, then a technology that measurably shortens it functions as an elasticity intervention rather than as a productivity tool, and it interacts directly with the composition limit — a shorter ramp widens the set of staffing models that can clear the proficiency bar. The two effects compound: shorter ramp reduces the cost of each hire, and higher retention reduces the number of ramps required.[10]

The connection between these findings and supply elasticity is a synthesis of the studies above against the learning-curve literature, not a result reported in any of them.

Documented limits

Two studies place firm limits on the claim.

Dell'Acqua and colleagues, in a field experiment with 758 consultants, identify a "jagged technological frontier" — a boundary where some tasks fall easily within AI capability while others, superficially similar in difficulty, fall entirely outside it. On tasks outside the frontier, consultants using AI performed 19 percentage points worse than those working without it, and the authors document mis-calibrated trust: participants over-relied on the system precisely where it was weakest.[12]

Wang and colleagues provide the customer-service analogue in a randomised field experiment on Alibaba's Taobao platform, where treated workers supervised an agentic AI system handling AI-eligible chats. Deployment reduced average chat duration with limited effect on retrial rates, but substantially lowered ratings on AI-eligible chats. Human intervention preserved quality in technical escalations — issues beyond the system's capability — but was markedly less effective in emotional escalations, where customers expressed frustration and where workers themselves showed lower engagement. Early intervention was essential. A positive spillover was observed on the chats the system did not handle, as treated workers redirected attention toward them.[13]

The read for a servicing operation is that automated capacity is real, is not free, and carries a quality cost concentrated in emotionally-loaded interactions — which is what a disruption event generates. See Human-AI Escalation Patterns in Production.

The condition under which the model fails

The argument turns on one empirical question that is rarely asked of an operation's own interaction data: does demand variance live in the simple work or the complex work?

If disruption generates predominantly straightforward re-booking, automation absorbs the variance and acts as a buffer around a stable human core; the elasticity gain is large. If disruption generates complex, emotionally-charged re-accommodation, that work sits at or beyond the jagged frontier,[12] attracts the quality penalty observed at Alibaba,[13] and the human residual becomes simultaneously harder and less fungible.

A related tension is worth holding openly. If automation absorbs the routine work, the remaining human portfolio is more specialised and therefore less fungible than before, so automation can reduce the elasticity of what is left. Whether the net effect is positive depends on where the variance sits. The question is testable from existing interaction data, and testing it before the investment case is written costs materially less than testing it afterwards. See Operational Volatility Index (OVIX).

Implications for operational design

  1. Value forecast investment by the decision it changes. Improvements that move a point estimate where no lever remains open have no operational return; improvements that produce a calibrated distribution feeding a live reallocation rule do.
  2. Design a skill chain rather than pursuing universal cross-training. The evidence indicates two skills per agent, correctly configured, delivers most of the available benefit.[6][7][8] The skill map is an engineering artifact.
  3. Treat workforce composition as a capacity instrument. Where distributional forecasting and dynamic reallocation are already in place, the mix of full-time, part-time, contingent, returner and automated supply is the binding constraint on response.
  4. Measure speed to proficiency explicitly and separately from speed to seat, and contract for it where delivery is outsourced. These are different goods and are commonly bought under one name.
  5. Distinguish co-sourcing as overflow from outsourcing as hiring in every commercial construct. Only the first buys elasticity.
  6. Examine contract form before supplier performance where quality underperforms in an outsourced estate.[9]
  7. Require every automation business case to state its elasticity contribution, not only its cost reduction. A case built solely on deflection optimises the wrong variable and is likely to be unwound when service quality moves.
  8. Locate the variance before committing the investment. Whether volatility sits in simple or complex work determines whether automation raises or lowers the elasticity of the residual human portfolio.

Maturity Model considerations

The practice differs sharply by maturity level.

  • Levels 1–2. Forecasts are point estimates and the workforce is uniformly full-time. Elasticity is effectively fixed, and forecast improvement is the only visible lever — which is why it absorbs disproportionate investment at these levels.
  • Level 3. Statistical and ML forecasting arrive, and intraday automation makes the first reallocation lever real. The ceiling becomes visible: accuracy improves and service outcomes move less than expected.
  • Level 4. Scenario simulation and dynamic skill routing make both instruments available. Composition emerges as the binding constraint, and staffing-model design moves from an HR topic to a planning one.
  • Level 5. Human and automated capacity are planned together, and automation is valued for its elasticity contribution and its effect on time to proficiency rather than for deflection alone.

Use this with Claude

The empirical test that decides which side of this argument an operation actually faces — how much of its demand variability is irreducible, how much its forecast already explains, and how much a better forecast could still reach — is set out in Doubly Stochastic Arrivals and Demand Variance Decomposition.

A ready-to-deploy instruction set and reference files for running it are at Wiki:Packs/Demand Variance Decomposition (CP-WFM-006).

See Also

For broader methodological context on workforce planning practice measured against the underlying theory, see Koole and Li.[14]

References

  1. ↑ Ibrahim, R., Ye, H., L'Ecuyer, P., Shen, H. (2016). Modeling and forecasting call center arrivals: A literature survey and a case study. International Journal of Forecasting 32(3), 865–874.
  2. ↑ Bassamboo, A., Randhawa, R.S., Zeevi, A. (2010). Capacity sizing under parameter uncertainty: Safety staffing principles revisited. Management Science 56(10), 1668–1686.
  3. ↑ 3.0 3.1 3.2 3.3 Koçağa, Y.L., Armony, M., Ward, A.R. (2015). Staffing call centers with uncertain arrival rates and co-sourcing. Production and Operations Management. Preprint: arXiv:1404.2938.
  4. ↑ 4.0 4.1 4.2 Kim, Y., Krishnan, R., Argote, L. (2012). The learning curve of IT knowledge workers in a computing call center. Information Systems Research.
  5. ↑ 5.0 5.1 5.2 Arlotto, A., Chick, S.E., Gans, N. (2014). Optimal hiring and retention policies for heterogeneous workers who learn. Management Science 60(1), 110–129.
  6. ↑ 6.0 6.1 6.2 Jordan, W.C., Graves, S.C. (1995). Principles on the benefits of manufacturing process flexibility. Management Science.
  7. ↑ 7.0 7.1 7.2 Wallace, R.B., Whitt, W. (2005). A staffing algorithm for call centers with skill-based routing. Manufacturing & Service Operations Management 7(4), 276–294.
  8. ↑ 8.0 8.1 Bassamboo, A., Randhawa, R.S., Van Mieghem, J.A. (2012). A little flexibility is all you need: On the asymptotic value of flexible capacity in parallel queuing systems. Operations Research.
  9. ↑ 9.0 9.1 Ren, Z.J., Zhou, Y.-P. (2008). Call center outsourcing: Coordinating staffing level and service quality. Management Science 54(2), 369–383.
  10. ↑ 10.0 10.1 Brynjolfsson, E., Li, D., Raymond, L. (2023). Generative AI at work. NBER Working Paper 31161. Published in Quarterly Journal of Economics 140(2), 889.
  11. ↑ Noy, S., Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science.
  12. ↑ 12.0 12.1 Dell'Acqua, F., McFowland III, E., Mollick, E.R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., Lakhani, K.R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. SSRN 4573321. Published in Organization Science (2026).
  13. ↑ 13.0 13.1 Wang, Y., Zhu, C., Feng, T., Lu, L.X., Jia, B. (2026). Agentic AI and human-in-the-loop interventions: Field experimental evidence from Alibaba's customer service operations. arXiv:2605.14830.
  14. ↑ Koole, G., Li, S. (2023). A practice-oriented overview of call center workforce planning. Stochastic Systems 13(4), 479–495. Preprint: arXiv:2101.10122.
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Supply_Elasticity_in_Workforce_Planning&oldid=3772"