Pooling Architecture in Service Workforces

The dedication penalty computed from Erlang C: it is a function of per-account volume, rises as the service target tightens, and saturates well below the square-root bound as the pool is split further. Sawtooth is integer staffing, not noise.

Service organisations sell exclusivity on a three-rung ladder — dedicated, designated, shared — and the middle rung does not survive contact with the mechanism. Dedicated and shared name a real, enforceable property: whether capacity is reserved to one account or available to many. Designated names no parameter at all. It is affinity presented as exclusivity and priced as eligibility.

The deeper problem is that the ladder bundles three things that are not the same thing. Exclusivity is commercial and binary. Eligibility is an operational hard constraint on who can take a piece of work. Affinity is a soft routing preference over who does take it when several people qualify. They have different mechanisms, different costs and different owners. Bundled into one ladder, none of them can be priced, and placement decisions — which work goes where, served by whom, under what commitment — become unresolvable, because the decision variable is a marketing tier rather than a set of independently settable parameters.

This page separates the three dimensions, re-derives the cost of exclusivity from the queueing model rather than asserting it, and argues that the width of a pool is set by a knowledge-capacity constraint rather than by service policy. It closes on what a constraint-based placement model needs from each dimension.

The three dimensions

Exclusivity

Exclusivity asks whether a unit of capacity is reserved to one account or shared across many. It is binary and commercial. It is what appears in a contract, what a buyer believes they are purchasing, and the only one of the three that carries a price today.

Its cost is the pooling penalty, and that cost is computable. Reserving capacity removes it from the shared pool, and a queueing system pays for every partition it creates. The size of the payment is derived in § Pooling economics below.

Exclusivity has no operational content. It says nothing about who is capable of the work, how long they take to become capable, or who handles a given piece of it. A dedicated pool of people unable to do the work is still dedicated.

Eligibility

Eligibility asks who can take a given piece of work. It is a hard constraint, set by skill, complexity, credentialing, regulatory permission, language, system access and account-specific process knowledge. Work arriving at a person outside the eligible set is not slow — it is unservable.

Eligibility shrinks the pool, so it costs the full safety term. A constraint that halves the eligible population for a class of work has the same staffing consequence as splitting the pool in two, and is paid for at the same rate. This is the expensive dimension, and it is almost never priced, because it is treated as an operational fact rather than a commercial choice.

Affinity

Affinity asks who preferentially takes work when several people are eligible. It is a soft ordering in the routing configuration: prefer this group, fall through to the wider pool when the preferred group is busy.

Affinity shrinks nothing. Because the fall-through is unconditional, the eligible population is unchanged and the safety term is unchanged. Its cost is second-order — a small loss from occasionally routing to a person who is not the marginal best choice, and the configuration overhead. What it buys is continuity: repeat contact with the same people, faster recognition of a returning problem, and an account-level relationship that survives individual interactions.

Affinity is the cheapest of the three and delivers most of what buyers of exclusivity say they actually want.

Dimension	Question	Type	Effect on pool size	Cost	Natural owner
Exclusivity	Reserved to one account?	Commercial, binary	Partitions the pool	Full pooling penalty	Commercial
Eligibility	Who can take it?	Operational, hard	Shrinks the eligible set	Full safety term	Operations / capability
Affinity	Who preferentially takes it?	Operational, soft	None	Near zero	Routing / service design

Why designated is not a tier

A tier is a commitment. A commitment that cannot be breached is not a commitment.

Designated commits that an account has a group of people who usually handle its work. There is no bound on "usually". No parameter is named, so no threshold can be tested, so no breach can be detected and no remedy can be triggered. A buyer paying a premium for designated capacity is paying for an intention. A provider selling it has sold something it cannot fail to deliver — which sounds like an easy obligation and is in fact the problem, because the price is real and the obligation is not.

Priced as eligibility, delivered as affinity: the premium is set as though the pool were being narrowed, while the implementation narrows nothing. When the arrangement is honoured — the preferred group really does take almost everything — the provider has silently created a near-dedicated pool and is carrying the dedication penalty without having priced it. When it is not honoured, the buyer has paid for continuity they did not receive. The tier is unstable in both directions.

The routing construct underneath is sound

The honest origin of designated is a routing pattern: agents carry a home queue as a primary skill and one or more adjacent queues as a lower-priority reserve skill, so they serve their own work first and absorb neighbouring work when idle. That construct is not a workaround. It is chaining.

Jordan and Graves (1995) established the central result in manufacturing: resources able to perform only two tasks, connected so the assignments form a single closed chain, capture close to the full benefit of resources able to perform every task. Limited, deliberately configured flexibility is nearly as good as universal flexibility, and vastly cheaper. Iravani, Van Oyen and Sims (2005) generalised the structural-flexibility argument beyond manufacturing, and Wallace and Whitt (2005) reproduced the result in a staffing context: with skill-based routing, giving each person a small number of skills — as few as two — achieves service close to that of a fully cross-trained workforce.

Primary-plus-reserve routing is therefore a well-founded design, not a compromise. What failed was writing a dynamic routing configuration into a static commercial commitment. A routing preference is meant to move — its whole value is that it yields under load. A contractual tier is meant to hold. Selling the first as if it were the second produces a tier that must either break its own mechanism or fail to mean anything.

The correction is not to abandon the routing. It is to stop selling it as exclusivity. Sell affinity as affinity, honestly priced at close to nothing, and sell exclusivity separately where a buyer genuinely wants reserved capacity and will pay the pooling penalty for it.

For the topology and design rules, see Chaining and Flexibility Design; for the routing mechanics, Skill-Based Routing and Multi-Skill Routing in WFM.

Pooling economics

The mechanism

Staffing a queue to a service target requires roughly the offered load plus a safety term:

$N\approx A+\beta \sqrt{A}$

where $A$ is offered load — arrival rate multiplied by average handle time, in erlangs — and $\beta$ is a quality-of-service coefficient. This is the square-root staffing law, formalised in the many-server heavy-traffic regime by Halfin and Whitt (1981) and developed into a dimensioning method by Borst, Mandelbaum and Reiman (2004). The first term is work that must be done. The second is capacity held against variability, and it is the only part that pooling can recover.

The consequence is immediate. Divide one pool into $k$ pools and the offered load divides $k$ ways, but the safety term is paid $k$ times over a base that has shrunk by a factor of $k$. Whitt (1999) analysed exactly this decision — partitioning customers into separate service groups — and Mandelbaum and Reiman (1998) characterised when pooling helps and when it does not in queueing networks. Pooling is not free in every direction: it can lengthen the tail for work that would have been served quickly in a small dedicated group, and it presumes the merged population is genuinely eligible for all the merged work. But on the staffing requirement, the direction is unambiguous.

See Pooling Theory for the general result and Multi-Skill Pooling and the Double-Counting Trap for the corresponding error in the opposite direction.

Re-deriving the penalty curve

The square-root law is an approximation, and under a service-level constraint rather than a pure delay-probability constraint the coefficient $\beta$ is not constant — fitting it across loads from 5 to 1,600 erlangs at an 80%-in-20-seconds target gives values falling from 1.34 to 0.43. A single-$\beta$ shortcut would therefore misstate the answer. The figures below are computed exactly instead, from Erlang C evaluated through the numerically stable Erlang B recursion, a standard formulation in the call-centre operations literature:

$B(0,A)=1,B(n,A)=\frac{AB(n-1,A)}{n+AB(n-1,A)}$

$C(N,A)=\frac{B(N,A)}{1-\frac{A}{N}(1-B(N,A))},\mathrm{S}\mathrm{L}(N)=1-C(N,A)e^{-(N-A)t/\mathrm{A}\mathrm{H}\mathrm{T}}$

Assumptions, stated in full: Poisson arrivals at a constant rate; exponential handle times; infinite patience, so no abandonment; no retrials; a stationary period, so no intraday shape — a material simplification, since time-varying arrivals change staffing requirements substantially; integer staffing; identical service targets and identical handle times across accounts; and no shrinkage, absence or occupancy ceiling — the numbers are pure queueing requirements, not rosters. Each account is assumed to have exactly the same offered load, which is the most favourable case for dedication; unequal accounts do worse. Where a figure below is quoted, average handle time is 300 seconds and the target is 80% answered in 20 seconds unless stated.

The dedication penalty is defined as the extra staff required to serve $k$ accounts in $k$ dedicated pools rather than one shared pool, as a percentage of the shared requirement.

Dedication penalty, one pool split into k = 10 dedicated pools. AHT 300 s, 80% in 20 s.

Contacts/hour per account	Offered load a (erlangs)	Staff per dedicated pool	Total dedicated (×10)	Pooled load	Pooled staff	Penalty
2	0.17	1	10	1.7	4	150%
12	1.00	3	30	10.0	14	114%
24	2.00	4	40	20.0	25	60%
48	4.00	7	70	40.0	46	52%
96	8.00	12	120	80.0	88	36%
192	16.0	20	200	160	170	18%
768	64.0	71	710	640	654	8.6%
3,072	256	267	2,670	2,560	2,578	3.6%
12,288	1,024	1,039	10,390	10,240	10,261	1.3%

The curve is not smooth. Integer staffing produces a sawtooth, sharply at low loads where a single agent is a large fraction of the requirement, and at loads below roughly 0.1 erlangs per account the penalty reaches its arithmetic maximum of $k-1$ — ten pools of one person each against a single person serving everything, a 900% penalty. That regime is not exotic. It is the regime of the small account.

Two properties follow. First, the dedication penalty is a function of per-account volume, not of total volume. A provider with very large aggregate volume and small individual accounts sits at the expensive end of the curve; total scale does not rescue it, because the partition is per account. Second, there is a break point. Holding the assumptions above:

• below roughly 5 erlangs per account (≈57 contacts/hour), exclusivity costs more than 50% extra staff;
• below roughly 13 erlangs (≈154/hour), more than 25%;
• above roughly 59 erlangs (≈705/hour), the penalty stays under 10%;
• above roughly 161 erlangs (≈1,937/hour), under 5%.

Above the break point, exclusivity is close to free and should be sold cheaply and often. Below it, exclusivity is frequently the single most expensive line in the operating model — and it is routinely conceded in negotiation at no price at all, because nobody has computed this curve for the account in question.

Splitting further has diminishing marginal damage. At 2 erlangs per account, splitting into 2 pools costs 14%, into 5 costs 43%, into 10 costs 60%, into 50 costs 85%. The penalty saturates well below the loose $\sqrt{k}-1$ bound implied by a fixed-$\beta$ square-root law, because under a service-level constraint the pooled requirement enjoys more than square-root benefit — the exponential term relaxes as the pool grows. The practical reading is that the first partition is the expensive one relative to what it buys, and the twentieth is nearly free. Providers already committed to per-account dedication have little marginal cost reason to resist one more; the decision that mattered was the first.

Two secondary results

The penalty rises as the service target tightens. At 2 erlangs per account and k = 10:

Target	Staff per dedicated pool	Total dedicated	Pooled staff	Penalty
70% in 30 s	4	40	24	67%
80% in 20 s	4	40	25	60%
90% in 20 s	5	50	26	92%
95% in 15 s	6	60	28	114%
99% in 10 s	7	70	32	119%

A tighter target increases the safety term, and the safety term is what dedication multiplies. Exclusivity and responsiveness are therefore not independent concessions. Granting both compounds; the second is more expensive than it looks because the first is already in force. See Service Level Target Selection and Erlang Sensitivity and the Staffing Cliff.

The penalty rises as handle time falls. Holding contacts per hour constant at 24 per account and k = 10:

AHT	Offered load a	Staff per dedicated pool	Total dedicated	Pooled staff	Penalty
600 s	4.00	7	70	47	49%
450 s	3.00	6	60	36	67%
300 s	2.00	4	40	25	60%
180 s	1.20	3	30	16	88%
60 s	0.40	2	20	6	233%

This is a coupling that is almost never accounted for. Automation that strips handle time makes exclusivity more expensive. Shorter handle time means less offered load per account, which means a smaller pool, which means a larger safety term as a proportion — the pool moves left along the penalty curve into the expensive region. The same is true of deflection, which removes volume rather than time but has the same effect on $A$. An automation business case built on handle-time reduction, in an estate committed to per-account dedication, is systematically overstated: part of the saving is consumed by a rising dedication penalty that nobody booked. The sawtooth in the table is integrality; the trend is the mechanism.

The corollary is a sequencing rule. Automation and exclusivity commitments should be decided together, because automation moves the break point. Work that comfortably supported a dedicated pool before automation may not support one after.

Pool width is a knowledge-capacity constraint

The number of accounts a single pool can serve is treated as a service-design choice. It is not. It is a capacity constraint on what a person can hold in usable, current memory.

The constraint is roughly multiplicative: accounts × complexity-per-account must stay under what a person can learn and keep current. Few accounts when each is intricate; many when each is not. This is why the same organisation supports pools of three accounts in one part of the estate and thirty in another without either being a policy decision — the work differs in complexity, and the width adjusted itself.

[ASSERTED] This framing is reasoned, not measured. The multiplicative form is a hypothesis; the constraint could equally be additive in distinct procedures, or dominated by a small number of high-frequency exceptions rather than by total account count. It is grounded in the cognitive-load literature — Sweller (1988) on the limits of working memory under problem-solving load — and in the observed shape of real skill maps, but no published study establishes an accounts-per-person capacity function for service work. What would settle it: hold error and escalation rates fixed, vary the number of accounts assigned to matched cohorts, and find the width at which quality degrades. Until that exists, this section is a model, not a result.

Two consequences follow if the framing holds, and both are consequential enough to be worth acting on under uncertainty.

Selling a fixed pool width strands the investment meant to change it

Abstraction layers, unified desktops, workflow standardisation, guided procedures and knowledge tooling are all capital spent on exactly one parameter: raising the number of accounts a person can hold. That is what "reducing complexity" means operationally.

A commercial commitment to a fixed pool width — this account is served by a group that serves no more than n accounts — is a contractual commitment against a parameter the organisation is simultaneously paying to move. It freezes the denominator of the ratio the investment exists to increase. The tooling can succeed completely and the operating model captures nothing, because width was sold as a guarantee rather than held as a variable.

Where a buyer genuinely requires bounded width, it should be priced as what it is: an option written against future productivity, not a service characteristic.

The right benefit metric for complexity-reduction tooling

If pool width is the constrained parameter, then accounts per person at held error and escalation rates is the correct denomination of benefit for complexity-reduction investment — better than handle-time reduction, for three reasons.

First, it measures the constraint that actually binds. Handle time is a symptom; width is the structural variable that determines whether pooling is available at all.

Second, it does not fight the pooling curve. Handle-time reduction shrinks offered load and, under dedication, pushes the pool toward the expensive end of the penalty curve. Width increases do the opposite: they enlarge the set of work a pool can absorb, which is a move along the curve toward cheaper territory.

Third, it connects to hiring. Simpler work widens the eligible labour pool — shorter time to proficiency, lower entry requirements, more locations qualified. That is a throughput and cost effect on recruitment that a handle-time metric does not capture at all. See Speed to Proficiency Curve and Cross-Training and Skill Mix Strategy.

The metric must be held at constant quality. Accounts per person rises trivially if error and escalation rates are allowed to drift, which is precisely the failure the constraint predicts when width is pushed past capacity.

Knowledge depreciation is the same problem from the other side

Pool width is bounded by what a person can keep current, and service knowledge does not stay current on its own. Darr, Argote and Epple (1995), studying productivity across franchise service units, found that knowledge acquired through experience depreciates rapidly and transfers within an ownership boundary but not across it — units under common ownership learned from each other; units under different ownership did not, despite doing identical work. Argote and Epple (1990) document the same depreciation pattern in manufacturing learning curves. Szulanski (1996) attributes much of the failure to transfer to the recipient's absorptive capacity and to the causal ambiguity of the knowledge itself, rather than to unwillingness to share.

Three implications for pooling:

• Codification is the only defence. Knowledge held only in individual experience depreciates and cannot be transferred, so it caps width permanently and disappears with attrition. Codified knowledge is the mechanism by which width can be raised at all. See Knowledge Management.
• Ownership boundaries block the transfer that width depends on. A pool spanning an ownership boundary — different providers, different entities — will not accumulate shared knowledge the way an internally pooled group does, so the achievable width is lower across the boundary than within it. This is a structural argument for placing pooling boundaries at ownership boundaries rather than across them.
• Rotation has a cost the roster does not show. Moving people between accounts resets partially depreciated knowledge. Affinity, which raises repeat exposure without constraining the eligible pool, is a cheap defence against that decay — and is a second reason to keep affinity even after exclusivity is unbundled from it.

The measurement grain is a hidden pooling decision

Where a service level is measured determines whether capacity can pool.

Measured across a team or a queue, capacity pools freely — a person can serve any work in scope, and the metric is satisfied by aggregate responsiveness. Measured per account, it cannot. A per-account service level is enforceable only if the organisation can guarantee per-account responsiveness, which requires either reserved capacity or a routing priority strong enough to be equivalent to it. The account has been dedicated, without anyone deciding to dedicate it, and the dedication penalty above is now being paid.

This is the single clearest case of a commercial choice silently determining an operational cost. Measurement grain is recorded as a responsiveness parameter — a reporting detail, negotiated by people who reasonably regard it as one — and is almost never recognised as a staffing commitment. It appears in no capacity model as an input. Yet at low per-account volume it can be worth more than the headline exclusivity tier, because it produces the same partition at no price.

Two practical consequences. A shared arrangement with per-account service measurement is not shared; it is dedicated with the price omitted. And the cheapest concession available in most negotiations is to grant a tight service target at a coarse measurement grain, which delivers real responsiveness while leaving the pool intact.

Two pricing models are two products

The same word — an outsourced or shared service arrangement — covers two commercial structures with opposite risk positions.

Fixed complement. The buyer purchases a specified number of people and accepts whatever throughput results. Idle risk sits with the buyer. Capacity is an input they own, so the arrangement is off the pooling curve entirely: pooling gains, if any, accrue to the buyer as higher throughput from the same complement, not to the provider as lower cost. Under this structure exclusivity is nearly meaningless as a commercial concept — the buyer has already bought the people.

Per transaction. The buyer pays for units of work and the provider decides how to staff. Idle risk sits with the provider, who is fully on the pooling curve: every point of the dedication penalty is a direct margin cost, and every pooling gain is retained. Under this structure exclusivity is expensive and should be priced explicitly.

Two products, radically different economics, one vocabulary. The failure mode is applying a tier structure designed for one to the other — most commonly, selling exclusivity premiums into a fixed-complement arrangement where the buyer already carries the risk that the premium supposedly compensates, or granting exclusivity free in a per-transaction arrangement where it lands straight on margin. Related: Performance-Based Vendor Allocation Design and BPO and Vendor Management for WFM.

Work shape and exclusivity are orthogonal

Commercial tiers routinely bundle how hard the work is with how exclusively it is served, as though complex work must be dedicated and simple work must be shared. The two are independent.

• Work shape — complexity, judgment content, consequence, synchronicity, time to proficiency — determines who is capable. It drives eligibility, ramp time, and which delivery location or node qualifies to do the work at all. See Service Chain Decomposition and Node Sourcing.
• Exclusivity determines whether the work can be pooled. It drives cost, through the penalty curve above.

Buying one does not imply the other. Decoupled, four things become possible that the ladder forecloses:

1. A named-person experience delivered cheaply on standardised work. High affinity, wide eligibility, no exclusivity. The buyer gets continuity; the provider keeps the pool. This is the product most buyers of "designated" were actually trying to purchase.
2. Complex work pooled. Where a buyer wants competence rather than exclusivity, hard work can sit in a narrow-eligibility pool shared across accounts. Eligibility is tight, exclusivity is zero, and the pooling gain is retained precisely where scarce skill makes it most valuable.
3. A price list for exclusivity. Once the penalty curve is computed per account, exclusivity has a number attached and stops being a free concession granted late in a negotiation.
4. Computable placement. Work shape maps to node and location; exclusivity enters the model as a separate constraint. Placement becomes an optimisation rather than a judgment call.

[VERIFY] The claim that commercial tiers in service contracting generally bundle these dimensions is a synthesis from the structure of the ladder and from how the tiers are described, not from a systematic survey of contract language. It should be re-derived from primary contract structures before being treated as settled.

Segment is a poor proxy for work shape

Placement decisions frequently fall back on segment or vertical as a stand-in for how the work behaves. It is a weak proxy. Segment is a go-to-market construct — it groups buyers by how they are sold to and who owns the relationship — and it is being asked to carry an operational decision it was never built for. Within any segment, work spans the full range of complexity, synchronicity and consequence; across segments, work of identical shape recurs.

Where a segment does predict work shape, the prediction is carried by an underlying operational variable — regulatory regime, transaction size, channel mix — which can be measured directly and used directly. Use the variable, not the segment.

Coverage floors

For any queue that must be staffed continuously, the binding constraint is the clock, not the volume.

An average month contains 730.5 hours. Continuous single-person coverage therefore requires 730.5 person-hours per month before any absence cover, holiday, training or attrition buffer. At 40 paid hours per week — 173.9 hours per average month — that is 4.20 people at zero shrinkage. At 30% shrinkage, 121.7 productive hours per person per month, it is 6.0 people. That is the floor for one continuously covered queue with a single person on at any moment, and it is independent of volume: it is the same whether the queue receives one contact an hour or none.

Business-hours coverage of ten hours a day, five days a week is 217.4 hours per month — 1.79 people on the same shrinkage assumption. The ratio between the two is 730.5 ÷ 217.4 = 3.36×, and it is invariant to the shrinkage assumption, which cancels.

The practical implication: any viability threshold expressed as a single volume number is wrong. Take a rule of thumb that a queue needs enough volume to keep its coverage complement reasonably occupied — say 70% — at 300-second handle time. Continuous coverage requires about 6,100 contacts per month to reach that; business-hours coverage requires about 1,800. These are the same rule producing answers a factor of 3.4 apart, and both are commonly quoted as though they were one number. State viability as a pair, keyed to the coverage model, or do not state it.

Two further notes. The 3.4× ratio understates the real gap, because continuous coverage additionally requires premium hours, a larger relief factor for single-point-of-failure protection, and supervision at hours when supervision is scarce. And below the volume threshold, a continuously covered queue is idle most of the time regardless of how well it is managed — no forecasting or scheduling improvement addresses a constraint set by the clock. The only structural fixes are to widen the pool so the coverage complement carries other work, to relax the coverage window, or to remove the requirement for live coverage.

Language coverage is a coverage-floor problem, not a speed problem

Where work can be made language-agnostic — asynchronous channels served through translation, so that any qualified person can handle any language — the gain is routinely justified as agent speed or headcount efficiency. That is the smaller effect and the wrong one.

The structural gain is removal of the coverage floor. A language served live must be covered during its hours by people who speak it, which imposes a separate 6.0-person or 1.79-person floor per language regardless of that language's volume. Nine languages at low volume each is nine floors. Made language-agnostic and asynchronous, they collapse into one pool with one floor — and the pool moves right along the penalty curve, gaining the pooling benefit as well. The saving is a different kind from a handle-time saving, and where volumes per language are small it is usually much larger.

The constraint is that this works only where the channel tolerates asynchrony and where translation quality is adequate for the consequence of the work. Synchronous, high-consequence work in a language keeps its floor.

What a placement model needs from this

A constraint-based placement model — one that computes where work should be served rather than selecting a tier — needs each dimension as a separate input:

Input	Source	Role in the model
Work shape	Measured per work type: complexity, synchronicity, consequence, time to proficiency	Determines eligible nodes and ramp cost
Eligibility constraints	Skill, credential, regulatory, language, system access	Hard constraints; shrink the feasible set
Exclusivity requirements	Commercial, per account, with a price	Partition constraints; cost via the penalty curve
Affinity preferences	Service design, per account	Soft objective term; near-zero cost
Measurement grain	Contractual, per account	Determines whether pooling is feasible at all
Pricing structure	Commercial, per account	Determines who carries idle risk, and therefore who owns the pooling gain
Coverage window	Contractual, per queue	Sets the floor complement independently of volume
Pool width capacity	Measured: accounts per person at held quality	Bounds feasible pool composition

The last row is the one that does not exist in most organisations. Every other input can be read off a contract or a routing configuration. Accounts per person at held error and escalation rates has to be measured, and until it is, pool composition remains a judgment call dressed as a policy.

This is where the dimensions become model inputs rather than commercial vocabulary, and it corresponds to the top of the WFM Labs Maturity Model — the level at which placement is computed from business objectives rather than chosen from a menu. Below that level the three dimensions can still be separated and priced independently, which is worth doing on its own terms; the model is what makes the separation operational.

Open questions

• Does the multiplicative width constraint hold? The accounts × complexity form is asserted. A cohort study varying account count at held error and escalation rates would settle it. [ASSERTED]
• What is affinity actually worth? The claim that affinity delivers most of the continuity benefit buyers seek from exclusivity is plausible and untested. Settling it requires a controlled comparison of satisfaction and resolution outcomes under affinity routing versus true dedication at matched volume.
• How much does unequal account size change the penalty? The curves here assume identical accounts, the most favourable case for dedication. Heterogeneous portfolios should do worse, and by how much is computable but not computed here.
• Where does the abandonment model change the conclusion? Erlang C assumes infinite patience. With abandonment, small pools shed load rather than queueing it, which flatters dedication on service level while degrading it on outcomes. The direction is known; the magnitude is not addressed here. See Erlang-A.

See also

• Pooling Theory
• Chaining and Flexibility Design
• Multi-Skill Pooling and the Double-Counting Trap
• Skill-Based Routing
• Service Chain Decomposition and Node Sourcing
• Cross-Training and Skill Mix Strategy
• Service Level Target Selection
• Erlang Sensitivity and the Staffing Cliff
• Vendor Operating Model Maturity
• Performance-Based Vendor Allocation Design
• Knowledge Management
• WFM Labs Maturity Model™

References

On the computations. Every figure in this article is computed from the Erlang C model stated in § Re-deriving the penalty curve, under the assumptions listed there, and from the coverage arithmetic in § Coverage floors. None is taken from a published measurement or a vendor figure. The parameters are swept rather than assumed, so the results are curves rather than points; readers substituting their own handle time, service target and account volumes should re-derive rather than transfer the percentages.