Occupancy and the Three Utilizations

From WFM Labs

Occupancy and the three utilizations are the time-based efficiency measures of a service workforce, defined on the four tiers of paid time set out in The Value of an Hour: doing the work, getting ready, the cost of being human, and the leak. Occupancy is held to a single queue definition and given its two jobs: a planning input that, with the service level objective and the workload, sets how much ready time must be staffed, and an output that shows how well schedules matched demand, judged against the occupancy that the queue's service level and workload allow, an occupancy-maximum computed with Erlang C. Utilization, which published sources define in several incompatible ways, is split into three named ratios, one financial and two operational, each with its denominator stated in its name. The split of ready-idle time into required readiness and excess idle connects the two: it decides how much of a queue's idle is the price of its service level and how much is leak. The page is the first in a series that states each planning definition in two flavors: the legacy flavor, as practised in operations where every arrival reaches a person, and the agentic flavor, for operations where software agents handle part of the arrivals and value-based planning becomes the guiding principle.

About this series: legacy and agentic flavors
Pages in this series restate a planning definition twice. The legacy flavor is the definition as it works today in a human-staffed operation; it is not deprecated, and most functions will run on it for years. The agentic flavor is the same definition read for an operation in which software agents take a share of the arrivals, people handle the residual, and the planning question shifts from "how many people" to "which work is worth a person's hour". Where the two flavors differ, the difference is stated; where they do not, the page says so. The series assumes the ledger on Agent Time Accounting Model and the register on The Definitions Register.

The four tiers on the time ledger

Four tiers of an illustrative paid period, and the three utilizations read off them. The numerator is teal, the denominator grey. The shares are illustrative, not a benchmark.

The four tiers, their value ratings and the automation zone between Tiers 1 and 2 are defined on The Value of an Hour. This section maps them onto the leaf buckets of the Agent Time Accounting Model, so that every ratio below can be computed from a closed ledger by labelling, without new measurement. Every paid hour lands in exactly one tier.

Tier Holds Value Ledger buckets it classifies Typical contents
Tier 1, doing the work Customer-facing or customer-serving productive time High: the hour the operation exists to sell In-work time by channel (ledger tier 6): talk, hold and after-contact work; chat focus time; touch time per item for asynchronous and back-office work Calls, chats, cases, claims, fulfilment, the work a customer would recognize as service
Tier 2, getting ready Enabling work that builds capability Positive: harder to price, but it moves handle time, quality, attrition and readiness Non-productive state classes of the enabling kind (ledger tier 4): training, coaching, development, team meetings, one-to-ones Planned, in-office, usually budgeted; what practitioner taxonomies call internal or discretionary shrinkage
Tier 3, the cost of being human Time the employer pays for because the workers are people Neutral: necessary, neither positive nor negative Off-schedule paid time (ledger tiers 1 to 2): paid leave, public holidays, paid sickness; and the personal state classes of ledger tier 4: breaks, lunch Planned external shrinkage in most taxonomies, plus unplanned absence where the function treats it as neutral
Tier 4, the leak Cost with no output Negative The residual; scheduled-not-logged time (lateness, no-show, early finish); unaccounted state time; system downtime; and ready-idle time beyond what service level requires (see below) Lost time, unexplained absence where the function treats it as negative, idle without assignment

Three rules make the tiers usable.

  • Every minute lands once. The tiers inherit the closure identity of the ledger: paid time equals the four tiers plus the published residual, and the residual sits in Tier 4 until it is explained.
  • The boundary between Tiers 1 and 2 is the automation zone. Software that takes an arrival moves minutes out of Tier 1; software that prepares, summarizes or coaches moves minutes out of Tier 2. The tiers therefore measure what automation did to the hour, not only what people did with it.
  • Ready-idle is split, not assigned whole. On a queue, part of the time an agent spends ready and waiting is the price of the service level, and part is excess. The split is defined in the occupancy section; only the excess is a leak.

Where a function must rule on an ambiguous item, it rules once and records the ruling against the activity code on Agent State Taxonomy and Activity Codes. System downtime recorded on an activity code is shrinkage like any other and negative, so it sits in Tier 4; the leak describes the hour, not the person. The case that usually needs a ruling is unplanned absence: Tier 3 if the function treats absence as a cost of doing business, Tier 4 if it treats it as negative, or Tier 3 up to the allowance built into the plan and Tier 4 above it. The ruling changes the leak number and must be visible.

Occupancy stays a queue measure

Occupancy is the share of an agent's ready time that is spent in work:

occupancy = in-work time ÷ (in-work time + ready-idle time)

where in-work time is talk, hold and after-contact work for voice, and focus or touch time for other channels, and ready-idle is time logged in and available with nothing in hand. Breaks, meetings, training, coaching and every other not-ready state are excluded from both numerator and denominator. This is the one definition that is stable across the academic review of call-centre operations, the industry association's own tutorial, and most practitioner sources, and it is the one that reconciles to queueing theory, where occupancy is the traffic intensity ρ = A ÷ N for an offered load A handled by N agents.[1][2][3]

The definition is deliberately narrow. It says nothing about paid time, shrinkage or value. What makes occupancy unusual among time measures is that it does two jobs.

  • As a planning input, it is the link in the Erlang model between the service level objective and the workload on each queue: together they set how much ready time has to be staffed so that an arriving contact finds someone free.
  • As an output, it shows how well the schedule matched the demand that actually arrived.

The output is where occupancy is most often misread. High occupancy is not evidence of efficiency: when it comes with a missed service level, it usually signals a mismatch of supply and demand, such as too few people for the volume that arrived, a forecast miss, or handle time or shrinkage above plan. Low occupancy with service well above target is the same mismatch in the other direction. Read alone, occupancy rewards the first and penalizes the second; read beside the service level and the occupancy-maximum below, it shows which one occurred. A function that lets not-ready time into the denominator, as at least one widely read practitioner article does, reports an occupancy that cannot be compared with its staffing model, and a function that reports a ratio of in-work to paid time under the name occupancy is reporting a utilization.[4] The confusion is documented on Agent Utilization vs Occupancy Definitions and Tradeoffs.

The occupancy-maximum test

The occupancy a queue can reach at its service level target rises with offered load. A flat guardrail is unreachable for small queues and exceeded by large ones unless they are capped. The three squares are illustrative actual readings.

Because occupancy is a result of service level and random arrival rather than a target in its own right, the useful question is not "is occupancy high" but "is this queue operating at the occupancy its service level and workload allow". Erlang C answers it. For an interval with offered load A (arrivals multiplied by handle time, divided by the interval length) and a service level target, the model gives the smallest number of agents N* that meets the target, and the occupancy implied at that staffing is

ρ* = A ÷ N*

The test compares the occupancy actually observed, ρ, with ρ*. The diagnostic, per queue and per interval of fifteen or thirty minutes:

  1. Compute A from arrivals and handle time.
  2. Find N* as the smallest N at which the Erlang C service level meets the target; set ρ* = A ÷ N*.
  3. Compute actual ρ from agent-seconds in work and ready states.
  4. Record the gap ρ − ρ*, the service level at the actual staffing, and the excess staff N − N*.
  5. Aggregate across intervals by weighting with agent-seconds, never by averaging percentages.
  6. Read a large negative gap as ready time above what service level needs (overstaffing, skill fragmentation or schedule shape), a gap near zero as a queue at its maximum, and a positive gap as understaffing, a missed service level or agents working faster than the handle time assumed.

Three properties of the model matter for the reading.

  • Scale raises the attainable occupancy. Under square-root staffing, N = A + β√A, so occupancy approaches one as load grows while the probability of waiting stays constant.[5] At a handle time of 300 seconds and a target of 80 percent in 20 seconds, a queue offering 10 contacts per half hour reaches about 42 percent, 40 contacts about 67 percent, 160 contacts about 83 percent and 640 contacts about 93 percent. The published staffing tables in the industry association's material show the same curve.[6] A flat cap therefore cannot be right for every queue, which is the argument of The Occupancy Trap.
  • A tighter target costs occupancy. At the same load, moving from 80 percent in 20 seconds to 90 percent in 20 seconds lowers ρ* by several points; relaxing to 80 percent in 60 seconds raises it.
  • The guardrail is a convention. Planning caps of 80 to 90 percent are widely used, and 85 percent is the number most often quoted, but the sources reviewed cite practitioner opinion rather than a study; workload is a recognized antecedent of exhaustion in the organizational literature without a validated threshold, and historical operator pools ran at 90 to 95 percent in busy periods.[7][8] A cap should therefore be published as a policy guardrail beside ρ*, not as a measured maximum.

Caveats: skip intervals with A below about two Erlangs, where integer staffing makes ρ* noisy; Erlang C ignores abandonment, so a queue with material abandonment should use Erlang A or a simulation for N*; multi-skill queues need a pooled or simulated N*; concurrent chat needs an effective handle time per agent slot.

Required readiness and the leak

Two queues with the same in-work time and the same total ready-idle. On the small queue most of the idle is the price of the service level; on the large queue most of it is leak.

The occupancy-maximum test makes operational the distinction The Value of an Hour draws between the slack a queue requires and the leak beyond it. On a queue, some ready-idle is structural: it is the capacity the service level target requires to be waiting when a random arrival comes, and no amount of good management removes it. The ledger therefore splits ready-idle into two buckets:

required readiness = in-work time × (1 ÷ ρ* − 1)
excess idle = ready-idle time − required readiness

Required readiness is costed with Tier 1 (it is the price of selling the Tier 1 hour at the promised service level) or shown as its own line; only excess idle enters Tier 4. Without the split, small queues look leaky by construction, which contradicts the occupancy-maximum test and punishes the queues that cannot be pooled. Back-office work with no arrival queue has no required readiness, and all of its idle is excess. No published shrinkage taxonomy treats ready-idle as shrinkage; the split is new in the sense that it makes the structural part explicit.[9]

Three utilizations: one financial, two operational

Published definitions of utilization disagree on three points: whether ready time counts as utilized (yes in the logged-in and shift-time family, no in the productive-hours family), whether the denominator is paid, scheduled or logged-in time, and whether training and meetings sit in the numerator, in the denominator only, or outside both.[10][11] The same day can therefore be reported at three different utilizations under one name. The remedy is not to pick one but to define the ratios that answer distinct questions, with a naming rule that says what each divides by. The first word names the denominator: financial utilization divides by every paid hour; operational utilization divides by available hours, the paid hours minus Tier 3. The plain name counts the work alone; with readiness adds Tier 2 to the numerator. U1, U2 and U3 are short codes for tables and charts.

Let T1 to T4 be the hours by tier and P the paid hours, with P = T1 + T2 + T3 + T4; available hours are P − T3. On a live queue, T1 includes required readiness (above), so the idle the service level needs counts as work in all three ratios.

Ratio Numerator Denominator Question it answers Reader Nearest published analogue
U1, financial utilization T1 P Of every hour we pay for, how much was the work? Finance; the annual plan; the price check on outsourced supply; the automation case Productive hours over paid hours; TEEP in manufacturing, which measures against all calendar time[12]
U2, operational utilization T1 P − T3 Of the hours people could have worked, how much was the work? Site and operations leaders; like-for-like comparison of sites and partners with different leave regimes Billable utilization in professional services, where leave and holidays are netted out of available hours; OEE, which removes planned downtime before measuring[13]
U3, operational utilization with readiness T1 + T2 P − T3 Of the hours people could have worked, how much went to work or to building the capability to do it? Workforce management leadership; learning and development; the budget line for enabling time "Resource utilization" (productive over available) in professional services; productive hours in healthcare staffing standards that count indirect work as productive

Three identities tie the ratios together and to occupancy.

  • U3 = 1 − T4 ÷ (P − T3). The third ratio is the leak measure: what is not work or enablement, on the available base, is leak.
  • For a voice queue, occupancy × (1 − shrinkage) gives time in work as a share of paid hours when shrinkage is defined on paid hours and occupancy as above. That is U1 before required readiness is added to Tier 1, which is why a function that already reports occupancy and shrinkage can produce most of U1 without new instruments.[14]
  • U1 ≤ U2 ≤ U3 always. A report that shows the three side by side shows, in the gaps between them, the cost of being human and the investment in readiness as fractions of the paid hour.

Worked example

An illustrative team-week of 1,000 paid hours, with the shares from the first diagram and a ready-idle split from the occupancy-maximum test:

Bucket Hours Tier
In work 550 Tier 1
Required readiness (ρ* gives 60 of the 100 idle hours as structural) 60 Tier 1, costed with the work and shown as its own line
Training, coaching, meetings 100 Tier 2
Breaks, lunch, leave, public holidays, sickness 250 Tier 3
Excess idle: ready time beyond what the service level needs 40 Tier 4
Paid 1,000

Lateness, system downtime and unaccounted time also sit in Tier 4; this illustrative week has none, so all 100 ready hours are idle and the test splits them 60 and 40.

Ratio All idle treated as leak Required readiness counted with the work
Occupancy (queue hours only) 550 ÷ (550 + 100) = 84.6 percent unchanged, 84.6 percent
U1, financial utilization 550 ÷ 1,000 = 55.0 percent 610 ÷ 1,000 = 61.0 percent
U2, operational utilization 550 ÷ 750 = 73.3 percent 610 ÷ 750 = 81.3 percent
U3, operational utilization with readiness 650 ÷ 750 = 86.7 percent 710 ÷ 750 = 94.7 percent
Leak share of available hours 100 ÷ 750 = 13.3 percent 40 ÷ 750 = 5.3 percent

The two columns are the point of the readiness split: a function that treats all idle as leak reports a leak of 13 percent on this team, and a function that applies the occupancy-maximum test reports 5 percent and can say where the other 8 points went. Those 8 points are readiness the service level needs; they move into Tier 1 in all three utilizations, and occupancy does not change because it already counts all ready time in its denominator.

Legacy flavor and agentic flavor

The same four tiers read in a human-only operation and in one where software agents take a share of the arrivals. Occupancy survives the transition because the arrivals that reach people are still random.

The four tiers and the three ratios are the same in both flavors. What changes is what fills Tier 1, what Tier 2 prepares people for, and what the planning question is.

Element Legacy flavor Agentic flavor
Tier 1 Handle time on every arrival, high and low value alike. One formula treats them the same Live work that reached a person after a software front-end took its share. The automation zone sits below Tier 1 rather than inside it, and Tier 1 is graded by value (retention, complaints and judgment at the top; routine status and balance work at the bottom) because the grade decides what is routed to a person at all
Tier 2 Training, coaching and meetings, budgeted as a shrinkage percentage Readiness for the work that remains: judgment, exceptions, and the supervision of software agents. A larger and more deliberate share of the human hour, with its own budget line rather than a shrinkage allowance
Tier 3 Breaks, lunch, leave Unchanged. People still need breaks and leave. The share of the human hour rises as Tier 1 shrinks, which is a reason to report U2 beside U1
Tier 4 Lost time and all idle beyond the schedule Excess idle only. Required readiness is costed, because the residual arrivals that reach people are fewer, more variable and more valuable, so the readiness to take them is a deliberate purchase
Occupancy Handle over handle plus ready, tested against Erlang C at the queue's service level The same definition and the same test. A software front-end changes the arrival rate and the mix that reach the human queue, and therefore N* and ρ*, but the arrivals that reach people remain random, so the occupancy a human queue can sustain at its service level is still an Erlang question. Software agents are capacity planned on a different model (they do not have a ready state in the human sense); the human queue behind them is planned as before, with a smaller and lumpier load
Utilization U1 for the plan, U2 for site comparison, U3 for the leak, if a function publishes more than the one it has All three, because the agentic case is argued in Tier 1 value per paid hour (U1) and defended in readiness investment (U3). A function that reports only U1 will under-invest in Tier 2 exactly when the work left to people needs it most
Planning principle Minimize paid hours for a service level Maximize value per human hour at a service level: route by the value and readiness grade of the work, then staff the human residual with the same queueing discipline

The agentic flavor does not retire occupancy. It narrows the population occupancy applies to (the people on a live queue), raises the value of each unit of their in-work time, and makes required readiness a line that a planner defends rather than a leak the planner apologizes for. The pages that describe the routing and handover side of the agentic case are Agentic AI Workforce Planning and The Agentic Handover Gate.

Adopting the definitions

  • Never publish "utilization" unqualified. The name says the denominator: financial (paid hours), operational (available hours), operational with readiness. A register entry per ratio, with numerator and denominator stated in ledger tiers, on The Definitions Register.
  • Publish U2 and U3 together. Each flatters something the other exposes: U2 hides leave differences between sites, U3 hides the cost of enabling time. Side by side they show the choice.
  • Keep unplanned absence visible. Remove only contractual, planned leave and breaks when computing P − T3; net out sickness and no-shows silently and U2 flatters the site with the worst attendance.
  • Grade every number. Occupancy from agent-state seconds is measured; ρ* is calculated; a tier share from a schedule rather than a state log is estimated; a leak that includes an unexplained residual is assumed, and carries that grade on Human Gates and Number Grades.
  • Reconcile before reporting. The tiers sum to paid or the report does not go out; the tests on Reconciliation Tests for Time Accounting apply.
  • In contracts, map every state code to a tier. Under hours-billed outsourcing, leave and holidays sit with the supplier and coaching is often billed separately, so a supplier's utilization is not comparable to the function's own unless both are restated on the same tiers and the same denominator.

Maturity Model Position

Four scales on this wiki use the word level; the launch page states which is which. On the WFM Labs Maturity Model, a Level 2 function can produce occupancy and U1 from its telephony and payroll systems; Level 3 closes the ledger, labels the four tiers and runs the occupancy-maximum test per queue; Level 4 reports U2 and U3 with the readiness split and uses the gap to ρ* as a routine planning signal; Level 5 plans software agents and people on one value-graded Tier 1 and treats required readiness as a purchased line.

See Also

References

  1. ↑ Gans, N., Koole, G. and Mandelbaum, A. (2003). "Telephone Call Centers: Tutorial, Review, and Research Prospects". Manufacturing & Service Operations Management 5(2), 79–141. Occupancy is defined as total active over total active plus total available; ρ = R/N is called "average system utilization or occupancy". https://aviman.technion.ac.il/files/References/Gans-Koole-Mandelbaum-CCReview.pdf
  2. ↑ Cleveland, B. and Harne, D. (2003). Call Center Metrics and KPIs, ICMI tutorial TU00002. "Occupancy is a result of service level and random call arrival and is not a driver of service level or budgets." https://www.icmi.com/files/StudentResourcePage/CCF/CCMetricsKPIs.pdf
  3. ↑ Rickwalder, R. (2015). "The Ugly Truth About Agent Occupancy (or Why 85% May Be Too Low)". Contact Center Pipeline, 17 Sep 2015. Chooses handle over handle plus waiting "because it allows me to reconcile back to my Erlang calculations". https://blog.contactcenterpipeline.com/2015/09/the-ugly-truth-about-agent-occupancy-or-why-85-may-be-too-low/
  4. ↑ LaReau, J. (2012, updated 2018). "Constructive Use of Occupancy". ICMI. Counts breaks, training and meetings as "occupied" time; the outlier among the sources reviewed. https://www.icmi.com/resources/2012/constructive-use-of-occupancy
  5. ↑ Halfin, S. and Whitt, W. (1981). "Heavy-Traffic Limits for Queues with Many Exponential Servers". Operations Research 29(3), 567–588.
  6. ↑ Cleveland, B. Harnessing Contact Center Dynamics, ICMI session 403. Staffing table at 210 seconds handle time and 80/20: 50 contacts per half hour needs 9 agents at 65 percent occupancy; 1,000 needs 124 at 94 percent. "When service level goes up occupancy goes down"; "Larger groups have higher occupancy". https://www.bradcleveland.com/wp-content/uploads/2025/03/Session-403-Brad-Cleveland-Harnessing-Contact-Center-Dynamics.pdf
  7. ↑ Aksin, Z., Armony, M. and Mehrotra, V. (2007). "The Modern Call Center: A Multi-Disciplinary Perspective on Operations Management Research". Production and Operations Management 16(6), 665–688, citing Cordes and Dougherty (1993) on work demands and emotional exhaustion. https://www.columbia.edu/~ww2040/4615S13/AAM07.pdf
  8. ↑ Coppell, A. (2020). "Why Occupancy Should Not Exceed 85%". Call Centre Helper. Expert opinion and anecdote; no study cited. https://www.callcentrehelper.com/occupancy-not-exceed-85-157102.htm
  9. ↑ Pearce, R. (2023). "How to Calculate Contact Centre Shrinkage". Call Centre Helper. External and internal shrinkage; waiting time is not shrinkage. https://www.callcentrehelper.com/how-to-calculate-contact-centre-shrinkage-90353.htm
  10. ↑ Mitchell, J. (2018). "What is the difference between occupancy and utilisation?" Call Centre Helper. Utilisation = logged-in time ÷ shift time. https://callcentrehelper.com/difference-occupancy-utilisation-132518.htm
  11. ↑ Two workforce-management vendors' knowledge bases define utilization as productive hours over paid hours, one adding "other productive activities" to the numerator and one keeping training and meetings in the denominator only (retrieved Wed 21 Oct 2026). The vendors are not named here; the point is the disagreement.
  12. ↑ OEE reference site, "Calculate OEE" and "TEEP". Availability = run time ÷ planned production time, where planned production time is the shift minus breaks; TEEP = OEE × utilization, measured against all time. https://www.oee.com/calculating-oee/ and https://www.oee.com/teep.html
  13. ↑ Professional-services utilization calculators state the choice explicitly: the same 28 billable hours read 70 percent on a gross 40-hour base and 87.5 percent net of 8 hours of leave, and one source advises stating "whether the denominator is gross capacity, net working time, or recorded hours" (retrieved Wed 21 Oct 2026).
  14. ↑ Watson, M. (2019). "Call Center Shrinkage". A workforce-management vendor's blog. Utilization = occupancy × (1 − shrinkage); 80 percent and 30 percent give 56 percent. https://blog.peopleware.com/forecasting/call-center-shrinkage
Retrieved from "https://wiki.wfmlabs.org/w/index.php?title=Occupancy_and_the_Three_Utilizations&oldid=5619"