The BPO Delivery Benchmark Report

The benchmark problem

Every contact-centre and outsourcing leader has been handed a slide with a number on it and a verdict attached. Your attrition is 34%; the benchmark is 30%, so you are failing. Your occupancy is 79%; the benchmark is 85%, so you are leaving money on the table. Your utilisation is 64%; the benchmark is 70%, so your consultants are underworked. The number arrives without its birth certificate, and the verdict follows without argument. This report is written against that slide.

The trouble is not that benchmarks are useless. It is that most of them are quoted with none of the three things that would make them mean anything. The first is the source: who measured this, across what population, in what year, and with what incentive. The second is the definition: attrition of whom, over what period, counted how; occupancy of what denominator; utilisation billed against which hours. The third, and the one almost never stated, is whether the number predicts anything at all, or whether it merely records something that has already happened. A benchmark stripped of those three things is not evidence. It is decoration with the authority of a decimal point.

A benchmark stripped of its source, its definition and its predictive power is not evidence. It is decoration with the authority of a decimal point.

We should be plain about what this document is, and what it is not, because that honesty is the whole point. ONX has no proprietary benchmark dataset. We do not run a panel of contact centres and publish their averages, and we are not going to pretend otherwise by quietly presenting a modelled figure as a measurement. What follows is therefore a benchmark of the published benchmarks: for each metric that people businesses are judged on, we go and find the best real, attributed, dated figure that exists, we say who produced it and how they defined it, and where only an industry convention exists we call it a convention rather than laundering it into a statistic. The reference value of this report is not the numbers it repeats; it is the discipline it models for reading them.

That discipline matters because delivery benchmarks are the vocabulary in which people businesses talk about their own performance. A business process outsourcer negotiating a renewal, a consultancy defending its rate card, a managed service provider explaining a missed service level: all of them reach for benchmarks to justify a position. When the benchmark is misread, the position is indefensible, and the reader on the other side of the table increasingly knows it. The metrics in the sections that follow are the ones that recur in those conversations, roughly in the order a delivery leader meets them: the workforce that leaves, the efficiency ratios that govern staffing, the utilisation that governs margin, the quality measures the customer feels, and the artificial intelligence that is quietly rewriting all of them.

The commercial stakes of getting this right are not abstract. A benchmark misread in a renewal negotiation costs real margin: a provider who concedes a service-level penalty because the buyer quoted an occupancy figure defined differently from its own has lost money to a definitional mismatch. A consultancy that benchmarks its utilisation against a published average without matching the definition can talk itself into cutting bench capacity it needed, or into complacency it could not afford. An outsourcer bidding an RFP against an attrition assumption drawn from the wrong survey misprices the work before it has won it. In each case the number was real; the reading was not; and the reading is what wrote the cheque. The whole industry runs on these comparisons, and the comparisons are only as sound as the definitions underneath them.

For each, we do the same three things. We give the real published range and its source. We say whether the metric leads outcomes (moves before the result, so it can be acted on) or lagsthem (records the result after it is fixed). And we say how it misleads when it is read on its own, because every one of these numbers does. If you take nothing else from this report, take the habit of asking those three questions of any benchmark before you let it decide anything. The rest is detail.

How to read a benchmark

Before the numbers, a short grammar for reading them, because the same figure can be a warning, a reassurance or a red herring depending on how it is held. There are three questions, and they are not optional.

Question one: what is the source, really?

A benchmark is a measurement, and every measurement has a measurer with a method and a motive. There is a hierarchy of trust here, and it is worth being explicit about it. A peer-reviewed study with a named sample and a stated method sits at the top. A reputable analyst firm publishing a dated survey with a described population (Metrigy, SPI Research, Deloitte, Gartner, ContactBabel) sits close behind. A vendor’s blog quoting “industry benchmarks” with no citation sits near the bottom, and a number that circulates with no traceable origin at all sits below the floor: it should not be used to decide anything. This is the same evidence hierarchythat governs any serious decision. A benchmark is a piece of external evidence, and its quality should be labelled before it is believed.

Question two: what exactly did it define?

Definitions are where benchmarks go to lie without meaning to. Attrition can mean annualised or monthly, voluntary or total, agents only or the whole site, first-year or tenured. Occupancy can be calculated against scheduled time or logged-in time. Utilisation can count billable hours against a 2,080-hour year or against available hours net of holiday. Two firms can report the “same” metric and be measuring genuinely different things, so a benchmark comparison is only valid when the definitions match. The single most common benchmark error in this industry is comparing a number to a published figure that was defined differently, and concluding you are ahead or behind when you are simply counting apples against oranges.

A worked example makes the trap concrete. Suppose your operation reports attrition of 26% and the slide says the benchmark is 30%, so you appear to be ahead. But your 26% counts voluntary resignations of agents who passed probation, annualised, while the published 30% counts all departures including probation failures across the whole site. You are not four points ahead; you are measuring a smaller thing and comparing it to a larger one, and your true, like-for-like number might be well over 40%. Nothing in the comparison flagged the mismatch, because both were labelled “attrition” and both looked like percentages. This is why the definition question is not pedantry: it is the difference between a benchmark that informs a decision and one that quietly corrupts it.

Two firms can report the same metric and measure genuinely different things. A benchmark comparison is only valid when the definitions match.

Question three: does it lead or lag?

The most useful distinction you can draw across a wall of metrics is between the ones that move before the outcome and the ones that move afterit. A leading indicator is a cause you can still act on: rising occupancy today predicts rising attrition next quarter, so it is a lever. A lagging indicator is a result already set: last quarter’s customer-satisfaction score records how delivery went, and no amount of staring at it changes what happened. Both matter, but they are used differently. You manage by leading indicators and you are judged on lagging ones, and a great deal of operational pain comes from treating a lagging number as if squeezing it directly would help. You cannot manage a business by pulling harder on the metrics that only tell you what already happened. The sections that follow label each metric as leading or lagging, and explain the mechanism, because the label is where the usefulness lives.

Hold those three questions in mind and most benchmark abuse becomes visible. The number without a source is dropped. The number with the wrong definition is set aside until it can be matched. And the lagging number is demoted from target to scoreboard, so that attention moves upstream to the leading metrics that can actually be changed. What remains is a much smaller set of numbers that genuinely deserve to influence a decision. Those are the ones the rest of this report is about.

Attrition: the headline that hides a definition

No metric is quoted more often, or more carelessly, than agent attrition. It is the number that appears in every article about the future of contact centres, usually with a single dramatic figure and no method attached. The reality is more instructive than any single figure, because the two most respected sources measure attrition at almost double each other, and both are right.

Metrigy, the analyst firm, tracks a clear and worsening trend. In its research base, contact-centre agent turnover rose from 21.8% in 2022 to 28.1% in 2023 to 31.2% in 2024, a steep climb over three years driven, in Metrigy’s reading, by the rising complexity of the interactions agents are asked to handle as the simple contacts are automated away.¹ Deloitte, in its Global Contact Center Survey, reports average annual agent attrition of roughly 52%.²Those two numbers, 31.2% and 52%, are both described as “contact-centre agent attrition,” and they differ by two-thirds.

There is a further layer the single headline hides, and it is the layer that actually matters for delivery: the distinction between early-tenure and tenured attrition. In many contact centres the great majority of leavers go in their first ninety days, before they have become productive, while the agents who pass their first few months stay for years. A blended annual attrition figure averages those two very different populations into one number that describes neither. An operation with 45% blended attrition that is almost entirely first-quarter churn has a hiring and onboarding problem; an operation with the same 45% spread evenly across tenures has a retention and management problem. The interventions are opposite, and the blended benchmark cannot tell you which you have. When you read an attrition benchmark, the useful question is not only how it compares but what tenure profile sits underneath it.

The temptation is to split the difference and quote something like “around 40%.” Resist it. Averaging two figures that measure different populations with different methods does not produce a truer number; it produces a fictional one with no source at all. The gap almost certainly reflects real definitional differences: the composition of each sample (Deloitte’s survey skews toward larger, often higher-attrition operations), whether the figure is voluntary or total departures, whether it counts annualised turnover or a headcount snapshot, and how each provider phrased the question to the people who answered it. Neither is wrong. They are answers to slightly different questions, and the honest move is to keep them apart and cite each for what it is.

Averaging two benchmarks that were defined differently does not produce a truer number. It produces a fictional one with no source at all.

What both agree on is direction, and direction is the part that leads. Attrition is one of the cleanest leading indicators a delivery operation has, which is why it belongs at the front of any benchmark discussion rather than buried in an appendix. A team losing people faster this quarter is a team that will, next quarter, carry more agents in their unproductive first weeks, resolve fewer contacts on the first attempt, score lower on quality, and cost more to run, because a new agent is an expensive agent for months. Attrition today is delivery quality tomorrow, and it is visible long before the delivery numbers move. That is what makes it worth watching as a cause rather than counting as a result. We develop the mechanism further in attrition as a leading indicator.

The mechanism by which attrition leads is worth spelling out, because it is what makes the metric a cause rather than a symptom. A departing agent takes with them not just a seat to be refilled but months of accumulated productivity, and the replacement arrives at zero. During the new agent’s ramp, they handle fewer contacts per hour, resolve fewer on the first attempt, need more supervisor support, and score lower on quality, all of which loads onto the colleagues around them. High attrition therefore does not cost a single replacement event; it keeps a rolling share of the workforce permanently in its least productive phase. An operation running 50% annual attrition may have a quarter of its agents in ramp at any moment. That is why the attrition line predicts the cost and quality lines: it is continuously feeding inexperience into the front of the operation, and the delivery numbers register it a quarter or two later.

There is one widely-circulated attrition-adjacent figure we deliberately do not use as a load-bearing claim: the frequently-quoted cost of “USD 10,000 to USD 20,000 to replace a single contact-centre agent.” It is repeated across the industry and attributed to various analysts, but it is genuinely hard to trace to a single primary publication with a stated method, and the true figure depends heavily on wage level, ramp time and geography. So we note it as a caution rather than a benchmark. The point of an attrition benchmark is not the exact percentage anyway; it is whether your own trend is bending in the wrong direction, measured consistently against itself, early enough to intervene.

Occupancy: the efficiency ratio people mistake for a target

If attrition is the metric that leads, occupancy is the metric that causes it, and the confusion between efficiency and health does more quiet damage in contact centres than almost anything else on the wallboard. Occupancy measures the proportion of an agent’s logged-in time that is spent actually handling contacts (talking, or in the after-call work a contact generates) rather than waiting for the next one. It is a measure of how tightly staffed the operation is against its demand.

The best public anchor for it comes from Call Centre Helper, whose analysis of more than 160,000 occupancy calculations put the average maximum occupancy target at 83.3%.³That figure is worth stating precisely because of how often the surrounding numbers are quoted loosely. The familiar rule that “75 to 85% occupancy is healthy” is an industry convention, a rule of thumb passed between workforce planners, not the output of a formal survey. It is a reasonable convention, and we repeat it, but we label it honestly: the band is received wisdom, and only the 83.3% average maximum is a measured figure.

The reason the ceiling exists is the reason occupancy is dangerous read alone. Occupancy and utilisation of people are not free to push upward, because the denominator is a human being. An agent held at 95% occupancy has almost no recovery time between contacts, takes the emotional load of one difficult interaction straight into the next, and burns out. High occupancy is, in effect, a way of borrowing capacity from the future by spending your people faster. It looks like efficiency this month and shows up as attrition, absence and quality erosion next quarter, at which point the operation is short-staffed and occupancy climbs again on the survivors. It is one of the most reliable doom loops in the industry.

There is a mathematical reason the ceiling is not arbitrary, and it is worth understanding because it explains why occupancy cannot simply be dialled up. The relationship between occupancy and customer waiting time is non-linear: as occupancy climbs toward 100%, waiting times do not rise gently, they rise steeply, because there is almost no idle agent available the moment a contact arrives. The queuing mathematics that workforce planners use (the Erlang models behind every staffing calculation) show that the last few points of occupancy are bought at a disproportionate cost in wait time and, therefore, in service level. This is why an operation cannot hold both very high occupancy and a demanding service level at once except by overstaffing against peaks, which lowers occupancy again. Occupancy, service level and staffing are three faces of one constraint, and the 83.3% average maximum is roughly where the trade-off stops being worth it.

High occupancy borrows capacity from the future. It reads as efficiency this month and arrives as attrition next quarter.

So how does occupancy lead or lag? It is best understood as a leading indicator of workforce risk that masquerades as a lagging measure of efficiency. Read as efficiency, it flatters the tightly-run site. Read as risk, sustained occupancy above the mid-80s is an early warning that the operation is over-extracting from its people and will pay for it. Occupancy is a guardrail, not a goal, and the operations that treat it as a target are usually the ones about to lose the staff who were hitting it. The related failure is chasing average handle time down in the same spirit, which we take apart in why average handle time is the wrong target: both are efficiency ratios that punish the operation when they are managed as objectives rather than watched as symptoms.

The practical reading is to hold occupancy next to attrition and to service level at the same time, never alone. An occupancy of 88% with flat attrition and comfortable service levels may be a genuinely efficient site. The same 88% alongside rising attrition and slipping service levels is a site consuming itself. The number is identical; the meaning is opposite; only the surrounding metrics tell you which you are looking at. That is the general lesson of this report in miniature, and occupancy is its clearest case.

Shrinkage: the planning number that is mostly convention

Shrinkage is the least glamorous number in the workforce planner’s toolkit and one of the most consequential, because it is the figure that turns a demand forecast into a staffing requirement. It captures all the paid time during which an agent is not available to handle contacts: holidays, sickness, training, coaching, breaks, meetings, system downtime and the rest. Get it wrong and every downstream number, the number of agents you hire, the shifts you build, the service level you promise, is wrong with it.

The benchmark almost everyone quotes for shrinkage is 30 to 35%. It is worth being scrupulous here about what that band is. It is an industry convention: the consensus range that experienced contact-centre practitioners report and that Call Centre Helper documents as the figure most centres come out at, alongside a reference to older benchmarking work that put an average near 35%.⁹It is not the output of a single current, universal survey with a stated method, and it should not be quoted as if it were. We repeat it because it is a reasonable planning starting point, and we label it because pretending it is a precise measurement would be exactly the error this report is against.

The reason the convention is dangerous as a default is that shrinkage is genuinely bimodal by function. A lean outbound sales operation may run shrinkage in the mid-20s. A complex technical-support or regulated-healthcare operation, with long training cycles and heavy coaching, can run in the high 30s or above. Applying a blanket 33% to both produces two wrong plans: one chronically overstaffed and burning margin, one chronically understaffed and burning people. The convention is a placeholder for a measured number, not a substitute for one.

The more useful cut through shrinkage than the headline total is the split between planned and unplanned. Planned shrinkage (holidays, scheduled training, coaching and meetings) is known in advance and can be scheduled around, so it costs capacity but not service level. Unplanned shrinkage (sickness, absence, unscheduled overrun) arrives without warning and hits the roster on the day, which is what actually breaks a service level. Two operations can both report 32% shrinkage and be in completely different health: one where most of it is planned and scheduled, and one where a third of it is unplanned absence that no forecast can absorb. A benchmark that reports only the total conceals the distinction that determines whether the number is manageable. The composition of shrinkage, in other words, predicts delivery risk far better than its level, and the composition is precisely what the 30 to 35% convention throws away.

A benchmark shrinkage figure is a placeholder for a measured one, useful only until you have your own.

Where does shrinkage sit on the leading-versus-lagging axis? It is neither, exactly, and that is worth naming rather than forcing. Shrinkage is a planning input: an assumption fed into the staffing model, whose accuracy is only revealed when the plan meets reality. Its benchmark value is not as a target to hit but as a sanity check on the assumption you are about to build a rota on. A planned shrinkage far below the conventional band should prompt the question of what has been left out; far above it, the question of what is being tolerated. Occupancy and shrinkage are usually discussed together because they are the two numbers that most quietly mislead, and we treat the pair in depth in occupancy and shrinkage: the numbers that mislead.

The connective point is that shrinkage is where a benchmark most obviously should be a temporary scaffold. You use the 30 to 35% convention on day one because you have nothing better; you replace it with your own measured, function-specific shrinkage the moment you can, because your operation’s real shrinkage is a fact and the convention is only a guess dressed up as one. The discipline of workforce planning, which we cover in what is workforce planning, is largely the discipline of retiring conventions in favour of measured facts as fast as the data allows.

Utilisation: the professional-services mirror

Step out of the contact centre and into the consultancy, the managed service provider or the professional-services arm of a BPO, and occupancy has a close cousin with a different name and the same failure modes: utilisation. Where occupancy asks what share of an agent’s logged-in time is spent on contacts, utilisation asks what share of a consultant’s available hours are billable. It is the metric on which people-based delivery margins are, in the end, won or lost, and it is under pressure.

The authoritative public source here is SPI Research’s annual Professional Services Maturity Benchmark. Its 2025 edition put average billable utilisation at 66.4%, down from 68.9% the year before, the lowest in the survey’s multi-decade history. High-performing firms, by contrast, reach about 81.2%.⁴The spread between the average and the leaders, nearly fifteen points, is the more instructive number than either alone, because it is the difference between a firm that keeps its people billable and one that does not.

Utilisation is where the “does it lead or lag?” question gets genuinely subtle, because it does both, in sequence. As a record of the quarter just gone, it lags: it tells you how much of the capacity you paid for actually earned revenue, after the fact. But it also leads the next outcome, because utilisation running below plan is an early signal of a pipeline that is not converting, a bench that is not being deployed, or scoping that is giving work away. Falling utilisation is margin erosion you can see a quarter before it lands in the accounts, if you are reading it as a cause rather than filing it as a result.

One reason a single utilisation benchmark misleads is that the right target genuinely differs by role, and a firm-wide average blends them into a figure that fits nobody. A junior consultant is expected to be almost entirely billable; a senior one carries business development, mentoring and oversight that are valuable and deliberately non-billable; a principal may be billable only part of the time by design. A professional-services firm is, in effect, a pyramid of different target utilisations, and its economics depend on the mix. Comparing your blended 66% to a published blended 66% tells you very little unless the two firms have the same shape of pyramid, which they rarely do. The benchmark that matters is utilisation by grade against your own model, not the single number that averages a deliberately uneven structure into a false uniformity.

Falling utilisation is margin erosion you can see a quarter before it reaches the accounts.

The way utilisation misleads read alone is the mirror image of occupancy. A very high utilisation figure, celebrated in isolation, can be the sign of an overworked delivery team with no slack for training, business development or recovery, which is to say the early stage of the same burnout loop that high occupancy produces in a contact centre. A very low figure can be genuine idleness, or it can be a firm deliberately holding capacity for a large mobilisation, or investing bench time in reskilling. The number does not tell you which. Utilisation is only interpretable next to the reason for it, and the firms that manage it well are managing the reasons, not the ratio.

What makes utilisation the clarifying case for this whole report is that it sits directly on the margin, the number that everything else in a people business eventually rolls up into. A benchmark like SPI’s 66.4% is useful not as a target to beat but as a question to ask: if your firm is materially below it, why, and is the reason a problem or a choice; if you are near the 81.2% of the leaders, are you there sustainably or by spending your people. The pressure on this number across the sector is real, and it is the clearest reason a delivery benchmark has to be read as a leading indicator of where the margin is heading, not a scoreline for where it has been.

Resolution, quality and the service level

The metrics so far have been about the operation looking at itself: how fast its people leave, how hard they are worked, how much of them is billable. This section is about the metrics the customer actually feels, which is where benchmarks are most likely to be quoted aspirationally and least likely to be defined carefully. Three matter most: first-contact resolution, customer satisfaction, and the service level.

First-contact resolution

First-contact resolution (FCR), the share of contacts resolved without the customer having to come back, is the single quality metric with the tightest link to both cost and loyalty, and it is measured more rigorously than most. SQM Group, which has benchmarked FCR for decades using a post-call survey method, puts the cross-industry average at around 70%, defines a good rate as 70 to 79%, and sets the world-class threshold at 80% or higher. Crucially, only about 5% of the 500-plus leading North American centres it benchmarks actually reach that world-class level.⁵

That 5% figure is the useful discipline. It means an unqualified claim of “80%-plus FCR” should be met with the question of how it was measured, because on a rigorous method it is rare. FCR is also the clearest leading indicator among the customer-facing metrics: a contact resolved the first time prevents the repeat contacts that inflate volume, and it lifts satisfaction and retention downstream. Raise FCR and cost, satisfaction and loyalty all move in your favour later. That is why it belongs at the centre of a delivery scorecard rather than the periphery, a case we make in contact-centre metrics that predict outcomes.

A word on the internal cousin of FCR, the quality or QA score, because it is where benchmarking breaks down entirely. Almost every operation grades a sample of contacts against a quality scorecard and reports an average, often a comfortable-looking figure in the high 80s or 90s. There is no external benchmark for this number worth the name, and there cannot be, because every operation writes its own scorecard: what counts as a quality failure in one centre is a pass in another, and the weighting is local. A published “industry average QA score” is therefore close to meaningless, an average of numbers that measure different things on different scales. QA is a genuinely useful internal instrument and an almost useless benchmark, and knowing the difference is the point. Compare your QA trend to itself and to your calibration, never to someone else’s number.

Customer satisfaction

Customer satisfaction is where benchmarking gets slippery, because “CSAT” means a dozen different things depending on the survey, the scale and the moment it is asked. The most rigorous public, cross-industry anchor is the American Customer Satisfaction Index (ACSI), a national measure built on around 200,000 customer responses. Its national score has sat at roughly 76.9 out of 100 and has barely moved since 2017.⁶ We cite it with an explicit caveat: the ACSI is a national cross-industry index of satisfaction with products and services, not a post-contact CSAT score for a contact centre, and the two should not be conflated. It is the best rigorous benchmark of the general satisfaction climate, not a target line for a support queue.

The honest position on operational CSAT is that there is no single authoritative public benchmark for it, because there is no single definition of it. A three-point post-call survey and a five-star app rating are both called CSAT and are not comparable. Customer satisfaction is a lagging indicator by nature: it records how delivery felt after it happened. It is essential to measure and dangerous to manage directly, because the lever is never the score itself but the resolution, effort and speed that produced it. Satisfaction is the scoreboard, not the game, and operations that try to raise the score by working on the score rather than on FCR and effort tend to move neither.

Service level

Finally, the service level, the metric most likely to be quoted as an unquestioned standard. The famous “80/20” target, answering 80% of contacts within 20 seconds, is genuinely the de facto industry standard, but it is worth knowing that it is a convention with a curious pedigree: it is traced to early telephony rather than derived from any study of what customers actually tolerate, and no research established that 20 seconds was optimal.¹⁰Many centres adopt it, in Call Centre Helper’s phrase, more or less arbitrarily. It is a reasonable default, but it is a default, and treating it as a law of nature is precisely the benchmark error this report warns against. Service level leads the customer’s experience of waiting, but chased in isolation it drives exactly the high-occupancy over-staffing-against-peaks behaviour that burns agents. Like every number here, it is only safe read in company.

Forecast accuracy: the benchmark that mostly is not one

There is a metric that sits underneath every staffing decision a contact centre makes and that has, remarkably, almost no rigorous public benchmark at all: forecast accuracy. It is worth a short section precisely because its absence from the trustworthy literature is itself instructive.

Forecast accuracy measures how closely the predicted contact volume matched what actually arrived, usually expressed as one hundred per cent minus the mean absolute percentage error. It governs everything downstream: forecast well and you staff close to demand; forecast badly and you are either paying for idle agents or missing your service level with queues out the door. Given how much rides on it, you would expect a well-established benchmark. There is not one.

What circulates instead is a set of conventions, and we are going to resist repeating them as though they were measurements. You will read that “95% forecast accuracy is the goal,” or that a good operation forecasts “within 5% at the interval level.” These are reasonable rules of thumb, and they appear widely, but they trace overwhelmingly to workforce-management vendors and consultancies rather than to a dated, methodical, public survey of a defined population. We could quote a crisp figure here and it would look authoritative on a slide. It would also be exactly the kind of unsourced, convenient number this report exists to refuse.

Where the only available benchmark is an unsourced convention, the honest figure is no figure, plus an explanation of why.

So the entry for forecast accuracy in this report is deliberately a gap with a label on it. The conventional targets exist; we name them as conventions; and we decline to present any of them as a benchmark statistic, because we cannot trace one to a primary source that states its method and population. An honest reference is defined as much by the numbers it refuses to invent as by the ones it cites.

The absence is itself a finding. A metric that everyone claims to manage but nobody can benchmark rigorously is a metric whose measurement is less standardised than its rhetoric. Forecast accuracy is defined a dozen ways in practice: at what granularity (interval, day, week), against which actuals (offered, answered, handled), and using which error measure. Those choices change the number so much that a single cross-industry figure would be meaningless even if someone gathered it. The honest conclusion is not that forecast accuracy does not matter, it plainly does, but that it is a metric you can only sensibly benchmark against yourself, under a definition you hold fixed. Where the industry has no comparable standard, the comparison you can trust is the internal one.

The practical guidance survives the missing benchmark intact, and is arguably better for it. Forecast accuracy is unambiguously a leading indicator: it is measured before the shift and it determines whether the shift will be over- or under-staffed. Because there is no trustworthy external line to hit, the only sound approach is to benchmark your forecast against your own history: track your accuracy consistently over time, at the interval level where it matters, and improve the trend. A forecast you measure against yourself, honestly, is worth more than a forecast you grade against a number nobody can source. This is the same logic that runs through workforce planning as a discipline: replace borrowed conventions with your own measured facts as fast as you can, and be candid about which is which in the meantime.

The AI overlay: when the benchmarks recalibrate

Every benchmark in this report was defined for a contact centre staffed and run by people. Artificial intelligence is now changing what those benchmarks measure, and the danger is not that the old numbers get better or worse but that they quietly stop meaning what they used to. This section is about reading benchmarks through that shift, because a figure compared across the boundary of automation is a figure compared against a moving definition.

Start with the best-evidenced effect. A peer-reviewed study of 5,172 customer-support agents, published in The Quarterly Journal of Economics, found that access to a generative-AI assistant raised agent productivity by 15% on average, with the gains concentrated among the least-experienced agents and the most experienced barely affected.⁷ That single finding should change how a leader reads a productivity or handle-time benchmark, because it means the average now hides a widening gap: AI compresses the distance between a novice and a veteran, so a site-level average that once implied a certain skill distribution no longer does.

The forward-looking claims are larger and should be read with the caution their horizon deserves. Gartner forecasts that autonomous AI systems will resolve 80% of common customer-service issues without human intervention by 2029, with a projected 30% reduction in operational costs.⁸ Deloitte reports that the most AI-centric contact centres in its survey are about 85% more profitable than their low-maturity peers.² Both are worth citing and both need their definitions held firmly. A forecast about 2029 is a projection, not a measurement, and the Deloitte figure is a correlation between AI maturity and profitability, not proof that the AI caused the profit rather than that already-capable operators adopted it first. The direction is credible; the causal certainty is not.

To see how sharply this bites, consider an illustrative case, offered to show the shape of the effect rather than to report a measured result. Imagine a support line where, before automation, half the contacts were simple password resets handled in two minutes and half were genuine problems taking ten. The average handle time is six minutes, and FCR is high because the resets always resolve first time. Now automation absorbs every reset. The human queue is entirely ten-minute problems: average handle time jumps to ten, and FCR drops because hard problems resolve first-time less often. Every headline metric has moved in the “wrong” direction, and the operation is doing exactly what it should. A leader reading those numbers against last year’s, or against a pre-automation industry benchmark, would conclude performance had collapsed. It had not. The work changed, and the benchmark did not keep up.

When AI removes the easy contacts, it does not just change the numbers on the wallboard. It changes what the numbers are counting.

Here is the mechanism that recalibrates every operational benchmark at once, and it is the reason this section exists. As AI resolves the simple, standardised, high-volume contacts, the contacts that remain for human agents are the hard ones: the escalations, the edge cases, the emotionally loaded conversations. That single shift moves every benchmark in the report. Average handle time rises, because the easy short calls are gone. First-contact resolution may fall, because what is left is genuinely harder to resolve in one touch. Occupancy pressure intensifies, because every remaining contact is heavier. A delivery leader who benchmarks this year’s human handle time against last year’s, or against an industry figure gathered before automation, is comparing two different jobs and calling it a trend.

This is also why AI raises the stakes on the evidence discipline rather than lowering them. The structural economics are shifting: as Harvard Business Review argues, generative AI automates precisely the standardised, monitorable tasks that made labour arbitrage rational in the first place, which reshapes what an outsourcing relationship is even for.¹¹ An operation deploying AI against benchmarks it has not re-based will make expensive mistakes, confidently, because the numbers will look familiar while measuring something new. We take up that failure mode directly in why AI without evidence creates expensive mistakes. The benchmark does not become useless under automation. It becomes something you have to re-earn.

The benchmark table

Gathered in one place, the metrics look less like a scorecard and more like a set of instruments, each calibrated differently and each needing to be read for what it is. The table below is the reference core of this report. It pairs each metric with its best real published range, the source that produced it, and, in the last column, the single most important property: whether the metric leads outcomes (and so can be acted on) or lags them (and so records what already happened). Where only a convention exists, the table says so rather than inventing a figure.

The way to use the table is not to read your own numbers across it and tick the boxes where you beat the range. It is to do three things in order. First, for each metric, confirm that your definition matches the source’s before you compare at all, because most of the value is destroyed at this step if you skip it. Second, note which of your metrics are worst relative to both their own recent trend and their matched benchmark, and treat those as candidates for the binding constraint rather than spreading attention evenly. Third, act only on the leading metrics in the final column, because those are the ones still upstream of the result. A metric marked “lags” belongs on the report you show the board, not on the list of things you change this week. The table is a filter, in other words, not a league table.

Read down the last column and the shape of the argument appears. The metrics that lead, attrition, occupancy-as-risk, FCR, forecast accuracy, are the ones a delivery leader can actually manage by, because they move before the outcome. The metrics that lag, customer satisfaction above all, are the ones the business is judged on, and they respond only when the leading metrics upstream of them are moved first. The reference value of this table is not the numbers in the middle columns; it is the discipline in the last one. A benchmark tells you where you stand. Only its position on the leading-versus-lagging axis tells you whether you can do anything about it.

Reading benchmarks as leading indicators of delivery risk

A table of benchmarks is a static thing, and delivery is not static. The final move this report wants to make is from the numbers to the decisions they should inform, because a benchmark that never changes a decision is just trivia with a citation. Three ideas turn a wall of benchmarks into an early-warning system for delivery risk, and each corresponds to a discipline ONX treats as foundational.

The first is the evidence hierarchy. Every benchmark in this report is external evidence, and external evidence sits below your own measured facts in the order of trust, because a published average was gathered somewhere else, on someone else’s population, with a definition that may not be yours. The correct use of a benchmark is therefore as a prior, not a verdict: a reasonable starting expectation to be replaced by your own measurement the moment you have it. An operation that runs on borrowed benchmarks it never re-bases is running on the weakest evidence available and calling it rigour.

The second is the principle that the worst constraint decides. A delivery operation does not fail at its average; it fails at its binding constraint. If forecast accuracy is excellent, service levels comfortable and quality high, but attrition is running away, then attrition is the constraint, and the healthy averages elsewhere are cold comfort. Reading benchmarks as a leading indicator of risk means finding the one metric that is worst relative to its own trend and its defined benchmark, and treating that as the thing that governs delivery, because it does. Delivery is set by the worst-behaving number, not the average of all of them, and a scorecard that reports the average hides the very thing that will break first.

A delivery operation does not fail at its average. It fails at its binding constraint, and the average is exactly where that constraint hides.

The third is the idea of a composite read, what ONX calls Delivery Confidence and describes more generally as business confidence. No single benchmark predicts whether a delivery commitment will hold. But read together, as leading indicators with their sources and definitions labelled and their worst constraint identified, they produce something a single number cannot: a defensible, evidence-weighted judgement about whether the work will be delivered at the quality and cost it was promised. That is the difference between a dashboard, which shows you numbers, and decision intelligence, which turns them into a governed judgement you can trace, challenge and learn from. We set out that discipline in full in what is decision intelligence.

Operationally, reading benchmarks as leading indicators is a cadence, not a one-off exercise. A useful rhythm is to hold each key metric next to three things at once: its own trend over the last several periods, the best-matched external benchmark with its definition confirmed, and the other metrics it moves with. Occupancy is read beside attrition and service level; utilisation beside pipeline and bench; FCR beside repeat-contact rate and satisfaction. The metric that is worst relative to its own trend and its matched benchmark is flagged as the binding constraint for the period, and it earns the attention, while the healthy metrics are noted and left alone. Done regularly, this turns a static scorecard into a moving early-warning system, one that surfaces the constraint before it becomes the crisis. The discipline is unglamorous and it is most of the value.

There is hard evidence that this way of working pays. McKinsey finds that only 20% of organisations say they excel at decision making, and 61% say most of the time they spend deciding is used ineffectively.¹² Bain reports a correlation above 95% between decision effectiveness and financial performance.¹³ For a people business, whose product is a chain of delivery judgements, that is not a soft finding: it says the quality of your decisions about attrition, staffing, utilisation and quality is, more or less, your margin. Benchmarks are the raw material of those decisions. Read as leading indicators of risk, with their provenance intact, they become an early warning system. Read as decorative averages on a slide, they become the reason the warning arrives too late.

This is the argument the sibling report, The State of Decision Intelligence for People Businesses 2026, makes at the level of the whole industry. This report makes it at the level of a single delivery scorecard: the benchmarks you are judged on are only worth anything if you know where each came from, what it defined, and whether it predicts the outcome you actually care about. Everything else is a number with the authority of a decimal point and none of the meaning.

Methodology & a note on honesty

A report about how to read benchmarks honestly has an unusually strict obligation to be honest about its own. So a plain account of how this one was assembled, and of the lines we would not cross.

The central constraint is stated at the top and repeated here because it is the whole premise: ONX holds no customer benchmark data. Nothing in this report is an ONX measurement. Every quantitative figure is a third-party published number drawn from a named, dated, public source, an analyst firm, a peer-reviewed journal or an established industry body, and each is listed in the References with a note describing what the figure is and its caveat. Where a metric is governed only by an industry convention rather than a survey, we say so explicitly and label the convention as a convention: this is true of the 75 to 85% occupancy band, the 30 to 35% shrinkage range, and the 80/20 service level, each of which is widely quoted as though it were a measurement and is not.

We were deliberate about disagreement. On agent attrition, Metrigy’s trend and Deloitte’s ~52% differ by roughly two-thirds, and we present both with their differing definitions rather than averaging them into a single false figure, because averaging measurements taken on different populations with different methods produces a number with no source. The same principle governs customer satisfaction: we anchor it to the ACSI national index while stating plainly that a national cross-industry index is not the same construct as a contact centre’s post-contact CSAT, and should not be read as one.

An honest reference is defined as much by the numbers it refuses to invent as by the ones it cites.

Two things we deliberately did not do. First, forecast accuracy: the conventional targets of “95%” or “within 5%” circulate widely, but we could not trace them to a dated, methodical, public survey of a defined population, so we name them as vendor conventions and decline to present any as a benchmark statistic. Second, the frequently-quoted cost of “USD 10,000 to 20,000 to replace a contact-centre agent”: it is repeated across the industry and attributed to various analysts, but it is very hard to pin to an original publication with a stated method, so we note it only as a caution and use it as a load-bearing claim nowhere. In both cases a gap with an explanation is more useful than a convenient number with no birth certificate.

A note on how sources were chosen, since the selection is itself a judgement. We preferred, in order: peer-reviewed research with a named sample and method; reputable analyst firms publishing dated surveys with a described population; and established industry bodies with a track record of measurement. Where a metric had several published figures, we chose the one whose method and population were most clearly stated, not the most flattering or the most dramatic, and we recorded the definition alongside the number in the References so the reader can judge the fit to their own operation. For contact-centre metrics specifically we leaned on sources that describe their measurement (SQM’s post-call survey method for FCR, Call Centre Helper’s large calculation base for occupancy), because a benchmark whose method is stated can be argued with, and one whose method is hidden can only be believed or ignored.

One element is our own framework rather than a measurement, and is labelled as such throughout: the reading of each metric as leading or lagging, and the three-part discipline (evidence hierarchy, worst constraint, composite confidence) in the closing sections, are a synthesis of how these numbers behave, not a benchmark in themselves. Any illustrative scenario in the text (for instance the occupancy doom loop) is offered to show the shape of a mechanism, not to report a specific measured outcome. We hold ourselves to the same evidence hierarchy we recommend: cited figures are cited, conventions are named as conventions, and our own framing is presented as framing.

This is the first edition. The AI figures in particular are moving quickly, and several are projections rather than measurements, so we expect to revise them. If you believe a figure here is wrong, or hold a better-sourced benchmark than the one we used, we want to know: a reference work earns its standing by being corrected, not by being defended.

References

Every figure in this report is sourced below. Third-party links open in a new tab. Where a figure is modelled rather than measured, or where a widely-quoted statistic could not be traced to its primary source, the report says so in the text.

1. What Metrigy’s Latest AI Data Reveals About Contact Center Staffing · Metrigy, 2024. Contact-centre agent turnover rising 21.8% (2022) to 28.1% (2023) to 31.2% (2024); Metrigy’s self-reported measure across its research base.
2. 2023 Global Contact Center Survey · Deloitte Digital, 2023. Average annual agent attrition around 52%; the most AI-centric contact centres reported being about 85% more profitable than low-maturity peers.
3. What Is the Ideal Occupancy Rate for a Contact Centre? · Call Centre Helper, 2023. An analysis of more than 160,000 calculations put the average maximum occupancy target at 83.3%; the 75 to 85% band is presented as convention.
4. Professional Services Maturity Benchmark 2025 · SPI Research, 2025. Average billable utilisation fell to 66.4% (2025) from 68.9% (2024), the lowest in the survey’s history; high-performing firms reach about 81.2%.
5. What is a Good First Call Resolution Rate? · SQM Group, 2025. Cross-industry FCR average around 70%; world-class threshold 80%+, reached by only about 5% of 500+ benchmarked North American centres; post-call survey method.
6. National ACSI (Q4 2025) · American Customer Satisfaction Index, 2026. National customer-satisfaction score 76.9 out of 100 (Q4 2025), based on about 200,000 customers; essentially flat since 2017. A national cross-industry index, not a contact-centre CSAT survey.
7. Generative AI at Work, The Quarterly Journal of Economics 140(2), 889-942 · Brynjolfsson, Li & Raymond, 2025. A study of 5,172 customer-support agents found a 15% average productivity gain, concentrated among the least-experienced agents; the most experienced gained little.
8. Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029 · Gartner, 2025. Forecast issued March 2025, with a projected 30% reduction in operational costs by 2029.
9. What is Call Centre Shrinkage and How to Calculate It? · Call Centre Helper, 2023. Reports the practitioner consensus that shrinkage typically comes out at 30 to 35%; references an older Dimension Data benchmark average near 35%. A convention, not a current universal measurement.
10. Is 80/20 Still a Reasonable Service Level? · Call Centre Helper, 2022. The 80/20 target (80% of calls answered in 20 seconds) is the de facto industry standard, a convention traced to early telephony rather than a derived optimum.
11. AI Is Rewriting the Economics of Outsourcing · Harvard Business Review, 2026. Generative AI automates precisely the standardised, monitorable tasks that made offshore labour arbitrage rational, changing what delivery benchmarks measure.
12. Decision making in the age of urgency · McKinsey & Company, 2019. Only 20% of organisations say they excel at decision making; 61% say most of the time they spend making decisions is used ineffectively.
13. Measuring decision effectiveness · Bain & Company, 2020. A 95%+ correlation between decision effectiveness and financial performance; top-quintile decision firms averaged materially higher total shareholder return.