Every service level in an outsourcing contract is a claim about who did what. The measured number sits in a row, the target beside it, and the colour of the cell reads as a verdict on the supplier being scored. Usually that reading holds. Occasionally the cell is a verdict on somebody else entirely, and the contract has nowhere to record it.
A current-state review of an outsourced IT service desk at a large European public-sector organisation carried one of those cells. Call abandonment rate, three-month average, 2.06%. The contract allowed anything under 8%. The industry average the review benchmarked against was 3.5%. On the face of the table the desk cleared both comfortably. The review then wrote down why the rate sat where it sat: it had steadied from the month before the measurement window opened, owing to lower call volumes and to support from the new extended user support team, a second supplier putting people in front of staff before those staff reached for the phone.
Neither of those causes is the desk under contract.
A fraction with somebody else's hand on the denominator
Abandonment rate is calls abandoned over calls offered. Every operations manager knows this and almost nobody says it in a governance meeting, because the number gets presented as a quality of the desk rather than an arithmetic result of two quantities the desk only half controls.
The desk controls the numerator, roughly. Staff the lines properly, hold the queue short, and fewer callers hang up. It does not control the denominator at all. Calls offered is a fact about demand, and demand here was being reduced deliberately by a different supplier whose people walked the floors and headed problems off before they became contacts.
The effect runs both ways at once, which is what makes it hard to argue about afterwards. Removing calls shrinks the denominator, which on its own pushes a rate up. It also shortens the queue for the calls that remain, which cuts the numerator. On this desk the second effect won, because the review recorded the rate steadying and attributed that to those two causes. The desk was answering into a quieter room, and a quieter room produces a better abandonment rate from the same people working at the same pace.
That is my reading of the mechanism. What is on the slide is the attribution itself, written plainly, in a document that scored the supplier as compliant on that service level.
The number that failed describes the same queue
On the same table sat abandonment time: 1 minute 32 seconds against a contract limit of under one minute and an industry average of 25 seconds. That one failed, at roughly 3.7 times the market figure by my arithmetic.
Read together they are not independent findings but two readings of one queue. The rate counts how many callers gave up. The time measures how long the ones who gave up were prepared to hang on first. A desk with a genuinely healthy queue produces a low rate and a short abandonment time, because nobody waits long enough to reach the point of giving up. This desk produced a low rate and a long abandonment time, which is the signature of a queue where the people who wait, wait badly, and where the count stayed low for a reason outside the queue's behaviour.
The scorecard treated them as two rows with two independent verdicts. One compliant, one breached. Nobody reconciled them, because the reporting format did not ask anyone to.
An abandonment rate is a fraction, and a second supplier had its hand on the denominator. The contract had a field for the fraction and no field for the hand.
The same cause showed up elsewhere with the opposite sign
The extended user support arrangement moved more than one number in that review. Its effect on cost per incident has already been written up: prevention took tickets out of the denominator and made the unit-cost metric look worse, which is set out in The Desk Answered in 17 Seconds and Solved Fewer Than Half the Calls.
The part worth adding is the asymmetry between the two. One intervention, by one party, made the client's own cost metric look worse and the incumbent supplier's quality metric look better, in the same months, from the same mechanism. Both facts were recorded in the same document, and nothing in the governance model was built to hold them as one. The responsibility matrix in that review ran across four roles, two on the client side and two on the incumbent's, covering service level monitoring, achievement and change. The party whose people were absorbing the demand appears in none of them.
Attribution failure rarely looks like a dispute. It is two true statements filed in separate places, each scoring a party that did not cause the movement being scored.
Deflection makes this the standing condition rather than the anomaly
I raise a decade-old measurement problem because the arrangement that produced it is now the default design for a service desk. An automated tier in front of a human tier is a demand suppressor, and far more powerful than a handful of floor-walkers. It answers contacts, resolves some, and passes the rest through. Every ratio computed on the human tier changes the day it is switched on, without a single person on that tier doing anything differently.
If the human tier is a supplier under a service level agreement, and the automated tier is built by the client or a platform vendor, then whoever owns the deflection owns the movement in the supplier's numbers. Contracts almost never say who that is, because they were drafted for a world with one queue and one party in front of it.
Both signs are available. Deflection that clears the simple contacts leaves the human tier a harder mix, so handle time rises, resolution rates fall, and a supplier gets penalised for a change in composition it did not choose. Deflection that shortens the queue improves answer speed and abandonment, and a supplier books credit it did not earn. In neither case does the scorecard know which happened.
What to settle before the automated tier goes in front
Four things, and all of them are contract drafting rather than engineering.
Report the denominator beside every ratio. A rate published without its volume cannot be audited. On this desk the rate and its benchmarks sat in one table, the volumes that explained the rate on another page, and the link between them survived as a line of commentary. Two columns instead of one closes it.
Name the party that owns deflection, and say whose scorecard moves with it. If nobody owns it, the answer defaults to the supplier whose numbers happen to move, which is the arrangement this review found.
Write the re-baselining trigger in advance. A step change in volume should reopen the targets automatically, on a rule agreed before anyone knows which direction it will cut. Renegotiating after the change is when both parties already know who benefits, and that conversation does not go well.
Keep at least one measure a third party cannot move. Absolute counts, duration per contact handled, resolution of the contacts the tier actually took. Ratios over demand are the ones that drift for reasons outside the contract.
A fifth item comes out of the review itself, and it shows how easily a definition slips. The review proposed a call-back option as the lever for bringing abandonment time down. Sensible, and it changes what abandonment means: a caller who accepts a call-back and puts the phone down has not been served yet and is no longer waiting. Whether that call lands in the numerator, leaves the denominator, or disappears from both is a definitional choice the contract had not made, so introducing the lever would have moved the metric again, in a direction chosen by whoever wrote the reporting query.
The average and the weeks it hides
One further point sits underneath all of this. The 2.06% was a three-month average taken across a window in which demand was being actively suppressed, and the review's own judgement was that steady state was not the risk. The risk was peak load during technology roll-outs, and the recommendation was workforce assessment and scheduling against the roll-out calendar.
An average computed over a quiet period says little about the weeks that matter, and least of all when the quiet was manufactured by someone outside the agreement. That is the version of this I expect to meet most often over the next two years: a desk reporting comfortably against inherited ratios, an automated tier holding the denominator down, and nobody able to say what the human tier will do the first week a large deployment lands on it.
Figures and findings are as recorded in a current-state review of an outsourced IT service desk at a large European public-sector organisation: its contract scoreboard benchmarked three ways against industry, contract target and measured performance, plus the review's own commentary on causes and levers. That review produced findings and recommendations ahead of a re-sourcing decision, not delivered results, and no corrected measurement regime was built. Attributing the abandonment rate to lower call volumes and the second supplier's extended user support is the review's own. Reading the rate and the abandonment time as two views of one queue, the arithmetic on the industry multiple, and everything said about deflection ownership and contract drafting are ours.
“An abandonment rate is a fraction, and a second supplier had its hand on the denominator. The contract had a field for the fraction and no field for the hand.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.