Service Level Agreements: Committing to a Turnaround

A turnaround commitment promises that a stated share of work reaches an outcome inside a stated time. The promise is only as honest as what starts its clock and who it leaves out. A commitment measured from the moment work reaches the system, rather than from the moment the customer starts, can be met comfortably while one applicant in seven sits outside it entirely.

Every attainment figure is a fraction, and almost all of the argument in it lives in the bottom half. The numerator is usually measured carefully and reported honestly. The bottom half is chosen, usually early, usually by the people who will later be measured against it, and usually without anybody in the room noticing that a choice was being made at all. Learning to read a turnaround commitment is mostly learning to ask what sits underneath the percentage before reacting to the percentage.

What does a turnaround commitment actually promise?

A small picture carries the whole idea. A customer leaves a kurta with a tailor two weeks before a wedding and the tailor says it will be ready on Thursday. The tailor has made a promise about that one kurta. When the tailor has grown into a workshop with forty people, he no longer promises Thursday to anybody; he puts a board on the wall instead: ninety-five of every hundred garments leave this shop within two days of being brought in. The board on the wall is a different animal. The board is not about the one kurta at all. The board states something about a population, and any individual customer could be one of the five.

A turnaround commitmentA promise that a stated share of work reaches an outcome inside a stated time. in a bank is the board on the wall, not the promise about a single kurta. A turnaround commitment has three moving parts and one large silence. The three parts are a share, a time and a population: how much of the work, inside how long, out of what. The silence is everything the share leaves over. If ninety-five in a hundred are promised two days, the other five have been promised nothing whatsoever. No time at all is committed for them. The remaining five are not promised three days, or a week. Nothing in the sentence reaches them at all.

A share commitment is completely silent about its own tail, and that silence is a design decision rather than an oversight. Committing to every single item is possible, and some contracts do it. The last few per cent are the ones that are strange, so the cost of covering them is usually enormous. So the sentence stops at ninety-five, and whoever reads the board is expected to notice.

Sumeru Bank Limited, an invented mid-sized Indian bank, runs a retail personal loan intake chain from a customer starting an application on a handset through to a decision. The bank commits to deciding 95 per cent of applications within 2 working daysSeven hours at this bank, and the unit both the commitment and the exception desk count in.. A working day here is seven hours. Both of those numbers, the 95 per cent and the 2 working days, are the bank's own choices. Neither number is anybody's rule, and neither is a threshold set by a supervisor. Both are what a group of people at one bank decided they could hold, written into a document and reported on every month.

What are the six things a commitment has to state before anybody can check it?

Here is the practical test. Hand the commitment to somebody who does not work in the team that reports on it and ask them to check the figure. If they can, the commitment is written properly. If they have to come back and ask a question first, the commitment is missing a part. There are six parts, and every question a checker asks turns out to be one of them.

1. What is counted. One what? One application, one document, one call, one payment. The unit sounds obvious until two teams count differently and their figures stop agreeing.
2. What starts the clock. The event elapsed time is counted from. The clock start does the most work of the six and gets the least attention, and it is set out in full below.
3. What stops the clock. The event that ends the measurement, usually an outcome being recorded somewhere.
4. What is left out, and how much of it there is. Not just the fact of an exclusion but its size. An exclusion without a count is unfalsifiable.
5. The share, and over what period. 95 per cent, read over a calendar month. A share read over a quarter and a share read over a day are different promises.
6. What happens when it is missed. Who is told, and whether anything is owed to anybody.

Five of those six are almost always written down. The fourth is almost never written down, so the fourth is the one worth asking for first. Nobody hides it. The people writing the commitment know perfectly well what they are counting, so stating what they are not counting feels like stating the obvious. Two years later, everyone who knew has moved on, and the figure is being read by people who were never told.

What starts the clock, and what does that one choice decide?

A queue at a government counter behaves the same way. There is a moment a person joins the queue, a moment a token is handed over, a moment the token is called, and a moment the work is done. If the counter measures itself from the token being called, it can report a superb average while people stand for two hours before getting a token. Nobody has lied. The measure is exactly what it says it is. The measure simply begins somewhere that leaves out the part of the experience that hurt.

At Sumeru Bank the clock startThe event from which elapsed time is counted, which decides who can appear in the figure at all. is the file reaching the decision engine. The file reaching the engine is the third of four points on the applicant's timeline. The applicant taps to start, completes digital onboarding by supplying identity, documents and consent, the completed file arrives at the decision engine, and a decision comes back. The clock starts at the third point and stops at the fourth.

Now, and this matters more than anything else in the case, that choice had a reason and the reason was a good one. The chain was designed so that the commitment covered what the bank controls. Everything before the third point depends on the applicant finding their documents, taking a photograph in decent light and getting through a consent screen. Committing to a turnaround over steps that wait on a customer is how teams end up promising things they cannot hold. Nobody at the bank was being clever, and the clock start was a defensible engineering decision made years before anybody read a monthly attainment figure.

The defensible decision quietly decided the population. In the month measured, 10,000 applications were started and 8,600 completed onboarding and reached the decision engine, being 86.0 per cent. The other 1,400 stopped somewhere in between. Because the clock cannot start for a file that never arrives, those 1,400 are not slow, not late, not breaching anything. The 1,400 sit outside the sentence. The clock start did not just decide when to begin counting. The same choice decided who was allowed into the count.

What stops the clock, and what arguments live there?

The clock stopThe event that ends the measurement, usually the outcome being recorded somewhere. looks like the easy half and it is where the most tedious arguments actually happen. At Sumeru the clock stops when the decision is recorded. Fine. Then the questions arrive, and each of them is genuine.

Is the decision recorded when the engine writes it, or when the applicant is told? The two events can be minutes apart or, on a bad evening, hours. If a person on the exception desk asks the applicant for one more document, does the clock keep running while they wait? Is a file returned to the applicant for correction a decision, a pause, or a fresh file with a fresh clock? Every one of those has a defensible answer, and every one of them moves attainment.

The trouble is never which answer a commitment picks, it is a commitment that has not picked one. The bank genuinely cannot decide a file whose payslip has not arrived, so stopping the clock while waiting on a customer is entirely reasonable. The pause stays reasonable exactly as long as it is written down and the paused time is counted somewhere a reader can see it. An unwritten pause rule is the one place where attainment can quietly become whatever the reporting team needs it to be, without a single person deciding to make it so.

Who is outside the commitment by construction?

Of the month's 10,000 applications, 1,400 never completed onboarding. The shortfall splits three ways: 620 were rejected by the liveness check on the selfie image, 480 abandoned at document upload, and 300 could not complete the consent step. 620 plus 480 plus 300 is 1,400. Why each group stopped is covered separately; what the attainment argument needs is the counts.

The 1,400 are excluded by constructionWork that cannot appear in a measure because of the way the measure is defined, rather than because of how it performed.. Exclusion by construction is not the same as exclusion by a rule somebody could relax, and the difference is worth slowing down on. There is no configuration to change. A clock that starts when a file reaches the engine simply has nothing to start on for a file that never reaches it. The exclusion is a consequence of the definition, and it would survive any amount of good intent.

The 620 rejected by the liveness check are the sharpest case in the whole month, and they deserve to be looked at squarely rather than waved past. Each of those 620 is a person who opened an application, was told no by a component, and then appeared in no turnaround figure the bank produces. Not as a miss, not as a slow case, not as a footnote. 620 people is twice the 310 that the same month's report calls a miss, and the report has a line for the 310 and no line at all for the 620. Cathy O'Neil's argument in Weapons of Math Destruction, 2016, is exactly this shape: the people a measure cannot see are usually the people it is failing, and the measure keeps improving as it loses them.

What happens when the same month is measured two ways?

Here is the month, and the two lines are worth reading slowly because they are the whole argument. Of the 8,600 files that reached the decision engine, 8,290 were decided inside 2 working days. The 8,290 are 96.4 per cent of the 8,600, against a commitment of 95 per cent, so the commitment was met with 1.4 points to spare and 310 files missed. Of the 10,000 applications started that month, the same 8,290 were decided inside 2 working days. The same 8,290 are 82.9 per cent of the 10,000, against the same 95 per cent, missing it by 12.1 points.

The numerator did not move. Nobody recounted anything. The same 8,290 decisions appear in both sentences, measured the same way, by the same system, on the same days. All that changed is the denominatorThe population the share is taken over, which is chosen rather than observed., and the 13.5 points between 96.4 and 82.9 are entirely a decision about who counts. Neither figure is wrong, and that is precisely what makes the pair difficult: a reader shown either one alone has been told something true and has not been told the denominator underneath it.

The natural next thought is that a wider window would fix it. If 2 working days produces 82.9 per cent from the customer's side, surely 3 or 4 would climb towards 95. The answer is worth predicting before it is checked.

The shape is the lesson. The curve measured from the engine crosses the promised line between the first and second working day and then flattens against the ceiling of 100. The curve measured from the first tap climbs the same way and flattens against a ceiling of 86.0 per cent. There it stops forever. Widening a window can only ever help the work that is inside the measure, and no window in the world reaches work that never entered it.

Which attainment report flatters, and which one can be checked?

Every month the figure went up as one line: attainment 96.4 per cent against a commitment of 95 per cent, status met. Everything in that line is true. The line is also unreadable, in the strict sense that a reader cannot get from it to anything they could verify or argue with. There is no way to tell what it is 96.4 per cent of.

The honest version is barely longer. An honest report states the count decided inside the window, both denominators with their attainmentThe share actually achieved against the commitment in the period measured. figures, the number that never entered the measure, and the clock start. Five lines. The fix for a flattering report is almost never a harder commitment. Publish the denominator beside the attainment every single time, and 96.4 per cent and 82.9 per cent arrive together, neither able to travel alone.

What does the commitment cost in capacity, and who pays for it?

A commitment is not held by intention. A commitment is held by having enough people to do the work inside the window, and a turnaround promise stops being a document and becomes a roster right there.

Trace the 310 misses. The chain sends 5,590 of the month's 8,600 files straight through with no person touching them, and those return a decision in about 4 minutes, median. Nothing in that path can take 2 working days. The other 3,010 go to a person on the exception desk. So every single one of the 310 misses came from the 3,010, or 10.3 per cent of the exception work, and the straight-through path contributed nothing to the miss count at all. A turnaround commitment on a chain like this is not a promise about the components, it is a promise about the desk.

The locked distribution makes a further point, and it is one worth noticing because it is easy to read past. The count decided inside one working day is 5,590, exactly the number that went straight through. Not one exception file closed inside a single working day. The whole of the desk's output lands on day two or later, and the commitment and the desk's natural rhythm were designed to sit right on top of each other.

Now put a roster against it. The desk handles 3,010 exceptions a month over 20 working days, so 150.5 arrive each working day. Seven people at an assumed 420 minutes each is 2,940 minutes a day, and at 19 minutes a case that closes 154.7 a day. The margin is 4.2 cases a day, being 2.7 per cent of capacity. The desk carries about 285 open cases at any moment, and 285 cases against 150.5 arriving a day is about 1.9 working days of work sitting in the queue. Against a commitment of 2 working days, that is a margin of about a tenth of a working day, or roughly 45 minutes. The queue depth and the margin are arithmetic on the bank's own locked counts rather than a measured waiting time, and they are worth stating as arithmetic.

Read that margin next to the desk's capacity margin of 2.7 per cent and the picture assembles itself. Seven people at an assumed fully loaded Rs 9,00,000/- a year each is about Rs 63,00,000/- a year, and that spend is what the 96.4 per cent is actually made of. A week carrying 3 per cent more volume turns a stable queue into a growing one, and a growing queue eats the 45 minutes long before anybody sees a red figure in a monthly pack. A turnaround commitment is a capacity promise wearing the clothes of a service promise, and it fails through the roster rather than through the technology.

What happens when it is missed, and who is told?

Two things get confused here and they are worth separating. A breachA single item that missed the committed time, as distinct from the commitment as a whole being missed. is one item that took longer than the committed time. The commitment being missed is the month's share falling below the promise. 310 out of 8,600 still leaves 96.4 per cent inside, so Sumeru had 310 breaches and no missed commitment. A breach and a missed commitment are different events with different audiences, and a report that only names the second has 310 stories with nowhere to go.

So ask the four questions in order. Is the breach recorded against the item, so somebody could count them? Is anybody inside the bank told, and how often? Is the applicant told that their decision is running late? Is anything owed when it is late, whether that is a fee waiver, a call, or simply an apology with a revised date? At Sumeru the first two have answers and the last two do not. When nothing at all follows from missing it, a commitment whose last two questions are blank is a target rather than a commitment.

Why does a comfortably met commitment reveal so little?

Think about ten shops in one mall that all share one delivery driver. On average, deliveries are fine. In practice, the shop at the end of the round waits three hours every single day and the shop by the entrance never waits at all. The average is a real number describing nobody's Tuesday.

The intake chain is exactly that. 5,590 files, being 65.0 per cent, come back in about 4 minutes. 3,010 files, being 35.0 per cent, take up to 2 working days. Taking those two path times at face value, the weighted average elapsed time is about 0.7 of a working day. No applicant has ever experienced 0.7 of a working day. Some got an answer while the tea was still hot and some waited two days. An attainment figure over a chain with two paths is an average over two populations that share nothing except a report line.

Which is why a green attainment figure is the start of a conversation rather than the end of one. A green figure shows that the desk kept up last month. The figure does not show what it was a share of, how the two paths fared separately, how close the queue ran to its margin, or how many people were never in the count. Every one of those is a question the figure cannot answer and none of them is hostile.

How do three different people actually use this figure?

Ismail Sheikh, who heads the exception desk, reads the commitment as a roster. For him the chain of arithmetic runs one way: exceptions a month, over working days, into arrivals a day, against people times minutes over handling time, into a margin. When somebody proposes tightening the window from 2 working days to 1, he does not argue about the customer promise. He points out that no exception file has ever closed inside one working day, so the proposal is a request to rebuild the desk rather than a request to try harder. The roster arithmetic is the number anybody is actually asking for when they propose a change to a committed window, so they should arrive with it already done.

Neelima Rao, reviewing the chain from the risk function, reads it as a single question: what is this a share of, and who cannot be in it. She does not need to doubt the numerator to find the problem, and she does not need to accuse anybody of anything. Asking for the denominator and the excluded count in writing takes one line in a report request and it converts a green figure into a checkable one. If the answer comes back as an exclusion with no count attached, that is her finding.

An applicant, who is the only person here without a report, experiences none of this. She experiences either four minutes or two days, and if she is one of the 620 she experiences a screen telling her no. The gap between what she experienced and what the pack said is the whole problem, and closing it costs one extra line of reporting rather than a new system.