Your SLA Dashboard Says 100%. Your Clients Are Angry.

Compliance and service are different measurements. Most processors only have the first one.

Share
Your SLA Dashboard Says 100%. Your Clients Are Angry.

Compliance and service are different measurements. Most processors only have the first one.

I have never seen an SLA report that was wrong about arithmetic. I have seen plenty that were wrong about service.

The distance between a dashboard reporting 98% and a client relationship that feels strained is rarely dishonesty. It is the accumulated effect of a dozen small definitional choices, each defensible on its own, that together move the number ten or twenty points away from anything a human experienced. Nobody decides to report a misleading figure. Everyone decides, separately and reasonably, where to put a boundary — and the boundaries compound.

This is a piece about where that distance comes from, and what it takes to close it.

Why the dashboard is probably wrong

Six definitional choices decide your compliance number before a single ticket is worked. Most organizations have never written them down.

What the clock runs on. Business hours or calendar hours. A ticket opened Friday at 5:10 p.m. is either fifteen hours old on Monday morning or sixty-four. Both are defensible. Nobody can tell which one your report used by looking at it. And if your service desk measures in business hours while your client's operations team lives in calendar hours, you are not disagreeing about performance — you are reporting in different units and calling it the same number.

When the clock starts. Creation, triage, or assignment. If it starts at assignment, a ticket that sits unassigned in a queue for six hours is free. That is the single most common quiet leak I have seen: a response-time metric that is excellent precisely because the measurement begins after the response.

When it stops. Resolution, client confirmation, or closure. "Resolved" declared by the provider is a different event from "working" confirmed by the client. Systems that auto-close after five days of client silence produce beautiful numbers and furious account managers.

The pause. Time waiting on the client, excluded from the clock. This one is legitimate in principle and the most abused in practice. Every hour of your own slowness can be laundered through a well-timed question. Set the ticket to pending-client at 4:45 p.m. on a Friday with a question your own team could have answered, and the weekend costs you nothing. I am not describing malice; I am describing an incentive. If pause time is not itself measured and reviewed, it will grow, and nobody will decide to grow it.

Reopens. A ticket resolved, reopened twice, and finally fixed nine days later. Is that one breach, or three resolutions of which the last two complied? Most tools count the final resolution against the original clock or, worse, start a fresh clock on each reopen. The client experienced nine days and three disappointments. The report can show 100%.

The denominator. Which tickets count at all. Enhancement requests, "not a defect," third-party dependency, duplicates, user error. Every exclusion is arguable and several are correct. But exclusions are applied after the fact, by the party being measured, to the tickets that went worst. That is not a measurement process; that is a filter with a plausible name.

None of these is fraud. Each is a reasonable answer to a real question. The problem is that six reasonable answers, chosen separately and never published, can move a compliance figure from the seventies into the high nineties without anyone doing anything differently.

The quietest lever: who sets severity

Everything above is visible once someone looks. This one usually isn't.

Severity determines which clock applies. A P1 with a four-hour target and a P2 with a twenty-four-hour target are the same incident with different consequences, and the difference between them is a judgment call made by a person under pressure.

When the provider classifies, severity drifts down. Not through dishonesty — through the ordinary gravity of a hundred small decisions made by people who know what the number costs them. A P1 reclassified to P2 is not a breach; it is a reclassification. It has its own justification, usually a sound one about scope or workaround availability. Do it often enough and you have a compliant service and an unhappy client, which is precisely the condition this essay is about.

When the client classifies, severity drifts up, and for the mirror reason: the only lever a client has to make something go faster is the severity field. A support organization where 60% of tickets arrive as P1 is not facing a crisis. It is facing a client who learned how the queue works.

The useful question is not who should own it. It is whether severity is reviewed by anyone who is not scored on it. Payments people understand this instinctively in other contexts — nobody lets the same team authorize and reconcile. The same separation applies to the field that decides the clock.

The minimum defensible arrangement: severity is set at intake against written criteria tied to business impact, not to how loud the reporter is; any change in severity after intake is logged with a reason and visible to the client; and the rate of downward reclassification is itself a reported metric. If that last number is rising while compliance is rising, you have found your gap.

The average is the enemy

Suppose every definition above is clean. The number can still mislead, because of how it is aggregated.

A monthly compliance percentage is a count of tickets that met target divided by tickets measured. It treats a breach by ten minutes and a breach by six days as the same event. It also treats a breach on a client's reconciliation window the same as one on a cosmetic defect in a back-office screen.

Nobody experiences an average. Clients experience specific incidents, and they remember the worst one. A month at 97% compliance with three catastrophic misses on the same client is, from inside that client's building, a bad month — and your report says it was better than last month.

Three changes fix most of this.

Report the distribution, not the mean. The 50th percentile tells you how the machine normally runs. The 95th tells you what your worst week looks like, and that is the number clients argue about. Average resolution time is the least informative statistic in the whole report and it is almost always the headline.

Report breaches by magnitude. Ten tickets that missed by under an hour is an operational wobble. Two that missed by a week is a structural failure, and it is invisible in a compliance percentage.

Report per client, not only in aggregate. Portfolio compliance is a number for your board. A client reads only their own row, and if the aggregate is carried by your five easy clients, the sixth one knows something your report does not say.

The test is simple: if the number went up and a specific client's experience got worse, the number is measuring the wrong thing at the wrong grain.

Your clients are not one client

A processor signs broadly similar SLAs with every institution it serves. Those identical terms produce genuinely different services, because the measurement rewards behaviors that only some clients have.

The mature client has its own operations function. Tickets arrive with logs, transaction references and a reproducible case. They respond to questions in minutes, which means the pause clock barely runs and their tickets look slow on your report even when you handled them well. They also know exactly how severity works and classify accordingly. These clients are expensive to serve and easy to measure — and because they read the report closely, they are the ones who catch the definitional games.

The small client files a ticket that says the system is not working. Clarifying it takes two exchanges over a day and a half, all of which is pause time, none of which counts against you. Their tickets show excellent compliance. Their actual experience is the worst in your portfolio, and your dashboard will never tell you. This is the most dangerous case precisely because the number looks good.

The client in a project phase sends change requests through the incident channel, because the incident channel is the one with a clock on it. If you classify them out, compliance is clean and the relationship corrodes. If you work them, your incident queue is full of work that was never incident work and your capacity model is wrong.

The client with a regulatory exposure has incidents that carry external reporting obligations with their own deadlines. Those deadlines do not pause for client response and do not care about your business-hours calendar. An incident that is comfortably inside your SLA can be comfortably outside their regulatory window, and they will escalate on that basis while your report shows compliance.

The conclusion is not that the SLA is useless. It is that the SLA is a contract, not a service model. The contract sets a floor that applies to everyone. The service model is how you allocate attention above that floor, and it has to differ by client because your clients differ. A processor that runs only on the contract is managing six relationships with one instrument and will be surprised by at least two of them.

Service versus efficiency

Here is the part that budget conversations never reach.

Queues do not behave linearly. A support team running at 95% utilization is not marginally slower than one at 70%. It is dramatically slower, because waiting time rises steeply as utilization approaches full and the queue stops draining between arrivals. This is arithmetic, not attitude. Which means a processor can breach its SLA with excellent engineers working hard, and the honest diagnosis is that the staffing model has no slack in it. Every efficiency program that raises utilization buys cost savings with response time, and the trade is usually made without anyone naming it.

Measurement distorts triage. Once compliance is the number the organization is judged on, the queue gets worked in order of time-to-breach rather than business impact. A cosmetic ticket forty minutes from its deadline jumps ahead of a serious one that breached yesterday and cannot be saved. That is rational behavior under the metric and indefensible as a service decision. If your team has ever said "that one's already breached, leave it," the metric is running the operation.

The cheapest fix is never the real one. There are three ways to improve compliance. Renegotiate the targets. Reclassify the tickets. Or fix the defects generating the volume. The first two are quick, cost almost nothing, and work. The third takes engineering budget, competes with roadmap commitments, and shows results in two quarters. Under cost pressure, organizations do the first two and report success. The volume does not move, and three years later the support cost is structural.

This is the tension at the center of the whole subject. Compliance can be bought cheaply, and service cannot. A processor serious about the difference has to be willing to hold deliberate slack in its support capacity, prioritize by impact even when it costs the number, and fund defect reduction out of the same budget line that pays for support. Each of those decisions looks like inefficiency on a quarterly view and like professionalism on a three-year one.

The useful frame for the executive conversation: an SLA is a risk instrument, not a cost line. The question is not "what does this level of service cost" but "what does a breach on our largest client actually cost us, and are we spending less than that to prevent it." Most organizations have never calculated the second number, which is why the first one always wins.

The gap you cannot close alone

A regional processor sits in the middle of a chain. Above it, bank clients with commercial leverage and tight expectations. Below it, a core platform vendor, usually in another country and another timezone, with its own support agreement.

The structural problem is that the SLA you sign upward is almost always tighter than the one you hold downward. A four-hour commitment to a bank on an incident whose fix lives inside vendor code, against a vendor agreement measured in business days on their calendar, is an open position. You have sold a level of service you cannot manufacture, and the difference is absorbed by your own people working around the problem.

This is survivable and extremely common. It stops being survivable when nobody has written down where the exposure is. The questions worth answering before a client asks them: which incident categories depend on vendor response, what the vendor's real response has been as opposed to its contracted one, and what the processor does in the window between its own deadline and the vendor's.

In a small market the problem has an additional edge. The pool of institutions is small, relationships are long, and your client may also be your shareholder's subsidiary, your competitor's partner, or the employer of someone you will work with in five years. An SLA dispute in that setting is never only contractual. Which argues for conservatism in what you commit to and generosity in what you disclose: the reputational cost of a missed target you flagged early is a fraction of the cost of one the client discovered themselves.

And there is a timezone fact that no contract fixes. If your vendor's engineering day starts when yours is half over, a certain class of incident has a built-in delay that no amount of local effort removes. That should be reflected in what you commit to, or financed by staffing that covers the gap. Pretending it is not there is the most expensive option, because the cost shows up as breaches you cannot explain.

What professional measurement looks like

Organizations move through roughly four stages, and most of the ones I have seen are at stage two while believing they are at stage four.

One: no measurement. Service quality is a matter of opinion and whoever complains loudest gets attention. Rarer than it used to be, but it still exists behind a dashboard nobody consults.

Two: a number nobody trusts. The dashboard exists, the percentage is high, and neither the client nor the support manager believes it reflects anything. Everyone has a story about a ticket the report counted as compliant. The number is produced for a meeting and forgotten the same afternoon. This is the common case.

Three: a number that is accurate and useless. The definitions have been cleaned up and the figure is honest. It still drives nothing, because it reports what happened without saying why, and nobody can act on a percentage.

Four: measurement that drives investment. Compliance is reported alongside volume by cause, repeat incidents, and the engineering work commissioned to remove them. The conversation in the governance meeting is about what is generating tickets, not about whether last month cleared the bar.

Two tests tell you where you actually are.

The reproducibility test. Can your client compute your number themselves from the ticket data you already give them, and land within a point of your figure? If yes, the number is a measurement. If no — because exclusions, pauses or reclassifications are applied in a spreadsheet only you hold — it is marketing, however accurate it happens to be. Publishing the method is what separates the two, and it costs nothing except the ability to adjust the method later.

The volume test. Is anyone accountable for ticket volume going down? Compliance measures how fast you clean up. It says nothing about how much mess arrives. An operation where compliance improves every quarter while volume grows every quarter is getting better at absorbing a problem it is not fixing. The mature version pairs every SLA report with a problem-management view: top recurring causes, what was remediated, what is funded next quarter.

Professionalization, in one sentence: the point is not to report a better number, it is to make the number boring, and then argue about the causes instead.

My judgment

An SLA is a floor, not a promise of quality.

One hundred percent compliance means you never fell below a level you negotiated, under definitions you chose, on a population you filtered. That is worth something — it is a real discipline and the absence of it is worse. But it answers a narrow question, and the organization usually hears it answering a broad one.

The broad question is whether the floor is in the right place, and whether you are above it for the right reasons. A processor can be fully compliant and structurally fragile: compliant because the clock is generous, the exclusions are wide, the severity drifts down and the worst client's experience is averaged into a portfolio figure. Every one of those is reversible. None of them reverses by itself, because the number looks fine.

The move that changes an operation is not a better dashboard. It is publishing the method to your clients and then living with it. The moment a client can reproduce your figure, every definitional convenience you were carrying becomes visible, and you have to fix the service instead of the measurement. That is uncomfortable for about two quarters and then it is the thing that distinguishes you from every competitor still reporting 99.4% to a room that does not believe it.

And if you only take one operational change from this: start reporting pause time and downward reclassification rate next to your compliance figure. Neither requires new tooling. Both are already in your ticket data. If they are flat, your number is probably real. If they are climbing while compliance climbs, you have your answer, and you found it before your client did.

Three questions

  1. Could your largest client reproduce your compliance figure from the ticket data you already send them — and if not, which step of the calculation would they be unable to see?
  2. Who in your organization can change a ticket's severity after intake, and is that rate reported anywhere?
  3. When your support team is at capacity, what gets worked first: the ticket closest to breaching, or the one with the most business impact? Answer with what actually happens, not with the policy.

Reply with even one line. The sharpest answers shape the next issue.