Sixteen parameters
and forty versions
A service quality standard is easy to download and impossible to apply in somebody else's form. It starts working once it describes not only good answers but also the exceptions: when a parameter does not count, what to do with a block that has lost all its parameters, and how to score an agent on a day when the product is down. My standard has been through more than forty versions, and the effect of every amendment was measured.
Full sampling instead of spot checks
Scoring covers every conversation of the week, not three selected ones. That is only possible with automation: the volume cannot be reviewed by hand, which is exactly why most service teams sample. The difference is not about effort, it is about trust: with a sample of three, an agent fairly says "I was unlucky", and there is nothing to argue with.
- Every agent gets an individual score and a review of their own misses, not a departmental average
- The score affects pay, so the scale has to be protected against chance
- The departmental average and the spread go into the shared report; individual scores are delivered privately
The rules that usually are not there
- Excluding a parameter. If the conversation gave no occasion for an action, the parameter is not counted and does not enter the percentage. Otherwise an agent is penalised for something that was never required
- Dropping a block. A block left with a single counted parameter is excluded in full: one score cannot represent a quarter of the result
- Redistributing weight. When a block drops out, the weight of the rest grows on its own: three at 33 per cent, four at 25, five at 20
- The incident window. When the product is down and the fix is not in the agents' hands, the parameters outside their control leave the base. Speed, tone and closing remain
The standard also has a ceiling above one hundred per cent: an action taken where it was not required is rewarded, up to a maximum of 101%. That looks odd until you meet the opposite effect: without such a rule agents stop doing anything extra, even when it helps the customer.
Every amendment is measured
Forty versions is not a sign of indecision. The standard is amended on the back of each week and the amendment is checked the following one: did the average move, did the contested parameter stop causing arguments, did a miss start growing in a neighbouring block.
Some amendments were reversed. One rule that required two conditions at once was withdrawn: it systematically understated the score for one category of contacts. Such reversals are recorded together with the reason, otherwise the rule comes back in six months.
The standard does not live apart from the documentation
Scoring an agent against a standard that contradicts the policy and the training course is pointless: they will answer from whichever document they read last. So the standard is built into a single source of knowledge that produces twenty-three policies, ten courses, a bank of 284 questions and a public FAQ section.
- An amendment to a rule reaches the policy, the course and the test; the links are formalised
- An audit gate with eight blocking checks: no document is published without an audit stamp
- Full-set audits for contradictions between documents, rather than proofreading one at a time
- The same standard applies to the AI agent: 97.7% on the same parameters