When the metric overtakes the goal: what teams start doing to hit the number
A green indicator can hide a customer problem. Examine what the number measures, how targets shape behaviour and which evidence reveals what is missing.
Article language: English

In this article
Imagine a report showing that first response times have fallen. Customers are writing more often because the first response now says, "We have passed this to the appropriate department." The team has performed the measured task quickly and correctly. The customer's need is still making its way through the company at an unknown pace. A green indicator can tell the truth and still tell too little.
A measure shows one selected part of the work. When it also becomes the expected result, people may adjust the order and manner of their work to suit it. Check which outcome for the customer that adaptation supports and what remains outside the measurement. Exceeding a numerical target alone does not establish that the whole service is working better.
First name the need the number is meant to describe
Suppose a company wants customers to wait less for useful help. It measures first response time because the system makes it easy to read. That can be useful information about contact. It does not automatically show the time until the matter is clarified, the number of repeat requests, or whether the customer could take the recommended step.
In its guide to performance measurement by regulators, the National Audit Office warns that an indicator can influence behaviour in unintended ways when the same number is both a measure and a target. It calls for a wider evidence base and context. The guide is from 2016 and written for regulators; its observation helps frame a question about the scope of a company's measure.
In our example, the customer wants help and the first response is one stage. Write that relationship down. A fall in response time can then be considered alongside what happens next. If a faster contact is followed just as quickly by the appropriate action, the result may be useful. If it merely moves the customer to another queue, further analysis is needed.
You do not have to choose between one indicator and an enormous dashboard. Start with a question the current number cannot answer. For a few cases, review the history and ask the person doing the work. Sometimes the end point is poorly defined; sometimes there is no information about what happened to the request afterwards.
Record the definition used by the system. Does first response mean a message from a person, an automatic acknowledgement, or any message sent? Does the clock run continuously or only during service hours? No single definition fits every need. But the people comparing results and making decisions from them must know which one is in use.
If the company has changed its configuration, check whether the data from before and after still mean the same thing. An automatic acknowledgement may improve the reported time without making human help any faster. Show the change in definition and separately assess whether the new message is useful. Otherwise a technical change to measurement can be presented as an improvement in the whole service.
Closing tickets can change the picture of the work
In our first fictional example, the support team is expected to close a set number of cases a day. Closing a case after handing it to another department improves the local result. The customer may contact the company again because they are still waiting. If the report treats the new message as a completely separate need, the full history disappears.
Check what closing means and why it is used. The adviser may be following the system's configuration. The company may itself have decided that a handoff ends the adviser's stage. Before treating the behaviour as a way around the rules, reconstruct those rules. People may be correctly meeting a local expectation whose relationship to the overall result has never been examined.
A complementary measure should reveal the missing part. It might be the proportion of cases returning with the same need within a defined period, or a review of the time taken to perform a particular action. Explain what counts as a return and which cases are compared. A new purchase is not the same as another question about a task that has not been completed.
When contact volumes rise, managing a seasonal peak in enquiries also matters. A change in the result may reflect a different mix of cases or team availability. Comparing two periods without that context can attribute an unproven effect to the indicator or to employees. The guide to seasonal peaks in enquiries (in Polish) explores this context.
A quick response can transfer work to the customer
The second example concerns first contact time. A company introduces an automatic reply asking customers to submit their details again in a different form. The system shows a quick acknowledgement. The customer must do extra work even though they have already supplied some of the information. To assess the help, look at that journey too.
Review a few cases from the first request to a useful next step. Does the message add information? Are the requested details needed for action? Does the customer know what to expect? Not every automatic reply is a problem. It can confirm receipt and explain the sequence. Its value depends on the part it actually plays in the service.
In Goals Gone Wild, the authors analyse the risk that overly narrow goals focus attention on a selected result at the expense of other important parts of a task. The 2009 discussion paper draws on earlier research. It does not prove that every numerical target is harmful; it supports care in defining the scope of a goal.
Our own implication for support is to check whether speed of contact helps the next action. If the message arrives immediately but the customer still does not know when their case will be checked, speed alone tells us little. If it provides the right information and reduces uncertainty, it can serve a useful purpose. Assessment requires the content and the course of the case, not just the time sent.

Allow for observation time when measuring returns. A case closed today has not yet had the same opportunity to return as one closed a week ago. Without that condition, the latest results can look better simply because the customer has not had time to write again. Set an observation window suited to the process and retain it in the description.
Look at cases that remain open too. A team may close many simple matters while postponing harder ones that wait longer and longer. Do not automatically attribute that to a numerical target. It is worth checking the choice of order, the kind of help needed and the reason for the hold-up. Reading a few old cases can add to a picture that the average for closed cases will not show.
The number of publications does not describe how useful they are
In the third example, a team uses AI to prepare content. It reports the number of finished articles, but "finished" means a draft saved in a folder. Source checking and editing sit in a different queue. An increase in files may therefore mean a larger backlog that someone still has to review.
Define the result at the point the measure is intended to serve. A draft, a checked text and material available to readers are different states. Each can be measured, but they should not be used interchangeably. If a company assesses readiness to publish, it should also see waiting materials, revisions and the point at which a person confirms readiness.
The same boundary matters when measuring time spent working with AI. Producing a first version faster does not yet show how long it takes to achieve a useful result. When some work passes to another person, a local saving can coexist with a new burden outside the report's view. For the wider process, see the guide to measuring work with AI (in Polish).
Keep ordinary cases that do not look good in a demonstration, too. A service demonstration without staging should reveal conditions, not only the best selected sample. Likewise, the scope of an indicator's observation should be visible. Excluding difficult cases needs justification, so that the result does not come to describe only the convenient part of the work. For a related perspective, see the guide to service demonstrations (in Polish).
For content, separately examine what happened to selected drafts. Some were accepted after minor editing, some needed new research, and some were set aside. These proportions need not represent the whole team. Describing individual paths can still help reveal whether the shared label "finished text" conceals materially different outcomes.
Keep material withdrawn after checking too. If it disappears from the register, the report may show only what went forward, leaving out the work needed to reject the rest. An honest definition explains what enters the calculation and what happens to cases outside it. The point is not to punish someone for rejecting a weak draft. The check may exist precisely to stop such material.
Ask the team which action makes the number worse while helping the customer. It could be a longer consultation, a source correction, or explaining an unusual need before replying. Do not assume every long activity is valuable. But check whether the evaluation system leaves room for work needed to reach a useful result. Also name the contribution of checking, preparation and sustaining collaboration if that necessary work disappears from the report. The article on recognising overlooked work (in Polish) discusses contributions such as checking and coordination.
Consider what a reasonable person will do to reach the target
Before introducing a measure, try a short exercise. Ask: if we want to improve this number, which actions become the easiest route? Which help the customer and which only change the record? This is a useful part of a project premortem, especially when the indicator will guide the evaluation of a trial. The guide to project premortems (in Polish) explores how to consider possible causes of failure before a project.
Do not assume bad intentions. A measure may direct attention to simple cases because they allow more matters to be closed. It may also shorten explanations if the company communicates that every long conversation is a failure. The investigation should examine the relationship between the rule and the work. Assessing an individual employee requires its own appropriate context.
Add a small piece of information that reveals a side effect and check whether it helps. For a first response, it could be what happens next to selected contacts. For AI drafts, the time and scope of checking before use. For lost B2B proposals, the reason and stage at which the opportunity was lost, rather than only the number of documents sent. Choose according to the goal, not a wish to fill a dashboard. The article on lost B2B proposals (in Polish) discusses examining individual decisions without assuming the cause of every loss.
A complementary measure can change behaviour too. If the number of returns is added to response time, the company should still enable customers to make contact when needed. Hiding the button or moving the conversation outside the system would improve the report without meeting the need. Assess how information is collected and whether difficult cases remain visible.
An AI usage audit should read what the report actually means
Aeneia Audit describes support for understanding AI use, information, costs and controls. Bring a report and examples, and discuss whether examining the measures and transferred work fits the agreed scope.
The criteria for an AI pilot should allow a fair comparison. Changing a definition midway through a trial may be necessary when an error is found, but the change must be documented. Do not present two series collected differently as one continuous result without explaining the difference. Related questions are covered in the guide to setting AI pilot criteria (in Polish).
Before the next comparison, establish which decision the report should support. A manager may need to choose the scope of a trial, while the person doing the work needs to find a blocker. The same view need not answer both questions. Keep a shared definition of the data and choose the level of detail for the reader's action, rather than creating competing versions of the same result.
An AI use register can retain the goal, the measure and the limits of its use. This makes it easier to assess later whether an increase still describes the same task. A good indicator helps ask the next question. If its colour begins to close the conversation, look at the work left on the other side of the chart. See also the guide to maintaining an AI use register (in Polish).








