How to Set KPI Targets and Thresholds Without Just Guessing

## The short answer

A KPI only becomes useful when it has a **target** (the level you are aiming for) and **thresholds** (the levels that trigger attention or action). Set them using a combination of historical baselines, capacity or capability limits, and a defensible improvement increment, never by reaching for a round number because it sounds ambitious. The most common failure is the arbitrary target: "let's aim for 95%" with no analysis behind it. Such targets are demoralising when impossible and meaningless when trivially easy.

## Start from a baseline, not an aspiration

Before you can sensibly say where you want to be, you must know where you are. Pull at least several periods of history for the metric and establish:

- the **typical level** (a median is usually more honest than a mean here),

- the **normal range of variation** from period to period, and

- any **seasonality or trend** that systematically moves the number.

This baseline does two things. It tells you what "normal" looks like, so you can distinguish a real signal from routine noise. And it gives you a realistic starting point from which to set an improvement, rather than plucking a target from the air.

## Choose a target you can justify

With a baseline in hand, a target should come from one of a few defensible sources.

1. **An external requirement.** A contractual SLA, a regulatory minimum, or a customer commitment sets a hard floor. These are non-negotiable and easy to defend.

2. **A capability ceiling.** Some metrics have a natural maximum, you cannot resolve tickets faster than your process physically allows without changing the process. Targets should respect these limits.

3. **A meaningful improvement increment.** Where you are simply trying to get better, set a target that is a stretch beyond the baseline but achievable within the period, informed by what past improvement efforts actually delivered.

4. **A benchmark, used carefully.** External benchmarks can inform a target, but only if the comparison is genuinely like-for-like. A benchmark from a different context can mislead more than it guides.

The test of a good target is whether you can explain *why* it is set where it is. If the only justification is "it felt right," reconsider.

## Define thresholds, not just a single line

A single target line answers "are we there?" but not "how worried should I be?" Thresholds add that nuance. A common and effective pattern is three zones.

- **Green:** at or beyond target, no action needed.

- **Amber:** below target but within an acceptable band, watch and investigate.

- **Red:** beyond an acceptable band, act now.

The placement of the amber and red boundaries should reflect your baseline variation. If a metric routinely wobbles within a certain range, the amber band should be wide enough not to fire on normal noise, otherwise you train people to ignore the colours. The red boundary should mark genuine abnormality, the level at which something has clearly gone wrong.

## Avoid the classic threshold mistakes

- **Round-number targets.** 90%, 95%, 100% feel natural but rarely correspond to anything real. Let the analysis, not the decimal system, choose the number.

- **Thresholds that fire constantly.** If a metric is red most weeks, the threshold is mis-set or the target is unrealistic. Alert fatigue is as damaging as no alerting at all.

- **Static targets in a changing world.** A target set two years ago may now be trivially easy or impossibly hard. Review targets on a fixed cadence.

- **One-size thresholds across segments.** A single threshold across very different cohorts hides problems in the weak ones and nags the strong ones. Segment where it matters.

- **Targets nobody agreed to.** A target imposed without the owning team's input is rarely owned. Set them collaboratively.

## Make targets dynamic where appropriate

For metrics with strong seasonality, a flat annual target is misleading, you will look heroic in peak season and disastrous in the trough for reasons that have nothing to do with performance. Where the pattern is predictable, set the target to follow the expected seasonal shape, or compare against the same period last year rather than a flat line. This keeps the signal honest.

## Close the loop

Targets are hypotheses about what is achievable. Treat them that way. At the end of each period, ask not only whether the target was hit but whether it was *the right target*. Consistently smashing a target suggests it was too soft; consistently missing it despite genuine effort suggests it was unrealistic or that something structural needs to change. Feed those lessons back into the next cycle.

## Where tooling helps

Calculating baselines, setting variation-aware thresholds, applying seasonal targets, and reviewing them on a cadence is laborious by hand and tends to decay. Analytics platforms can derive normal ranges automatically and flag thresholds that fire too often. At neart.ai we build enterprise-grade products in this area, and the consistent finding is that thresholds tuned to real variation are trusted, while arbitrary ones are ignored.

## Takeaway

Never set a target by reaching for a round number. Establish a historical baseline, choose a target you can justify from a requirement, a capability limit, or a defensible improvement, then set amber and red thresholds wide enough to ignore normal noise and tight enough to catch real problems. Review them every cycle and treat each target as a hypothesis to test.