Input and Output Metrics
Every team can hit its targets while the business misses its goal. When that happens, the failure may sit in the model connecting the targets to the outcome. The dashboard records whether the targets were met. It cannot, by itself, establish that the right things were asked of them.
This is the problem underneath the distinction between input metrics and output metrics.
An output metric measures a result: revenue, retention, signups. An input metric tracks an activity or condition expected to contribute to that result. The distinction is useful because a team needs something it can act on. “Grow revenue” names the desired outcome; it leaves the operating decisions largely unspecified.
But an input metric is a claim about causality. Choose one, assign a target, and you are making a prediction: moving this measure, under these conditions, should help move that outcome. The target makes the work explicit. The causal claim often remains implicit.
Control has a boundary
The input-metric discipline described by Working Backwards includes discovering useful inputs, examining their interactions, and reviewing them alongside outputs. It gives teams a way to connect daily work to a larger objective. The difficulty begins when those provisional relationships become fixed targets that no one is responsible for revisiting.
Consider a team responsible for renewals. It can change onboarding, respond to problems, and help customers use the product. It can influence renewal rates. It cannot decide whether a customer renews. Budgets, procurement, competing products, and switching costs sit outside its control.
Moving a metric upstream does not make it controllable. Adoption is upstream of renewal, but adoption still requires customer behavior. Even a measure that looks close to the work can contain decisions made by someone else.
There are three things worth keeping separate:
- The action: what the team can do, such as removing a specific obstacle in onboarding.
- The observed response: what happens next, such as more customers completing a useful workflow.
- The business outcome: what that response is expected to support, such as renewal or expansion.
The labels input and output depend on where you draw the boundary. Renewal rate is an output of several customer and product processes; it also contributes to recurring revenue. One team's output can be another team's input. A label does not remove the dependency.
This changes accountability. Teams need responsibility for their actions and for investigating the responses they observe. Leadership needs responsibility for the model that connects those responses to the business outcome. Otherwise, an uncertain relationship becomes a performance obligation nobody has agreed to examine.
The inputs interact
Take a hypothetical enterprise software company trying to grow annual recurring revenue. Its dashboard tracks qualified pipeline, proof-of-concept win rate, deal cycle length, adoption depth, and renewals.
These measures describe parts of a possible revenue model. Prospects enter the pipeline, some buy, some customers adopt the product deeply enough to get continuing value, and some renew or expand. Satisfied customers may also become references that help attract new prospects.
One possible feedback path looks like this:
This is a hypothesis about how the business works. Each arrow needs a reason to exist, and each transition takes time. Adoption today might affect a renewal months from now. A shorter sales cycle can bring revenue forward; if it depends on skipping implementation planning, it can also leave the customer less prepared to use what they bought.
The inputs do not necessarily improve together. Increasing one can weaken another. Nor does every improvement matter equally: if implementation capacity is limiting successful adoption, adding more pipeline may grow the queue without relieving the constraint.
Assigning a metric to each team does not assign ownership of those interactions. Someone has to notice when one team's improvement becomes another team's problem.
A green dashboard can conceal deterioration
Suppose sales is rewarded for closing quickly and customer success for keeping support escalations low. Sales reduces time spent preparing customers for implementation. Some customers struggle to get started, then stop trying. Escalations stay low because those customers are barely using the product.
Both measures can look healthy while the conditions for renewal deteriorate. This is one way a proxy fails: the same reading can be produced by two different underlying states. Low support demand might mean a reliable product with capable users. It might mean a product people have abandoned.
Incentives make the ambiguity harder to resolve. A sales team can shorten deal cycles by discounting. Marketing can increase reported pipeline by loosening qualification. Customer success can send more training emails without changing what customers can do. Each action moves a visible measure. Whether it helps the business depends on what was sacrificed and whether the expected response followed. When pressure to optimize a proxy pulls it away from the goal it represents, we have a Goodhart problem.
The people closest to the work often see those tradeoffs before leadership does. A dashboard can erase that context precisely when the target makes the context most important.
The question to ask of an input metric is: how could this improve while the outcome gets worse? The answer identifies what else needs to remain visible. Faster sales may need to be read alongside margin and implementation readiness. Fewer escalations may need to be read alongside actual product use. Adding measures is useful only if they distinguish states the original metric conflates.
Movement is not attribution
A customer renews after an adoption program. The team credits the program, and leadership gains confidence in the model. But the customer may have renewed because replacing the product was too expensive or procurement could not complete an alternative purchase in time.
The reverse is also possible. Adoption improves, but a budget cut ends the contract. A useful intervention can coexist with an unfavorable outcome.
Observing an input and an output move together cannot settle which explanation is right. The metric records a result without supplying the counterfactual: what would have happened without the intervention? Input metrics give you operational focus. They don't give you causal proof.
That uncertainty should affect how confidently the organization rewards, repeats, or abandons an action. Where feasible, a controlled experiment can help separate the intervention from other causes. Where it is not feasible, comparisons across similar customers, direct investigation, and an explicit account of competing explanations can still improve judgment. They do not become causal proof merely because an experiment is inconvenient.
Timing matters too. Judging adoption work before the affected contracts reach renewal can reject the model too early. Extending the deadline whenever renewal fails makes the model impossible to challenge. The expected delay belongs in the claim before the result arrives.
Keep the model open to revision
The original attraction of input metrics is operational focus. People can choose work, assign responsibility, and see whether something changed. That focus becomes brittle when the model is treated as settled.
In a new market, the connection between an action and an outcome may be exactly what the business is trying to discover. Fixing permanent input targets too early can reward consistent execution of a mistaken theory. In a mature business, a once-useful relationship can weaken as customers, competitors, or the product change.
For each important input, make the causal claim inspectable:
- What action can the owner take, and what response do we expect it to produce?
- Why should that response contribute to the outcome, and over what timescale?
- What could improve the measure without improving the underlying condition?
- What evidence would make us change the intervention, the metric, or the model?
Return to the adoption example. Sending onboarding emails is an action. Completing a workflow is a response. Renewing is a later outcome. If emails increase but useful workflow completion does not, the first link needs examination. If completion increases but renewals do not improve after the relevant renewal cycle, investigate the next link and the conditions around it. More activity at the beginning cannot resolve an unexplained break further along.
The review needs both the people who own the work and the people who own the outcome. The former can explain what the measure hides. The latter can decide whether the model still justifies the allocation of effort. Accountability includes surfacing evidence that the target has stopped being useful.
An input metric earns its place through an explicit, revisable explanation of how the work contributes to the result. Leadership owns that explanation, including the obligation to change it when the evidence no longer supports it.