← All insights
AI

The Rumble Strip at Work

Feedback that arrives too late may still teach a lesson, but it cannot correct the decision that just produced the error... In operational systems, its greatest value often comes when [feedback] arrives early enough for the user to correct course.

How Real-Time Feedback Keeps Routine Work Moving, and Interrupts Errors Before They Spread

Series note: This is the fourth article in the Efficient by Design series on using behavioral science to redesign operational systems. The first article introduced the Detective Mindset, the second introduced the Decision Diagnostic, and the third applied Defaults. This article turns to the next Choice Architect tool: Feedback.

Feedback that arrives too late may still teach a lesson, but it cannot correct the decision that just produced the error. For operational feedback, timing matters: the signal has to arrive while the action is still reversible. Otherwise, it is closer to an autopsy than a warning.  Most organizations have no shortage of the latter usually called monthly reports, quarterly reviews, annual evaluations, and after-action reminders. Those documents are useful for cataloging errors. They are not useful for stopping them. By then the error has often already traveled downstream, and it may be too late to fix.

Operational feedback has a different job. It appears while the action can still be corrected. Thaler and Sunstein (2008) identify feedback as a core tool of choice architecture. In operational systems, its greatest value often comes when it arrives early enough for the user to correct course. The same logic shows up in a seat-belt chime, or in a rumble strip that vibrates when a tire drifts toward the shoulder. Neither waits for a monthly report. That is the kind of feedback choice architects need inside business systems.

Feedback Is as Much About Timing as It Is About Evaluation

In the language of the Decision Diagnostic, feedback can either support fast, automatic processing or interrupt it. Sometimes the signal is so intuitive that System 1 can correct without much conscious effort (Kahneman, 2011): a green row tells the analyst everything is on track, a flip to red makes an ageing invoice stand out.  These visual cues make the preferred action obvious. Other times the system needs a stronger interruption. A persistent warning, a confirmation, or a blocked Save button can create enough friction for deliberate review before the user continues.

The distinction matters because not every error needs more thinking. Routine work depends on automaticity. The goal is not to force System 2 into every click; that would make the process slower and more exhausting. The goal is to make the state of the system obvious when the routine path is safe, then become louder when something unusual or risky requires attention. Three design elements usually do this work: attention features such as color or contrast, persistent signals that remain until the deviation is corrected, and constraints that block or tax the error path. They often travel together, but they are not interchangeable.

Research on feedback cautions against treating more feedback as automatically better. Kluger and DeNisi's (1996) meta-analysis of 607 effect sizes found that feedback interventions improved performance on average (d = .41), but over 38 percent of the observed effects were negative. Their feedback intervention theory locates the problem in attention: feedback loses value as it pulls focus up a hierarchy, away from the task and toward the self. Hattie and Timperley's (2007) review of educational research makes a related distinction across four levels: task, process, self-regulation, and the self, with person-level feedback the least useful. That literature is about classrooms, not operations, but the implication is still practical: put the signal on the work, not on the worker.

What Good Operational Feedback Does

For operational choice architecture, I use four working tests for operational feedback. They are a design checklist, not a validated scale. The signal should arrive in real time, be tied to the exact action occurring now, reference a goal the user already understands, and make the deviation noticeable enough to prompt correction. A vague reminder sent tomorrow fails all four tests. A red field that appears while a questionable number is being entered does not.

That does not mean every signal should be red, loud, or annoying. If the desired correction can happen automatically, a subtle environmental cue may be enough to keep the user moving. If the consequences are larger, or the habitual response is likely to continue unchecked, the signal may need to persist or add friction. The Architect's task is to match the strength of the feedback to the risk of continuing on the wrong path.

Case Study: Thirty Hours on One IT Help Ticket

The following teaching case is adapted from an information-technology service environment. Identifying details and some figures have been altered or combined to protect confidentiality and illustrate the behavioral principle.

Engineers working after-hours incidents were expected to record billable time directly on service tickets. The form contained separate fields for days, hours, and minutes. The engineers knew how much time they had worked, and they understood why accurate billing mattered. Yet account teams began finding entries for 15, 20, even 30 hours on a single update. In many cases, the engineer had entered the right number in the wrong box: 30 minutes became 30 hours, or hours landed in the Days field.

The obvious response would have been retraining. But the pattern did not look like a knowledge problem. These were experienced engineers working quickly and touching tickets repeatedly during triage and resolution. The intention was correct; execution failed. Norman (2013) and Reason (1990) distinguish this kind of failure from a mistake in judgment: a slip occurs when a person intends the right action but performs it incorrectly.

The Decision Diagnostic therefore pointed toward Feedback. The desired action was simple: enter accurate billable time. The current error was also simple: correct digits were landing in the wrong field. The problem occurred during a fast, repetitive task with no immediate signal that anything unusual had happened. The system accepted an obviously implausible value with the same calm response it gave a normal one.

The Fix: Make the Error Feel Wrong While It Is Still Reversible

We redesigned the time-entry area around plausibility rather than memory. First, the Days field disappeared because it was not needed for ordinary ticket updates. Then the system began checking entries against the context it already knew. If a value exceeded a typical range, exceeded the time the ticket had been open, or exceeded the time since the previous update, the time-entry section changed visibly, and the system asked the engineer to verify the entry.

A questionable entry such as 30 hours triggered a bold message: “You reported 30 hours on this update. Is that right?” The engineer could choose Fix Entry and return immediately to the field. If the value really was correct, the engineer could confirm it, but doing so required a short justification. The point was not to shame the engineer. It was to make an implausible value harder to accept automatically than to correct.

This redesign used more than one tool. The visible warning was feedback. The grayed-out Save state was a constraint, not a nudge: the user could not simply continue. Norman (2013) calls this kind of block a forcing function. Feedback tells the user what is happening; a constraint changes what actions are available. In this case they worked together because a 30-hour billing slip was consequential enough to justify a brief interruption.

In the first months after the check went live, the internal error rate on billed hours fell from doule digits to under 1 percent, while billing cleanup dropped from roughly ten hours a week to about one hour of spot-checking. Those were internal operating measures, not the results of a controlled experiment, so other changes during the same period could have contributed. What the case demonstrates is the design logic: the engineers did not need to remember harder. They needed the system to tell them, at the right moment, that the value they had entered did not make sense.

Support System 1 Until the Exception Appears

The most useful part of this case is not the red warning. It is what happened before the warning appeared: nothing. Engineers could continue entering ordinary time without extra clicks, confirmations, or interruptions. The system supported the routine path until it detected an exception. Only then did it become intrusive.

That is the same support-or-interrupt principle from the Decision Diagnostic. Good feedback does not turn every transaction into a checkpoint. It lets System 1 handle familiar, low-risk work and reserves stronger signals for the moments when automatic processing is likely to miss something important. A dashboard that flashes constantly becomes background noise.  Human-factors research shows that frequent, unreliable, or low-value alarms can weaken attention and trust in the warning system (Wickens et al., 2004). A warning that appears only when the system detects a meaningful deviation has a better chance of earning attention.

A Practical Feedback Test

Before adding another alert, ask four questions. First, what exact action is drifting off course? Second, can the system detect the deviation while correction is still possible? Third, what does the user need to notice in that moment? Fourth, how strong must the signal be to produce a useful correction without creating alert fatigue?

Then measure the outcome. Did the warning reduce the error, rework, or downstream cleanup? Did users simply click through it? Did the alert create a new bottleneck? Feedback is a design hypothesis, not a decorative layer. If the signal does not change the behavior you intended, redesign the signal rather than blaming the user for ignoring it.

The Architect Moves Feedback Upstream

Managers often discover problems downstream: billing finds the bad time entry, finance finds the missing approval, or a client notices the missed response target. The Architect asks whether that information could have appeared earlier, at the decision point where someone still had the ability to change course.

Training builds knowledge. Reporting helps organizations learn. Operational feedback does something different: it makes the current state visible while the decision is still alive. Sometimes that means a quiet cue that lets automatic behavior stay on track. Sometimes it means a rumble strip strong enough to interrupt the routine. Either way, the principle is the same: do not wait until after the crash to tell people they were drifting.

In the next article, we will turn to Mapping: how to make future consequences easier to understand before a decision locks them in.