Every business generates far more data than anyone reads. Order records, timestamps, page views, call logs, delivery times, stock movements. Almost all of it is accurate and almost none of it is looked at.
A signal is different. A signal is a specific condition that, when it occurs, means someone should do something. Data is continuous and passive; a signal is discrete and demands a response.
The gap between the two is where most reporting effort is wasted — producing more data in the hope that signals will emerge from it. They generally do not, because a signal has to be defined in advance.
Anatomy of a signal
A properly constructed signal has four parts. Missing any one of them turns it back into data.
A condition. A precise, testable statement. Not "sales are down" but "weekly enquiries from the commercial segment fell below 20 for two consecutive weeks."
A threshold decided in advance. Set before looking. Thresholds chosen after the fact are always chosen to justify a conclusion someone already reached.
A recipient. One named person, not a distribution list. Signals sent to groups are received by nobody.
A defined response. What that person does. Even "review the last ten and report back by Friday" is enough. Without this, the signal produces awareness, which is not action.
Types of signal worth having
Threshold signals. A number crosses a line. Cash below a floor, response time above a ceiling, stock under a reorder point. Simple, reliable, and the right choice where the acceptable range is genuinely known.
Trend signals. A direction sustained over a defined period. More robust than single-point thresholds, and better suited to noisy figures. The design requirement is deciding how many periods count as a trend — before you start.
Absence signals. Something that should have happened did not. No enquiries from a source that normally produces them. A recurring order that did not arrive. A report that did not run. These are the most commonly missing category and often the most valuable, because absence is invisible by nature — nothing appears on a chart to prompt the question.
Comparison signals. One unit behaving unlike its peers. One branch, one technician, one product line diverging from the others. Useful because they are self-normalising: a general market shift moves everyone and triggers nothing.
Composite signals. Two conditions together. Enquiry volume steady but conversion falling. Revenue up and margin down. These frequently carry more meaning than either component alone, and they are the ones a person watching a dashboard will not spot.
Why more data does not produce more signals
Attention does not scale. Adding a chart to a report does not add a reader. Beyond a certain density, additional information reduces the chance that any specific item is noticed.
Humans are poor at noticing absence. We reliably see what is present and unreliably see what is missing. A machine checking "did this happen?" is far better at this than a person reviewing a report.
Variation looks like meaning. Given enough numbers, some will move interestingly by chance. Reviewing many metrics without pre-set thresholds generates false alarms, and false alarms train people to ignore the whole report.
Nobody reviews everything, every period. Real reviews are partial and driven by what was interesting last time. Signals do not depend on someone being diligent on a Tuesday.
Building a small signal set
Start from decisions, not from data. Ask what the business would want to be told about promptly if it happened. The list is usually short — five to ten items — and most of it is already known intuitively by whoever runs the operation.
Typical entries in a small business:
- A quote over a certain value has had no contact for a week
- Enquiries from the main channel dropped below the usual range for two weeks
- Any customer account is more than 60 days overdue
- Median first-response time exceeded the published promise for a week
- A recurring customer has not ordered within their normal interval
- Any job has been in the same status for more than X days
- Margin on any service line fell outside its expected band for a month
Then, for each: threshold, recipient, response. Write them down together in one document. That document is more valuable than most dashboards, and it can usually be implemented with the alerting features already present in the systems you own.
Maintaining them
Signals decay. Review them at a set interval and ask two questions.
Did each one fire? A signal that has never fired is either well-calibrated for a stable condition or badly calibrated and useless. Check which.
Was each firing acted upon? A signal repeatedly ignored is worse than no signal, because it teaches people that alerts are noise. Either fix the threshold or remove it.
Aim for a low but non-zero rate. Signals that fire constantly get filtered out; signals that never fire are not being tested.
Signal design
- The condition is precise enough to be evaluated automatically
- The threshold was chosen before examining the data
- One named person receives it
- The expected response is written down
- Absence conditions are included, not just threshold breaches
- The total number of active signals is small enough to take seriously
- Firing frequency is reviewed periodically
- Signals that are consistently ignored are recalibrated or removed
Most businesses could delete half their reporting and add six well-designed signals, and would make better and faster decisions as a result. The bottleneck was never the availability of data.