Labour is the entire cost base
In manned guarding, staff cost dominates. Margins are set by how efficiently hours are allocated, how much is spent on overtime and agency cover, and whether contracts are priced for what they actually consume.
That makes the most valuable applications unglamorous: rostering, absence cover and contract-level cost analysis. All rely on data the business already generates through its scheduling and time systems.
Contract profitability at shift level
Many guarding firms know profitability at contract level annually and nothing in between. Analysing at shift level frequently shows that a contract judged acceptable overall is loss-making on particular shifts.
- Actual hours worked against contracted hours
- Overtime and agency premiums by site and shift
- Travel time and cost where staff cover multiple sites
- Cover rate - how often the rostered person is not the one who attends
- Training and vetting cost amortised per site
The cover rate is often the revealing figure. A site where a third of shifts are covered by someone other than the regular guard costs more in overtime and delivers a worse service, and it usually indicates something specific about that site.
Rostering as an optimisation problem
Building a roster is optimisation rather than prediction: assign people to shifts subject to hours regulations, licence requirements, site-specific vetting, travel and preferences.
Machine learning contributes the inputs - predicted absence, predicted cover requirements, predicted travel times - and an optimiser builds the schedule. Attempting the roster itself with a model usually produces something that violates a constraint, which makes it useless.
Alarm triage
Where a firm operates monitoring, false alarms consume response capacity and erode client relationships. Predicting which activations are likely genuine, from sensor patterns, site history, time and weather, allows better prioritisation.
| Use | Assessment |
|---|---|
| Prioritise response order | Reasonable and useful |
| Flag sites with recurring false alarms for engineering | Often the biggest win |
| Adjust patrol frequency by predicted risk | Reasonable with human oversight |
| Decide not to respond at all | Not advisable - liability and duty of care |
The second row frequently matters most. A site generating repeated false activations usually has a fixable cause - a sensor positioned badly, a door not closing properly - and identifying it removes the problem rather than managing it.
Keep analysis away from individual profiling
It is technically possible to analyse patrol patterns, response times and incident handling at individual level. Doing so as a performance surveillance tool damages trust and tends to degrade the data as people adapt to what is measured.
Aggregate and site-level analysis delivers the operational benefit without that. Where individual data is used, it should be for scheduling fit and development rather than monitoring.
In guarding, the margin is in the roster. Everything else is a rounding error.