How IIoT deployments go wrong — and what prevents each one.
Industrial IoT projects rarely fail on the sensor or the dashboard. They fail in the middle: on radio coverage nobody surveyed, on integration through a path that was never agreed, and on data nobody owns. Each of these is cheap to prevent and expensive to retrofit.
| What goes wrong | How you notice | The fix |
|---|---|---|
| Connectivity chosen before the site survey | Coverage gaps appear after the hardware is installed | Survey first. Concrete, steel, distance and what the building is full of decide the radio — not the datasheet range |
| Everything streamed to the cloud | Data volume and cost climb while the useful signal gets harder to find | Decide at the edge what is worth sending. Raw telemetry is cheap to generate and expensive to keep forever |
| No plan for the outage | Gaps fall exactly across the events you needed to examine | Buffer locally and backfill on reconnect — on most South African sites this is a design requirement, not an optional feature |
| Integration through an undocumented path | Readings stop after somebody makes a control change | Read through a documented interface and agree in writing who may change it and who gets told |
| Tags named by whoever installed them | Nobody can find a point six months later | Agree the naming convention before commissioning. Renaming after integration is a migration, not an edit |
| Clocks not synchronised across sources | Correlating two systems produces conclusions that cannot be right | One clock source for everything. Without it, cause and effect cannot be demonstrated |
| Thresholds set on go-live day | Alarm fatigue within weeks; alerts acknowledged without being read | Set thresholds after a baseline period and review them once more a month later |
| Security treated as a network question only | An exposed controller turns up in an audit | Segment the network, expose nothing directly, and stay read-only wherever control is not actually required |
| No owner for the numbers | The dashboard ages and nobody notices | Name the person whose weekly meeting uses the figure. Reports that inform nobody get quietly switched off |
Two of these decide whether a deployment survives its second year. The integration path is the first: a reading taken through an undocumented route works perfectly until a controls change breaks it, and by then nobody remembers what it was reading from. Ownership is the second, and it is not a technical problem — a platform with no named consumer for its numbers degrades quietly, because nobody is inconvenienced when it stops being right.