Assessing IoT readiness in logistics: can your sensing estate carry an AI decision?

Stage 1 is not an absence of sensors — it is an absence of continuous observation. Almost every operator here already owns telemetry hardware: tractors have telematics because insurers and, in the United States, the ELD rule require it; handhelds scan at pick and at delivery; the gatehouse writes trailer numbers on a sheet. None of it produces a stream. It produces a handful of discrete confirmations per shipment, separated by hours or days in which the operation genuinely does not know where its freight is or what condition it is in.

The characteristic experience of stage 1 is the reconstruction. When something goes wrong — a late delivery, a rejected pallet, a temperature claim — the operation assembles a timeline afterwards from scans, driver recollection, gate sheets and emails. It is usually good enough to settle the dispute and never good enough to have prevented it, because every element was recorded after intervention was still possible. Operators call this a visibility problem; it is more precisely a sampling problem, in which the world is sampled a few times per journey at moments chosen for administrative convenience rather than for decision value.

The stage is cheap to leave and expensive to occupy, and the expense is invisible because it is distributed: the dispatcher phoning for positions, the detention charge nobody can contest without arrival evidence, the reefer claim settled because the box's temperature cannot be proven, and the fact that no AI proposal can be evaluated honestly because there is no history to backtest against. The move out is not a fleet-wide hardware programme. It is one decision, instrumented properly, on a scope small enough to finish.

Stage 2 is where most logistics operators sit, and it is genuine progress: the estate now observes continuously. Reefer controllers log supply and return air temperature, tractor telematics reports position and engine state, some trailers carry solar trackers, and a retail customer's mandate has put RFID tags on a slice of the freight. Each of those populations arrived as its own project with its own supplier, and each terminates in a portal. The telemetry exists; it does not exist anywhere a model, a rule or a TMS field can reach without a human in between.

Three defects make portal-bound telemetry unusable for decisions, and operators tend to fix only the first. Access: the data sits behind a login, exportable as CSV on the vendor's schedule. Identity: the reading belongs to a device serial, and mapping that serial to a trailer, a shipment and an order is a hand-maintained relationship that decays every time an asset is repaired, swapped or re-tagged. Time: each population stamps events with its own clock and convention — device, gateway or ingest time, some local and some UTC — so events from two populations cannot be reliably ordered. Fixing the API without fixing identity and time buys a fast feed of unjoinable readings.

The stage has a political failure mode too. Because each portal shows its own population looking healthy, nobody owns the question of whether the estate as a whole observes what the operation needs. The fleet team reports 98% device uptime, the cold-chain team reports zero missed alarms, and the transport desk still cannot say where a specific customer's pallet is, because the answer requires joining three portals that share no key. Every number is true and the operation is still blind.

Stage 3 is the first stage at which the estate produces observations rather than readings. The technical difference is small and the operational difference is total: an observation states that a specific physical thing was in a specific state at a specific known time, on a clock the rest of the operation shares. That is the unit a model can consume, a rule can fire on and an auditor can accept. Getting there takes three pieces of work that have nothing to do with buying sensors — buffered ingest, identity resolution and time normalisation — plus one conceptual step: agreeing a canonical event.

The canonical event is where GS1's vocabulary earns its keep. An operator that adopts an EPCIS-shaped model — what object, at what location, at what time, in what business step — stops arguing about whose gate-in is the real one: there is one event type with defined precedence and deduplication, and the tracker read, the gate camera and the driver's scan are three observations of it. The identification keys matter as much: an SSCC for the logistic unit and an asset identifier for the trailer or container let a reading survive a cross-dock and a trailer swap, which a device serial does not.

What stage 3 has not done is characterise error. The stream is complete, joined and timely, and nobody can say how wrong any individual number might be. Probe placement inside a reefer is unrecorded; the last calibration date sits in a maintenance system nobody joins; the position feed's accuracy comes from a datasheet rather than from measurement; firmware has been updated over the air twice this year and nobody checked whether the sampling interval changed. Models trained on that stream inherit its unmeasured error, present outputs with unearned confidence, and degrade in ways monitoring cannot see, because the data keeps arriving on time.

Stage 4 imports asset-management discipline into what most operators still treat as a consumables budget. Devices become a managed population with a register, a lifecycle and an owner: where each unit is fitted and to what, which firmware it runs, when it was last verified, how much battery and data budget remains, and which decisions depend on it. That last field is what changes behaviour — once a device is known to feed a decision, its failure stops being a maintenance ticket and becomes an incident with a named consequence.

The quality gate is the stage's defining artefact. It sits between the resolved stream and the decision and enforces the sensing specification: are the required inputs present, in tolerance and fresh enough? A gate has three outcomes, not two — pass, degrade (widened bounds or a simpler schedule-based rule), or suppress and escalate. Operators who build only pass and fail switch the capability off during exactly the disruptions when it would have been most valuable, because a binary gate cannot express 'still useful, less certain'.

Trust also has a physical dimension that software teams underestimate. A reefer probe moved 30 centimetres during a repair changes the meaning of every subsequent reading; a tag re-applied to a different face of a case changes read rates; an over-the-air update that halves a sampling interval to save battery quietly rewrites the physics a model was trained on. Stage 4 treats these as configuration changes with a review, a record and a re-baseline — which is why the stage is reached jointly by an engineering team and a maintenance function, never by either alone.

Stage 5 is narrower than it sounds, and the operators who reach it are the least interested in calling it autonomy. What distinguishes it is that sensing has become a managed portfolio with a feedback loop of its own: each device class is reviewed against the decisions it feeds and the value those decisions produce, and classes no decision reads are retired rather than renewed. That single discipline reverses the accumulation dynamic every large estate suffers, in which each project adds a device population and none is ever switched off.

The closed loops themselves are deliberately small: a reefer setpoint corrected within a defined band when ambient conditions and door events predict an excursion; a dock appointment moved when an arrival prediction shifts past a threshold; a returnable-asset repositioning triggered by tag counts falling below a depot minimum. Each has a stated bound, an escalation path outside it and a reconstructable trail. Anything with regulatory exposure — releasing a pharmaceutical consignment, closing a temperature deviation, changing a dangerous-goods classification — stays with a person indefinitely, and correctly so.

What most clearly separates stage 5 from a well-run stage 4 is designed degradation. Telemetry always stops eventually: a vessel crosses a satellite gap, a roaming agreement changes, a battery cohort reaches end of life, a storm takes out a depot gateway. At stage 5 each decision has a specified behaviour for that moment — fall back to schedule, widen the bounds, suppress and escalate, freeze the last verified state — exercised on purpose rather than discovered during an incident. Sustaining the stage is a physics and governance problem rather than a modelling one: battery chemistry, spectrum availability, certification and driver-privacy law all bound what an estate may observe, and those bounds move.