Fleet analytics programs rarely fail because the model was wrong. They fail because someone built an analysis on a stream that could not carry it. The position record was too coarse for the dwell question. The event feed was a vendor threshold, not a measurement. The engine data was there but the odometer it was joined against came from a different clock. Sorting telematics streams by what they can actually settle is most of the work, and it is work that happens before any modeling starts.
Three sources in one box, three different data qualities
A telematics unit looks like one device and behaves like three. Keeping them separate in your head is the fastest way to stop over-trusting a number.
The satellite receiver. Position, ground speed, heading, and a timestamp, produced independently of the vehicle. It knows nothing about the truck and everything about where the truck is. It degrades in exactly the places logistics cares about most: under dock canopies, inside container yards, in downtown street grids, in tunnels. It answers where and when. It never answers why.
The vehicle network. On a Class 7 or Class 8 truck this is SAE J1939, the communications standard the trucking industry built so components from different manufacturers could share one bus. The National Motor Freight Traffic Association, the freight industry's research body, names J1939 directly as the standard that makes mixing and matching parts practical and, in the same breath, widens the attack surface. Off that bus come engine hours, the ECM odometer, fuel rate, engine speed, coolant and oil readings, brake and ABS status, PTO engagement, and diagnostic trouble codes. Light-duty vans in a mixed fleet speak a different dialect entirely, which is why mixed fleets have harder data problems than pure tractor fleets.
The device's own sensors. An accelerometer, a gyroscope, sometimes a camera. This is where hard-braking, harsh-cornering, and rapid-acceleration events come from. They do not come from the truck. They are the vendor's interpretation of a motion trace against a threshold the vendor selected, and two units from two vendors on the same tractor will disagree. That does not make the stream useless. It makes it a derived product, and derived products need their assumptions written down before anyone builds a driver scorecard on top of them.
Start with the streams somebody already audits
The cheapest high-quality data in a fleet is the data a regulator already forces you to produce correctly. The collection cost is sunk, the format is specified, the definitions are settled, and somebody outside the company has an interest in the record being right. Two streams qualify.
Hours of service. Under 49 CFR 395.8, motor carriers have had to install electronic logging devices and require drivers to use them since December 18, 2017. The regulation lists narrow exceptions: drivers who complete a record of duty status on not more than 8 days within any 30-day period, driveaway-towaway operations where the vehicle being driven is itself the shipment, driveaway-towaway of a motor home or recreation vehicle trailer, and vehicles manufactured before model year 2000 as reflected in the VIN on the registration. Everyone else is producing duty-status data continuously, whether or not anyone in the office is reading it.
What that data encodes is a hard scheduling constraint. 49 CFR 395.3 sets it out: a driver may not drive without first taking 10 consecutive hours off duty; may not drive after a period of 14 consecutive hours after coming on duty following those 10 hours; may drive a total of 11 hours inside that window; may not drive if more than 8 hours of driving time have passed without at least a consecutive 30-minute interruption; and is capped at 60 hours in any 7 consecutive days or 70 hours in any 8, with a 34-hour or longer off-duty period ending the applicable period. Dispatch software that does not consume live duty-status data is guessing at the single constraint that most reliably breaks a delivery promise.
Jurisdictional distance and fuel. The IFTA Procedures Manual is unusually specific about what a GPS-derived distance record has to contain, and it is worth reading the actual text rather than a vendor's summary. Section P540.200 requires that where distance records are produced by a vehicle tracking system using latitude and longitude, a record be created and maintained at a minimum every 10 minutes when the vehicle's engine is on, containing the date and time of each system reading, the latitude and longitude to a minimum of four decimal places, the ECM odometer reading at each reading, and the vehicle identification or unit number. Section P510 requires licensees to retain those records for four years following the date the return was due or filed, whichever is later. Section P520 puts the burden of proof in an audit on the licensee, not the jurisdiction.
The penalty for failing that standard is written down, which makes it one of the few places in fleet analytics where the value of good data has a number attached. Under P570.100, when records do not meet the adequacy criterion, the base jurisdiction imposes an additional assessment by reducing the licensee's reported fleet fuel economy to 4.00 miles per gallon, or by reducing the reported average fuel consumption factor by 20 percent, or by increasing jurisdictional distance by 20 percent. A fleet running 6.5 miles per gallon that gets assessed at 4.00 is paying fuel tax on a fiction, across every jurisdiction, for every quarter in the audit period.
The precision trap
Here is the mistake that quietly wrecks more fleet data projects than any modeling error: assuming that because a position log exists, it is good enough for a position question.
FMCSA's published guidance on the ELD rule states that during on-duty driving periods, recorded location accuracy is approximately within a one-mile radius, and that when a driver operates a commercial vehicle under personal conveyance, position reporting is approximately within a ten-mile radius. That is a deliberate privacy design. It is also, at one mile, roughly a hundred and forty times coarser than the four-decimal-place requirement IFTA sets for distance records, which resolves to about eleven meters of latitude.
The same satellite chip is producing both. The difference is entirely in what the vendor chose to store and expose. A platform that persists only what the hours-of-service rule requires has thrown away the resolution your detention analysis, your yard-dwell analysis, and your geofence arrival timestamps depend on, and it cannot be recovered after the fact. Ask, in writing, what position resolution and what sampling interval the platform retains in its export, and ask before you sign rather than during the first analysis.
| Stream | Where it originates | Practical resolution | What it can settle |
|---|---|---|---|
| Duty status | ELD, per 49 CFR 395 subpart B | Status changes with position to roughly a one-mile radius on duty | Available hours, restart timing, whether a load is legally deliverable today |
| IFTA distance | Vehicle tracking system plus ECM odometer | Four decimal places of latitude and longitude, at least every 10 minutes engine-on | Jurisdictional miles, lane-level routing, dwell and detention at a facility |
| ECM parameters | SAE J1939 vehicle network | Sub-second on the bus, usually downsampled hard by the telematics unit | Fuel rate, idle and PTO hours, engine load, true odometer |
| Fault codes | Vehicle ECUs, read off the bus | Event-driven, with vendor-specific decoding | That a threshold was crossed. Not that a failure is coming |
| Motion events | Device accelerometer and gyroscope | Vendor threshold applied to a raw trace, rarely exposed | Relative comparison within one vendor's fleet, nothing across vendors |
| Video | Road-facing and driver-facing cameras | Clip-triggered, retention set by policy | What actually happened in a specific incident |
Ranking the streams by decision value
The ranking below is our own reading of the field, and the axis is narrow on purpose: how often a stream changes a decision that costs real money, discounted by how much engineering it takes before the raw record can be trusted. A stream can be fascinating and rank low because nothing changes when you look at it. Treat this as a starting order of operations for a fleet that has not built anything yet, not as a measured statistic.
Telematics streams by decision value — first build
Editorial weighting from public sources and practitioner reading — illustrative, not a measured statistic.
Fault codes are not failure predictions
A diagnostic trouble code is a report that an electronic control unit observed a value outside a band its manufacturer set. It is a threshold crossing. Fleets that pipe every code straight into a work-order queue discover within a month that the queue is mostly noise, and the shop learns to ignore it, which is worse than never having built it.
The signal that survives contact with a maintenance department is the pattern, not the event: which codes recur on which unit, at what mileage since the last relevant service, under what duty cycle, at what ambient temperature. That is a modest data-engineering job and a small amount of statistics. It is also the honest version of what most vendors market as predictive maintenance.
The economics deserve stating plainly, because they are arithmetic and they decide whether the project is worth doing. Suppose a fleet of 500 tractors, and a component that fails on 3 percent of units per year — 15 failures. A detector that catches 10 of those 15 and raises a false alarm on 10 percent of the healthy 485 produces about 58 flags to find 10 real problems. Roughly one flag in six is a catch. Whether that is excellent or worthless depends on one ratio: the cost of an unnecessary inspection against the cost of a roadside failure with a loaded trailer, a towed tractor, a missed appointment, and a service failure on the customer scorecard. Fleets that know that ratio for their top ten failure modes can decide in an afternoon which ones are worth a model. Fleets that do not know it will argue about model accuracy for a year and never ship.
Safety data and the scoreboard that actually counts
Every telematics platform sells a driver safety score. Very few operators notice that it is not the score their insurer, their brokers, or their shippers can see.
FMCSA's Safety Measurement System organizes roadside inspection and crash data into seven Behavior Analysis and Safety Improvement Categories: Unsafe Driving, Crash Indicator, Hours-of-Service Compliance, Vehicle Maintenance, Controlled Substances and Alcohol, Hazardous Materials Compliance, and Driver Fitness. That is the public-facing record built from what an inspector wrote down at a scale house, not from what an accelerometer inferred in the cab.
This reframes the useful question about a driver-behavior stream. It is not "does this score correlate with crashes," which is nearly unanswerable at single-fleet sample sizes. It is "does acting on this stream change our inspection and violation record." That question is answerable within a year at moderate fleet size, using data the fleet already receives, and almost nobody runs the measurement.
Two of the seven categories are directly reachable with the work described in this article. Hours-of-Service Compliance is preventable at dispatch, before the truck moves, if planning consumes live duty status instead of yesterday's summary. Vehicle Maintenance is preventable in the shop, and it responds to exactly the recurring-fault-pattern work above. Unsafe Driving is behavioral, moves slowly, and is where coaching programs live. A fleet that wants a measurable safety result inside twelve months should spend its engineering on the first two.
Cameras, consent, and the program that dies in month three
Driver-facing video is the richest stream a fleet can collect and carries the largest adoption risk. Two separate issues get conflated and both need answers before procurement.
The first is legal. Several states regulate the collection and storage of biometric identifiers under statutes distinct from general privacy law, and a driver-facing system that performs facial recognition, gaze tracking, or drowsiness inference from facial geometry can fall inside those definitions depending on how the vendor implements it. Some of these statutes carry a private right of action. This is a question for counsel and for the vendor's technical documentation before a purchase order, not a question to discover after a thousand units are installed. It is also genuinely in flux: state legislatures have been active in this area and the answer that was right two years ago may not be right now.
The second is operational and is the one that actually kills programs. Who can pull a clip, and on what trigger. How long clips are retained. Whether footage is reviewed for coaching only or is admissible in a disciplinary process. Whether it will be produced in litigation. A written policy that drivers see before installation is worth more to the program's survival than any feature on the camera. Fleets that install first and write the policy after an incident spend the next year rebuilding trust they did not have to lose.
Fuel and idle: the clearest arithmetic in the building
Idle time is the one place in fleet analytics where the whole chain fits on one page. Engine-on-not-moving hours come off the vehicle network. Fuel burn per idle hour is measurable on the same bus. Fuel price is known. The intervention is a policy change, an auxiliary power unit, shore power at the terminal, or a conversation with a driver. Nothing in that chain requires a model, and the result is defensible to a CFO in a single table.
The same is true of fuel economy ranked by unit, by driver, and by lane. With enough observations per cell to be stable, a ranking is a management tool. It tells a maintenance manager which tractors are drifting and a terminal manager which lanes are quietly expensive. Every fleet should build this before it builds anything with the word predictive in the name, because it is fast, it is cheap, and it establishes whether the underlying data joins correctly. If unit-level fuel economy comes out implausible, the problem is in the pipeline, and you have found it for the price of a week rather than the price of a modeling project.
Fuel data also travels outward. EPA's SmartWay program, now past its twentieth year, works with more than 4,000 partner shippers, carriers, and logistics companies on freight efficiency, and its Excellence Award recognizes the top two percent of participating shippers and truck carriers. When a shipper customer participates, a carrier's fuel and emissions reporting stops being an internal exercise and becomes part of the commercial relationship. That changes what "good enough" means for the data.
Regulatory reporting is a data problem before it is a mechanical one
State emissions programs are increasingly built on periodic data submission rather than on a physical inspection alone, and they reach fleets that are not domiciled in the state. California's Clean Truck Check program is in effect and, per the Air Resources Board, applies to subject vehicles including vehicles registered outside California when operating in the state. Compliance involves reporting the vehicle, paying an annual compliance fee, and submitting emissions test results.
For an out-of-state fleet that runs into California occasionally, that is an obligation triggered by an operating fact the telematics system already knows and the compliance department often does not. Connecting jurisdictional presence to a reporting obligation is a small, high-return piece of engineering. Requirements in this area change, and the right source is the agency's own current guidance rather than a vendor's summary or an article, including this one.
The vendor questions to ask before the contract, not after
FMCSA's ELD registration list is a self-certification list. The agency states it plainly on the page: the listed devices are self-certified by the manufacturer, and FMCSA does not endorse any electronic logging device. There is a separate Revoked ELDs list for devices removed from the registry. Registration is therefore evidence that a manufacturer made a claim, not evidence that anyone verified it.
That has a direct operational consequence, and it is written into 49 CFR 395.34. When an ELD malfunctions, the driver notes the malfunction and provides written notice to the carrier within 24 hours, reconstructs the record of duty status for the current 24-hour period and the previous 7 consecutive days on paper, and continues manual records until the device is fixed. The carrier must correct the malfunction within 8 days of discovery or of the driver's notification, whichever comes first, with an extension available by notifying the FMCSA Division Administrator within 5 days. That 8-day clock is a service-level requirement in disguise, and it belongs in the contract rather than in the operations manager's inbox.
The rest of the diligence is about getting your own data back out. Dashboards are not data. Ask for these in writing.
- Full-resolution position records with timestamps, not summarized trip rows — and the retained decimal precision stated as a number
- ECM parameter samples with the source identity preserved, so a fuel-rate value can be traced back to the bus parameter it came from
- Event records that include the threshold value the vendor applied, because an event without its threshold cannot be compared to anything
- Bulk historical export on termination, in a documented format, at no additional charge, with a stated delivery window
- API rate limits and retention windows written into the agreement, not published on a help page the vendor can change
- A stated replacement path and timeline if a device is removed from the FMCSA registered list
Scoping a first build that can fail cheap
The pattern that works is small, historical, and aimed at one decision. The pattern that does not work is a platform.
Pick a decision, not a dataset. "Should this tractor go into the shop this week" is a decision. "Build a maintenance dashboard" is not. The decision determines which streams matter and, more usefully, which ones you can ignore.
Write the counterfactual before writing code. What will the operation do differently when the system says something, what does that action cost, and what does it save when it is right. If nobody can answer, the analysis will be interesting and unused.
Reconstruct twelve months of history first. Before a single live integration, pull the export and rebuild the last year. This does two things: it tells you whether the historical record is even joinable, and it lets you measure a proposed rule against outcomes that already happened, at zero operational risk. Most projects that die in production would have died here for a hundredth of the cost.
Evaluate on the operational metric. Not model accuracy. Roadside failures avoided, detention hours recovered, violations per inspection, gallons per mile against the prior period. If the operational metric does not move, the model being accurate is not a defense.
Agree in advance on the number that kills it. A threshold written down at the start is a discipline. The same threshold proposed after six months of investment is an argument nobody wins.
Bottom line
The streams worth acting on first are the ones a regulator already forces you to produce, because their definitions are settled and somebody outside the company has an interest in their accuracy. The stream most likely to be over-trusted is the vendor's behavior score, because it is a derived product with an undisclosed threshold. The mistake that costs the most is assuming a position log built for a compliance purpose has the resolution to answer an operational one. Get the hours, the miles, the fuel, and the idle right, prove the pipeline on a year of history, and the harder work becomes tractable. Skip that and the harder work stays a demo.
Common objections
Our telematics vendor already gives us all of this in their dashboard.
A dashboard answers the questions the vendor anticipated. The value in fleet data is usually in a question specific to your lanes, your customers, and your equipment mix, and that question needs the underlying records. The test is simple: request a raw export and see what precision, what sampling interval, and what history actually arrives. That result tells you more about the platform than any feature list.
We are too small for this to pay off.
The compliance-adjacent work scales down further than people expect, because the penalty exposure is proportional. An IFTA adequacy assessment applies to the fleet as a whole regardless of its size, and idle arithmetic is the same arithmetic on 40 tractors as on 400. What does not scale down is a modeling project on a rare failure mode, where a small fleet may not generate enough events per year to learn anything.
Drivers will not accept more monitoring.
Most of what is described here is data the fleet already collects for regulatory reasons and is not new monitoring at all. The genuine tension is around cameras and behavior scoring. The programs that hold up are the ones where the retention policy, the review triggers, and the coaching-versus-discipline line were written and communicated before installation.
We tried predictive maintenance and it did not work.
The common cause is a model built without the cost ratio behind it. A detector that flags six units to find one real problem is a success when a roadside failure is expensive and an inspection is cheap, and a failure when the reverse is true. Establishing that ratio per failure mode, before modeling, changes which components are worth the effort and usually shortens the list considerably.
Frequently asked questions
Duty status and available hours, then jurisdictional distance and fuel, then idle and PTO hours. These are already collected for regulatory reasons, their definitions are fixed, and each maps to a decision someone makes every day. Behavior scores and video are valuable but should come after the basic streams are proven to join correctly.
Generally no. FMCSA guidance describes recorded location accuracy during on-duty driving as approximately within a one-mile radius, and roughly a ten-mile radius under personal conveyance. IFTA, by contrast, requires latitude and longitude to at least four decimal places at 10-minute intervals when the engine is on. If the platform only retains the coarser record, facility-level dwell analysis will not hold up.
For IFTA purposes, the Procedures Manual requires licensees to retain records of operations for four years following the date the return was due or was filed, whichever is later, plus any period covered by waivers or jeopardy assessments. The burden of proof in an audit is on the licensee. Other retention obligations, including hours-of-service supporting documents and any video policy, run on their own schedules.
Under 49 CFR 395.34 the driver notes the malfunction, gives the carrier written notice within 24 hours, and reconstructs the record of duty status for the current 24-hour period and the previous 7 consecutive days on paper, continuing manually until it is fixed. The carrier has 8 days from discovery or driver notification, whichever is first, to correct it, with an extension available by notifying the FMCSA Division Administrator within 5 days.
No. The registration list is built on manufacturer self-certification, and FMCSA states directly that it does not endorse any electronic logging device. The agency maintains a separate Revoked ELDs list for devices removed from the registry, which is worth checking during a contract as well as before signing one.
The Safety Measurement System is built from roadside inspection and crash records across seven categories, including Hours-of-Service Compliance and Vehicle Maintenance. Those two are the categories most directly reachable by data work: hours violations are preventable at dispatch, and maintenance violations respond to recurring-fault-pattern analysis in the shop.