Kodiak’s 99% Isn’t a Safety Factor: What Hurdles Do Driverless Trucks Really Need to Overcome?

In news about autonomous driving, numbers can easily overshadow explanations. “99%” sounds almost perfect, but without knowing what the denominator is, it has almost no informational value. For a heavy-duty truck designed to operate without a driver in the cab, this distinction is particularly important and cannot be overlooked.

On October 30, 2026, Kodiak announced that its Autonomy Readiness Measure for long-distance autonomous driving had reached 99%. The company defines this as the proportion of claims and evidence that have been substantially completed in the safety validation process. This is not an accident avoidance rate; it does not mean that 99 out of every 100 trips are safe; nor is it a safety certification issued by a regulatory agency. The company still plans to launch long-distance driverless commercial operations between Dallas and Houston, Texas, by the end of 2026, following the completion of its safety validation.Kodiak's Announcement

This news is worth paying attention to, but what’s noteworthy isn’t the numbers themselves—it’s how autonomous driving companies are beginning to describe the evidence that they are “ready.”

Just because progress is nearing completion doesn’t mean the risks disappear proportionally.

This can be understood through a somewhat imperfect analogy: just because 99% of a building’s acceptance documentation has been completed does not mean that the building has eliminated 99% of risks. The remaining items may be simple, or they may involve core requirements; the quality and scope of the completed documents must also be considered.

Similarly, while the progress of safety validation helps the public understand the project’s advancement, it cannot, on its own, answer the question of whether the system is capable of handling all road conditions. It is primarily a project and evidence management metric, not a figure that can be directly used to rank the system against other vehicles“ ”safety scores.”

For readers, the most effective questions to ask are not “Why is there still a gap of 1%?” but rather: What operating conditions does the analysis cover? What does “completion” mean? Which data comes from simulations, and which from actual vehicles? When boundary conditions change, which verification steps need to be repeated?

These issues are more important than chasing numbers. This is because the commercial promise of autonomous driving lies not in putting on a one-time demonstration, but in consistently providing repeatable services under agreed-upon conditions.

Why should we take into account the possibility that “the system is functioning normally, but errors may still occur”?

In a previous technical white paper, Kodiak used PRA to assess risk, incorporating scenario frequency, collision probability, and severity into its analysis; its internal BreakPoint tool proactively identifies system vulnerabilities during simulations. The company emphasizes that the issues identified are fed back into the risk assessment and engineering iteration processes.Kodiak's Technical Note on PRA and BreakPoint

This provides a useful introduction to the topic: problems with autonomous driving don’t only occur when “sensors fail” or “computers crash.” Even when all components are functioning as designed, certain combinations of scenarios may fall outside the scope of what the system has been thoroughly tested for.

From an analytical perspective, actively seeking out failure conditions is more persuasive than simply repeating successful scenarios. The fact that a turn is executed smoothly under normal traffic conditions does not replace the need to check for obstructions, illegally parked vehicles, or sudden changes. The value of validation lies in transforming “the unexpected” into “what we already know and have a basis for handling.”

However, simulations should not be viewed as complete substitutes for reality. How the model is configured, the validity of the input data, and how the boundaries are defined all influence the conclusions. For any simulation-driven safety claim, it is important to consider both how real-world data is used to calibrate the simulation and whether the original evidence remains valid after the system is modified.

Once it was observed that the cab was unoccupied, the liability for the business changed.

According to the partnership agreement announced by Kodiak and IKEA in September 2026, the two companies already have experience with deliveries involving safety observers; the planned initial unmanned route spans 219 miles along the I-45 corridor in Texas and is part of a more comprehensive delivery chain. Previous operations and the planned unmanned operations are not at the same stage and should not be conflated to suggest that “fully unmanned delivery is already in place.”Announcement Regarding the Partnership Between Kodiak and IKEA

Behind this distinction lies a change in the structure of responsibility. When an observer is on board, the company still maintains a requirement for human intervention on-site; once the driver’s cab is unoccupied, the handling of anomalies can no longer be vaguely left to the principle that “humans will resolve the issue when necessary.”

From the customer’s perspective, what they are purchasing is not just an algorithm that can drive, but a complete set of transportation services. How departure conditions are confirmed, how vehicle malfunctions are handled, how maintenance windows are scheduled, and how delays in cargo delivery are communicated—all of these factors contribute to whether the service can be trusted.

This means that when commercializing autonomous driving, the comparison should not be limited to driving capabilities alone; it must also take into account the operational systems. Technical demonstrations most easily showcase a few minutes of a vehicle on the road; real customers, however, care about whether services are completed as scheduled every day, whether issues can be managed when they arise, and how much coordination is required to restore service.

Why Restricting Routes Is Not Simple Evidence of “Incompetence”

The public often regards versatility as the sole measure of advancement: a system is considered truly mature only if it can operate anywhere; one that operates only on specific routes seems to be merely a work in progress. For safety-critical systems, this view is overly simplistic.

Providing reliable service within a clearly defined scope can, in and of itself, have commercial value. Defining specific routes allows companies to place road characteristics, operating hours, handover procedures, and maintenance systems within a more concrete validation framework. While this does not prove that the system is applicable elsewhere, it should not, for that reason, negate the value of the existing use case.

It is important to exercise caution when deriving universal capabilities directly from a limited environment. Evidence gathered on the Texas Corridor cannot automatically serve as a substitute for validation under different road conditions, weather conditions, or operational models. Expanding the scope should be treated as a new engineering effort, rather than simply enlarging the service area on a map.

Therefore, when assessing a company’s future progress, it is helpful to evaluate “completion of the first route” and “expansion of the scope of replicability” separately. The former validates whether the company can get off the ground, while the latter determines the cost and pace of growth.

What types of subsequent evidence can clients and observers view?

The first category is operational boundaries. Does the company clearly specify the conditions under which it provides services, the conditions under which it suspends them, and how it safely handles situations after a suspension? Clear boundaries do not weaken the technology; rather, they give its capabilities a verifiable meaning.

The second category is sustained operations. Don’t just focus on the initial footage of an unmanned vehicle taking its first ride; you should also consider whether subsequent services are stable, whether maintenance can be planned, and whether business partners are willing to continue the collaboration. There is a long way to go in terms of management between a single success and sustained delivery.

The third category is change management. After software updates, changes to vehicle configurations, or route changes, how can a company determine whether previous conclusions still hold true? If corresponding validation logic is provided with every expansion of capabilities, it will be easier for the outside world to understand how growth and safety can be reconciled.

The fourth category is the transparency of metrics. Percentage of progress, mileage, on-time delivery, and risk assessment each address different questions. Combining them into a single “leadership” ranking can easily lead to distortions; only by examining them separately can one identify true strengths and unknown factors.

After 99%, the real news will be how it’s implemented

This announcement from Kodiak indicates that the project is nearing its launch, but whether it will go live on schedule and how stably it will operate remains to be seen. The startup team, led by Founder and CEO Don Burnette, faces the challenge not only of getting the vehicles on the road but also of turning its engineering commitments into a service that customers can rely on over the long term.

For readers familiar with “physical AI,” this may be more noteworthy than yet another model ranking: when software output directly corresponds to real-world motion, the evaluation criteria must shift from “appearing intelligent” to “being reliably, evidence-based, and sustainable within a clearly defined scope.”

99% can tell us how far along the project is, but it cannot replace our judgment about the next steps. What really matters is not rounding up that last percentage point to a nice round number, but ensuring that every completed run stands up to the same rigorous scrutiny.