From Reports to Real-Time Intervention: Building a Voice Agent for Missed Vehicle Inspections

Before a commercial vehicle heads out, the driver is expected to inspect it: check the tires, lights, windshield, fluid leaks, and other components that determine whether it is safe to operate.

The results are captured in a Driver Vehicle Inspection Report, or DVIR. If something is wrong, the driver documents it so the fleet can address the problem before the vehicle continues on its route.

On Samsara, customers complete hundreds of millions of DVIR workflows every year. At that scale, recording the inspection is only part of the challenge. Fleet managers also need to know when an inspection has been missed—or when a vehicle previously marked unsafe is back on the road.

A report can tell a manager that this happened. But by the time someone finds that report, the vehicle may already be miles into its route.

That led us to a different question: what if the system could intervene while there was still time to change the outcome?

Latent space: find the moment that matter

The underlying problem was not a lack of data. Samsara already knew:

• Whether a vehicle had started moving

• Whether the driver had completed a DVIR

• Whether the vehicle had previously been marked unsafe

• Whether it had traveled beyond a yard or other geofenced area

But individual signals are noisy.

A vehicle moving a short distance inside a yard is different from one beginning its route. A missed inspection is different from a vehicle being driven despite an unresolved safety issue. And an accurate signal is still of limited value if it remains buried in a dashboard, waiting for someone to find it.

The real signal was not any single event. It was the combination of movement, time, location, inspection status, and severity.

Once we could distinguish the moments that demanded action from the much larger volume of routine activity, we saw an opportunity to invert the operating model. Instead of asking fleet managers to continually search reports for exceptions, the system could watch for those exceptions and initiate the response.

Our alpha: a bounded, closed-loop agent

The most useful agent workflows are not always the most open-ended.

A missed DVIR was a strong early candidate precisely because the response was narrow, repetitive, and measurable:

• If an inspection was missing, contact the driver.

• If the driver completed it, close the incident.

• If it remained incomplete, escalate to a fleet manager.

• If the vehicle was already marked unsafe, treat the incident with greater urgency.

The workflow had a clear objective, a constrained set of actions, an observable success condition, and a human fallback.

The agent was not being asked to decide whether a vehicle was safe. It was closing the response loop around an existing safety process: detect the issue, prompt the driver, observe the outcome, and escalate when automation could not resolve it.

That combination made it possible to remove repetitive work from fleet managers without removing them from the decisions that required human judgment.

The result

We built a proactive DVIR workflow inside Samsara’s Incident Center and paired it with an in-cab voice agent that could reach the driver directly.

The workflow operates in four stages:

1. Detect: The system observes that a vehicle has started moving.

2. Validate: After a short grace period, it checks whether the driver has submitted a DVIR.

3. Intervene: If the conditions are met, the Incident Center creates an incident. An automation then triggers a two-way call to the driver through the in-cab device, prompting them to complete the inspection.

4. Resolve or escalate: After the call, the system continues checking for completion. If the driver submits the DVIR, the incident resolves automatically. If not, it escalates to a fleet manager after a defined period. 

One of the important design decisions was knowing when not to intervene. Triggering the moment a vehicle moved could create false alarms when someone was simply repositioning it within a yard. We have built in spatiotemporal thresholds that gave the system enough context to distinguish routine movement from the start of a route.

That tradeoff matters in any real-time agent. Act too quickly and unnecessary interventions erode trust. Act too slowly and the system becomes another report describing what already happened.

Technically, the workflow brings together several parts of the Samsara platform. Real-time telemetry and event detection run in the cloud. Automation connects the incident workflow to the voice agent. The in-cab device becomes the channel for reaching the driver. Incident Center then tracks the outcome and brings in a fleet manager when the bounded workflow cannot close the loop on its own.

Moving from reporting to intervention

The Missed DVIR Agent did not require a new source of truth. It required connecting signals we already had to an action the system could take.

That is the broader opportunity for agents in the physical world. Reports explain what happened. Agents can help change what happens next.

When the signal is reliable, the action is bounded, the outcome is observable, and a human fallback exists, a passive workflow can become a closed loop.

See you in the field.