Rapid root-cause diagnosis on field-deployed heavy autonomous vehicles, where every hour of downtime at a customer site costs millions. ASI and SoftBank Robotics partnered with Gritmind to build AI-powered software that compresses an investigation that historically took three and a half days into one that takes twenty minutes.
Why this mattered
ASI automates machinery for agriculture, logistics, construction, and other heavy industrial markets. Their customers operate vehicles in places where downtime is paid for in millions per hour. When a truck stops unexpectedly, ASI’s engineers treat it as an incident. Each incident generates gigabytes of log data covering minutes of vehicle time. Determining root cause used to take anywhere from two days to two weeks of senior engineering work, while the truck stood idle and losses grew.
Why this mattered
ASI automates machinery for agriculture, logistics, construction, and other heavy industrial markets. Their customers operate vehicles in places where downtime is paid for in millions per hour. When a truck stops unexpectedly, ASI’s engineers treat it as an incident. Each incident generates gigabytes of log data covering minutes of vehicle time. Determining root cause used to take anywhere from two days to two weeks of senior engineering work, while the truck stood idle and losses grew.
After a major round of investment, ASI’s leadership saw an opportunity to leapfrog the existing approach to incident response: not optimize the steps, but rebuild the workflow AI-first. Gritmind built the system that did it: a multi-agent AI workflow that runs the investigation end-to-end.
What Gritmind built
A multi-agent AI system that investigates an autonomous-vehicle incident end-to-end:
A multi-agent AI system that investigates an autonomous-vehicle incident end-to-end:
- Parses raw vehicle logs automatically, at gigabyte scale per incident.
- Reconstructs the incident timeline.
- Hands off to specialist sub-agents by domain (brake, throttle, planning, and others) to examine their part of the logs and report back.
- Surfaces candidate root causes with click-through evidence to the source log lines.
- Visualizes how events propagated across vehicle subsystems (the team calls this the evidence-board flowchart).
- Lets the engineer pin the verified findings. The user owns the final summary, not the model.
How we built it
Three decisions compressed the timeline.
Three decisions compressed the timeline.
- Twenty-five user interviews in three days, on-site, with engineers in the room.
Gritmind ran user research at ASI’s site with engineers attending every session. Bringing engineering into the user interviews rather than translating findings through documentation is what made the rest of the build possible. The team had usable personas and pattern recognition before they left. - Engineers building from observation, not tickets.
During the interviews, a Gritmind engineer noticed a recurring pain point: users couldn’t easily see how a failure had propagated across vehicle subsystems. He turned that observation into a feature, the evidence-board flowchart, that no PM had requested and no designer had specified. - AI-native from day one.
On this project, AI showed up from day one, both in the process and in the product. During discovery, the team used AI to synthesize user-research notes and draft personas in real time. Then the team carried that AI-native posture through the proof of concept itself, where a multi-agent architecture became the system, not a layer on top of one.
The product decision that made it land
Early usability testing surfaced something more interesting than accuracy. It surfaced calibrated trust. Some engineers were skeptical of AI conclusions. Others trusted them too quickly. In one case, a user validated a wrong answer by asking the AI itself to confirm it.
Early usability testing surfaced something more interesting than accuracy. It surfaced calibrated trust. Some engineers were skeptical of AI conclusions. Others trusted them too quickly. In one case, a user validated a wrong answer by asking the AI itself to confirm it.
Gritmind made an intentional product change. The AI no longer writes the incident summary. It produces a list of candidate problems with supporting evidence. The engineer pins the ones they’ve verified. The summary is the engineer’s, not the model’s. Skeptical users get the verification surface they want. Optimistic users are forced to slow down enough to check. Trust but verify is the team’s North Star on the product.
Outcomes
The proof of concept was running end to end three weeks after the on-site workshop. Verified results:
The proof of concept was running end to end three weeks after the on-site workshop. Verified results:
- 3.5 days → 20 minutes. A documented past incident from ASI's 2-to-14-day historical investigation range was reproduced and resolved in twenty minutes, verified accurate against the original engineering write-up.
- 30 to 40 minutes by a non-specialist. A customer-support engineer, not senior engineering, pulled an untouched incident from the queue, used Sherlog with light coaching, and identified a CAN bus wiring drop-out on the vehicle in roughly half an hour, confirmed against control-room data. The same incident, run through the standard investigation path, would have escalated to senior engineering and entered the 2-to-14-day queue.
- One hour vs. seven days. A support-team incident that had been queued seven days waiting on engineering produced an actionable answer inside a one-hour user training session.
- In use on live customer incidents. Support engineers are requesting for tooling to become the standard.
“If you can get to seventy-five trust and twenty-five verify, you’re saving millions and millions of dollars and tons of hours of engineering time.”
- Rod Lewis (ASI Development Manager)
What’s next with ASI
With the system in active use, Gritmind and ASI are turning to the next set of high-value problems: ingesting on-vehicle black-box data so investigations don’t require a manual data transfer, optimizing the cost profile so the system scales economically across customer fleets, and extending the same AI-native pattern to adjacent operational workflows where senior engineering time is the bottleneck. The savings show up in two ledgers, engineer-hours pulled off incident archeology and customer downtime avoided.
With the system in active use, Gritmind and ASI are turning to the next set of high-value problems: ingesting on-vehicle black-box data so investigations don’t require a manual data transfer, optimizing the cost profile so the system scales economically across customer fleets, and extending the same AI-native pattern to adjacent operational workflows where senior engineering time is the bottleneck. The savings show up in two ledgers, engineer-hours pulled off incident archeology and customer downtime avoided.
Why this project, and what it means for yours
Some workflows absorb AI as a layer. Others have to be rebuilt around it. Gritmind’s work is in knowing which call is right, and operating on the harder side when the math says rebuild. ASI Root Cause is what that second call looks like when the math works: senior engineering hours pulled off log archeology, customer downtime measured in millions per hour, and a leadership team willing to skip the middle and ask for the end-state. If those conditions describe a system in your operation, the next conversation is worth having.
Some workflows absorb AI as a layer. Others have to be rebuilt around it. Gritmind’s work is in knowing which call is right, and operating on the harder side when the math says rebuild. ASI Root Cause is what that second call looks like when the math works: senior engineering hours pulled off log archeology, customer downtime measured in millions per hour, and a leadership team willing to skip the middle and ask for the end-state. If those conditions describe a system in your operation, the next conversation is worth having.
