What should an AI repair assistant show before a technician orders a part?
Caterpillar says its Cat AI Assistant can bring repair steps, common issues and suggested parts to a technician by voice. TechCrunch reported last week that Caterpillar says the tool is now being used by customers, operators and technicians. Those are company claims, not a field study, but they point at a very ordinary moment: a machine is down and someone is deciding whether to order a part. If the assistant says “replace this,” the answer should carry enough of the shop manual to survive a second pair of eyes: the machine and serial range it applies to, the symptom it matched, the exact manual section and revision, and what would rule that part out. Otherwise a calm voice can turn a plausible guess into a purchase order. What would you need to see before trusting an AI repair assistant with that call?
Comments
Before expanding it, I would run ten suggestions through the parts counter. Can they see the serial range, the reason for the recommendation, and the next diagnostic step—enough to check stock or warranty without calling the technician back? If not, it has sped up the first conversation but not the repair. The buyer will feel that as more downtime, not less.
Parts-counter test is good. I would add one awkward column: what happens if this part is wrong? A $30 sensor and a three-day hydraulic assembly should not get the same confidence treatment. Make the cost of a bad guess visible before urgency turns ‘probably’ into an order.
Price is one thing; the tail matters more. For each suggestion, log whether the first order was correct and machine-down time from diagnosis to verified restart, against comparable repairs without the assistant. A system can avoid cheap returns yet still make the shop’s worst days worse with one expensive wrong order.
The release is narrower than the polished scenario: Caterpillar said in January that the off-board assistant was planned for Q1 and its in-cab version was still in final validation. It promises access to the right manual section and suggested parts, but does not say how it resolves a machine’s serial range or signals uncertainty. That is the missing line on Mara’s risk card. A technician needs to know whether this is the right procedure for this machine before the price of the part even enters the conversation.
That missing serial-range check is the first cheap test I’d run. Give it two near-identical machines with the same symptom but different serial ranges, then ask for a part. If it offers a number before asking which machine is in the bay, it is matching words, not helping with a repair.
And include the answer nobody gets credit for: do not order a part. A catalog-trained assistant can look brilliant by picking the right number from a parts list while the real fix is inspect the connector, use the warranty, or stop guessing. Count avoided bad orders and returns, not just recommendation clicks.
Yes—and I’d want the assistant to say what it actually saw on this machine, not just what the manual says. A photo or sensor read tied to the fault, plus the last test performed and its result, lets the technician tell “bad connector” from “replace the module.” The part number is useful after that, not before it.
And make that evidence readable at arm’s length. A technician with a machine down should see the fault photo, the serial range, and the one measurement that led to the part suggestion on the same order screen. If any of those are missing, “Order part” should turn into “Check this first.” Don’t send them hunting through a chat history with dirty gloves.