What AI agents change in the parts business

AI agents are AI systems that can carry out a sequence of steps toward a goal: read, look things up, decide, act and check their work. In the parts business, much of the costly work looks exactly like that. Processing an engineering change, onboarding a supplier's catalog or answering a dealer who has no part number are each a dozen small decisions in a row. Agents make that work continuous instead of periodic, with people approving what goes out.

The research group METR measures AI progress in a way that maps well to real work: the length of task, measured by how long it takes a skilled person, that an AI system can complete successfully half the time. In 2019, GPT-2 managed tasks that take a person about 3 seconds. GPT-4, in 2023, handled tasks of about 6 minutes. By late 2025, leading models were completing tasks that take a skilled person around 4 hours. Over six years, that length has doubled about every seven months.

METR's tasks are software and reasoning problems, and the measure has wide error bars. The direction is still hard to miss: AI has moved from finishing a sentence to finishing a job.

Most of what keeps a parts catalog accurate is a chain of small, connected decisions. A few examples:

• An engineering change. A supplier discontinues a valve. Someone has to find every unit built with it, identify the replacement, confirm it fits each configuration, write the supersession, update the description and any kits, and flag open orders.
• A new supplier catalog. A component maker publishes a 400-page PDF. Someone has to match its numbers to the OEM's parts, pull out specifications and images, and note which parts it doesn't cover.
• A question without a part number. A dealer describes a part from a photo and a few words. Someone has to find the unit, narrow the candidates, ask a clarifying question and confirm what's orderable.

Today, each of these waits for a person with time. That's why catalogs fall behind between releases.

Here's how the first example plays out with an agent doing the legwork. The steps are illustrative.

1. Notice. The change is released in the OEM's engineering system.
2. Read. The agent reads the change: old part, new part, effective date, reason.
3. Find the machines. It identifies every unit whose build record includes the old part, by serial range and configuration.
4. Check fit. It confirms the new part works on each of those configurations, and separates the units where it doesn't.
5. Draft the updates. It writes the supersession, updates the part's description and specifications, and adjusts any kits that include it.
6. Ask for approval. It sends a summary to the OEM's parts team, with the evidence for each decision.
7. Publish and record. Once approved, the updates go live everywhere the catalog appears, with a record of who approved what and when.

What used to wait for the next catalog release can now be done the week the change is made.

Agents are good at doing the steps. Deciding what the business should do is still a person's job. In practice, that means three things.

• Approval before publishing. The OEM's team reviews changes before they reach dealers and fleets, especially anything that affects safety or fit.
• Evidence with every decision. Each proposed change shows the records it relied on, so a reviewer can check it in seconds.
• A full record. Every change can be traced: what triggered it, what the agent proposed, who approved it.

As trust builds, an OEM can let routine changes, such as a corrected description, publish with lighter review, and keep closer review for the rest.

• OEM catalog teams spend their time reviewing and deciding instead of re-keying. The catalog keeps pace with engineering.
• Dealers find the current part the first time, because the supersession is already in place when they look.
• Fleets can see changes that affect their own machines, such as a superseded component or a recall, attached to each asset.