AI for auto parts chains: one model for every store's returns, warranty and catalog data

A return starts as a conversation at the counter and a few words typed under time pressure. The reason code, if there is one, is picked from a list by whoever is on shift. New hires, busy Saturdays and stores with their own habits all leave their mark. The result is returns data that can't be compared store to store, and supplier warranty claims built on reasons nobody trusts.

The training data is headquarters' own history: returns, warranty returns or catalog items that its own analysts have already coded. That's the most consistent version of the coding standard that exists.

StayCharted learns from it, checks the examples for contradictions first, and tests the model on cases it never saw. The report shows how often it's right for each reason or category and which ones it confuses, before any store sees an answer. The model is published once and every channel asks it:

A store sees the answer, its confidence, and nothing else. Its access code reaches that one model. It can't see the training data, other models or other stores' answers. Store staff don't need a StayCharted account, and headquarters can stop a store's code at any time. Independently owned stores can use it the same way, without access to anything else in the chain's account.

• Return reason coding: the counter note coded to the chain's return reasons, so returns can be counted by reason, part number, supplier and store.
• Warranty returns to suppliers: returns coded to the defect reasons a supplier's warranty program uses, so the claims sent to suppliers are consistent.
• Part terminology classification for supplier product data: new items from suppliers, classified into the chain's own category tree or the industry's part terminology (the Auto Care Association's PCdb, used with PIES). A missing or wrong part type can make a product harder to find in category-based searches.
• Photos of returned and damaged parts: a store's photo of a part that came back, or arrived damaged, classified with a picture model, from the phone app.
• Store feedback and requests: messages from stores to headquarters routed to the right team on arrival.

The model files text and photos into categories. It doesn't decide fitment, check a VIN, or match part numbers to a vehicle; those stay with your catalog and lookup systems.

One coding standard, applied the same way everywhere. Every store uses the same published model and coding standard. Different descriptions can still produce different answers; use the report and review queue to find and correct inconsistencies.

Every unsure answer in one place. Answers below the confidence cutoff go to headquarters' Review Queue, marked with the store's code name, so the people who own the standard decide them.

Corrections that improve the model for every store. Each decision becomes a training example. When headquarters publishes the next version, every store's next answer comes from it. The published version keeps answering while the next one trains.

Three of our public benchmarks are close to these jobs:

None of these is a parts chain's own returns data. Counter notes are short, and your reasons may overlap, so your report will differ. It shows, for your own reasons and categories, how often the model is right and which share of answers can go straight through.

Each store's access code has its own limits, set by headquarters: how many phones can use it (10 by default), answers a day across those phones (500 by default), when it expires (90 days by default), whether photos are allowed, and pause, turn back on, revoke or remove a phone. Every answer counts against headquarters' workspace, so the chain sees and pays for the network's use in one place. Systems that call the API use API keys headquarters creates and can revoke.

Access codes require Essentials or Business and the workspace Owner’s permission. Network usage draws from workspace allowances; contact us to agree capacity and limits for a large store network.

• Your own database schema, holding only your organization's data.
• Encrypted at rest and in transit, hosted on AWS in the United States.
• Used only for your models. Your data trains your models and nobody else's.
• Your own cloud, if you prefer. On the Business plan, models can be trained in your own Amazon Bedrock or Microsoft Foundry account.

Security and your data →

Training in your cloud still sends data through StayCharted to prepare and run the workflow. See how cloud training works.

1. Week 1: export a few thousand returns your returns team has coded, with the counter note and the reason. Remove customer names, phone numbers and card details.
2. Week 1: train, read the report per reason, and pick the confidence cutoff.
3. Weeks 2 to 3: give access codes to ten stores, or connect the returns screen of one region to the API. The returns team works the Review Queue a few times a week.
4. Week 4: publish the version trained on their corrections and compare the reports.

Start free → · Talk to us about a store network pilot →