A, B, C, R: how AI condition grading actually works
AI condition grading is the automated process of assigning a returned item a condition grade from photos taken at receiving, using a computer vision model that outputs a grade, a confidence score, and detected condition tags. At areturnz we grade every returned parcel on a four-tier A/B/C/R scale: A means resellable as new, B means light work or open-box, C means liquidation, and R means destroy or recycle. That single grade is the hinge the rest of returns processing turns on, because it tells you what an item is worth and where it should go before a human has to guess.
What AI condition grading actually is
Condition grading answers one question: what state is this returned item in, and is it good enough to sell again? Traditionally an operator eyeballs the item and writes down a subjective call. The problem is consistency. Two operators grading the same lightly-worn jacket will often disagree, one calling it new-with-tags and the other calling it open-box, and those two decisions send the jacket to two different places with two different recovery values. Automated grading removes that variance by applying the same trained model to every item the same way.
Here is the pipeline in order:
- Photos at receiving. Every returned parcel is photographed as it is opened: the outer label, the opened parcel, the item itself, and any visible defect. This is the raw evidence the grade is built on, and it is captured on every single return, not a sample.
- Vision model inference. The images run through a computer vision model that reads the item against its expected new condition and produces a proposed grade of A, B, C, or R.
- Confidence score and detected tags. Alongside the grade, the model emits a confidence score and a set of condition tags such as scuffs, missing packaging, worn soles, or liquid damage. The tags make the grade legible, so a reviewer sees why a jacket landed at B rather than just being told it did.
- Verification checks. The same vision step runs content verification to catch swaps and mismatches (the wrong item in the box) and quantity or completeness checks (a two-piece set returned with one piece missing).
The output is not just a letter. It is a grade plus a confidence number plus tags plus a full custody chain, bundled as evidence you can inspect in the dashboard or pull through a signed-JSON API. You can see a worked example on our evidence sample.
The A/B/C/R grade rubric
Each grade maps to a condition band and a typical disposition. The grade describes the item; the disposition is the decision that follows from it. Here is the rubric we grade against:
| Grade | Meaning | Typical disposition | Example |
|---|---|---|---|
| A | Resellable as new. Item and packaging intact, no visible wear. | Restock to primary inventory | Unopened sealed electronics returned as wrong-size order |
| B | Light work or open-box. Fully functional, minor cosmetic or packaging issues. | Refurbish, repackage, resell as open-box | Jacket tried on once, tags removed, no wear |
| C | Liquidation grade. Usable but not fit for primary channel. | Route to liquidation or secondary market | Shoes with visible scuffing and creased soles |
| R | Destroy or recycle. Damaged, contaminated, or unsafe to resell. | Recycle, dispose, or certified destruction | Item with liquid damage or a broken seal on a consumable |
Grades feed rules, not gut. Disposition rules combine the grade with item attributes (category, value, hazmat flags, brand policy) to route each return automatically, with any operator override logged against the record. We go deeper on that handoff in how disposition rules turn grades into decisions.
The confidence score and human-in-the-loop
Automated grading does not mean unattended grading. Every proposed grade carries a confidence score, and that number decides whether a human ever sees it:
- High confidence auto-accepts. When the model is clearly sure, the grade is accepted and the return moves straight into disposition. This is what keeps a median cycle time near 48 hours.
- Borderline goes to review. When confidence sits below threshold, or when a tag conflicts with the grade, the item is routed to a human operator who confirms or changes the call.
- Every override is logged. When an operator disagrees with the model, the change is recorded against the return with who, when, and what. That log is both an audit trail and the training signal that keeps the model honest.
Across more than 180,000 returns processed, our AI-vs-operator match accuracy runs about 99.6 percent, meaning the model and a human reach the same grade in nearly every case. The small remainder is exactly where human review earns its place. If you want the mechanics of how thresholds are set, read what a confidence score actually decides on the line.
Why consistency is the real payoff
The headline benefit of AI condition grading is not speed, it is repeatability. A model that grades the same jacket the same way every time removes the operator-to-operator drift that quietly leaks recovery value. When grading is consistent, your restock decisions are consistent, your liquidation pools are cleanly sorted, and your finance team can trust the recovery numbers. Consistency also makes the grade defensible, because every grade ships with the photos and tags that produced it. That is the difference between a claim and proof on every return.
Where grading fits in the wider system
Grading is one stage in a processing node. Photos flow in at receiving, the model grades and tags, rules disposition the item, and webhooks (return.received, graded, dispositioned, restocked) notify your systems at each step. The full spec of how a node captures, grades, and routes is documented on our node spec, and the standardized four-tier grade is what makes that pipeline portable across categories and tenants.
Frequently asked questions
What does A/B/C/R mean in returns grading?
A is resellable as new, B is light work or open-box, C is liquidation grade, and R is destroy or recycle. Each grade maps to a default disposition, so the letter is a shorthand for both the item condition and the recommended next step.
How accurate is AI condition grading?
In our operation the AI grade matches the operator grade about 99.6 percent of the time across more than 180,000 returns processed. High-confidence grades auto-accept and borderline cases go to human review, so accuracy is a blend of model output and logged human confirmation.
What is a confidence score in grading?
The confidence score is how sure the vision model is about the grade it assigned. High scores auto-accept and route straight to disposition, while low scores flag the item for a human operator to confirm or override, keeping questionable calls out of automated flow.
Does a human ever review the AI grade?
Yes. Borderline confidence scores and grade-tag conflicts are routed to a human operator, and every override the operator makes is logged against the return record. That log doubles as an audit trail and a training signal for the model.
What are detected tags and why do they matter?
Detected tags are the specific conditions the model finds, such as scuffs, missing packaging, or liquid damage. They explain why an item received its grade, make the grade auditable, and give reviewers and downstream buyers concrete detail rather than a bare letter.
Related reading: AI Condition Grading Explained: The ABCR System Behind Every Return
