A refurbished device is judged twice: once by the test bench and once by the buyer. AI can help make those judgements consistent, but only if it is fed the right evidence.
The most common mistake is asking a model to grade a device from a photo alone.
What AI can see
A model can inspect images for cracks, scratches, dents and discoloration. It can compare a device against a reference set and flag marks that a tired technician might miss.
It can also read structured test results: battery health, display response, camera output, charging behaviour and sensor readings. That is often more reliable than image analysis because the data is already measured.
The useful system combines both. Test results establish function; images describe appearance; the grade rule combines them.
Functional data vs cosmetic data
Functional failures are binary in most cases. A camera that does not capture, a port that does not enumerate or a battery below the cutoff is a known finding.
Cosmetic condition is a spectrum. A scratch has a location, length, depth and visibility. Two technicians can describe the same mark differently, which is why photo evidence and a device map matter.
AI can normalize the cosmetic description, but the business still has to define what counts as grade A, B or C.
Grade consistency
The value of AI grading is not a smarter letter. It is the same letter for the same evidence, regardless of who is on shift.
That requires a rule set with a version, a fixed input order and an exception queue for boundary cases. The model should produce findings; the rule engine should produce the grade.
When a grade is disputed, the report should show the test readings, the cosmetic evidence and the rule version that was applied. A model score without that trail is not defensible.
False positives
Image models produce false positives: a reflection read as a scratch, a shadow read as a dent, a screen protector read as damage. The cost is a device downgraded or sent for unnecessary repair.
Use a confidence threshold and route low-confidence images to a person. Keep a sample of reviewed images to measure the false-positive rate over time.
Human review
A person should approve any grade that changes the price, the channel or the warranty. The reviewer needs the evidence in one screen, not a folder of photos and a separate test report.
Review is also how the system learns. Every override is a data point about a rule that is too strict, too loose or poorly described.
A sensible pilot
Pick one product category and 100 devices. Grade them with the current process, then with the assisted process. Compare the grade distribution, the disputes and the time per device.
If the assisted process does not reduce disputes or time, do not scale it. The goal is a better decision, not a model in the workflow.
Building the reference set
An image model needs a reference set that reflects your actual inventory. Collect at least 100 devices per category, with photos taken under the same lighting and angles the bench will use.
Label the defects, not just the grade. A grade A device with one faint scratch and a grade B device with three deep scratches are different inputs; the model needs the defect labels to learn the difference.
Include the awkward cases: curved edges, screen protectors, reflections, cases, stickers and dust. A clean studio set will not prepare the model for a used-device bench.
Version the reference set. When the camera, lighting or device mix changes, the old labels may no longer describe the new images.
A useful rule is to keep the model out of the final price. Let it produce findings and a suggested grade, then let the rule engine and the reviewer set the commercial outcome. That separation makes it easier to change the model without changing the pricing policy, and easier to explain a grade when a buyer asks.
FAQ
Can AI grade a phone from photos alone?
It can describe cosmetic condition, but it cannot confirm function. A grade should combine hardware tests and cosmetic evidence.
Will AI replace technicians?
No. It reduces repetitive inspection and makes evidence easier to review. Technicians still handle exceptions, repairs and final approval.
How do we start?
Use one category and a small batch. Measure grade consistency, dispute rate and time per device before expanding.
For the product side, see device grading software, mobile QC software and the refurbishment workflow.
