Where AI actually pays off in a recycling or scrap operation (and where it does not)

Every scrap and recycling operation gets pitched AI now. Some of the pitches are real. Many are a camera and a dashboard with a large price tag. The useful question is not “should we use AI” but “which specific task costs us money today, and would a model do it better than the current way.”

This is a practical way to sort the good bets from the poor ones.

Where does AI actually earn its place in a scrap yard?

Look for work with three traits: it is done many times a day, the inputs are messy, and a wrong answer is cheap to catch. Those are the jobs where modern language and vision models are strong.

Reading scale tickets, paperwork and emails

Many yards still key scale tickets twice, once at the scale and once in the back office. Inbound paperwork arrives as photos, PDFs and email attachments: mill certs, bills of lading, tickets from other yards, customer POs. A model can read these, pull out the weight, material, supplier and reference numbers, and fill the form a person would have typed.

The key design choice is to have a person confirm anything the model is unsure about. Done right, clerks stop typing and start reviewing, and errors drop because the review is faster than the original keying.

Matching and reconciling

Reconciliation is dull and expensive. Does the weight on the mill’s settlement match your outbound ticket? Does the invoice match the contract price? Models are good at fuzzy matching across documents that do not share a clean key, then flagging the ones that do not agree for a human to look at.

Flagging exceptions

Instead of asking a person to scan every transaction, let software surface the odd ones: a tare weight that differs from the truck’s history, a price outside the contract band, a supplier who suddenly delivers a different mix. Much of this is plain rules and statistics, not AI. Use whichever is simpler.

Drafting and summarizing

Quote requests, supplier follow-ups, daily summaries for the owner, notes from a call turned into a CRM entry. Low risk, real time saved, and easy to review.

Answering questions about your own data

A well-built assistant on top of clean operational data lets a manager ask “which suppliers lost us money on copper last quarter” without waiting for a report. This only works if the data underneath is sound.

Where is AI a poor fit?

Be skeptical when the pitch skips the problem it is meant to solve.

• Replacing the grader’s judgment on a hard load. Vision can help flag a contaminated load or confirm a material class. It does not replace an experienced person looking at a mixed load, especially for grade disputes with money attached. Treat it as a second opinion.
• Predicting commodity prices. Metals markets respond to events no model on your data can see. A tool that promises to time the market should make you suspicious. A tool that tracks your position against current prices and your contracts is another matter, and that one is plain software.
• Safety-critical decisions. Radioactive source detection, battery fire risk and similar hazards belong with certified detection equipment and procedures. AI can help log and trend events. It should not be the control.
• Fixing bad data. If material codes are inconsistent and the same supplier exists under five names, a model will produce confident nonsense. Clean the data first. That work is unglamorous and it pays for everything else.
• Physical sorting hardware you do not have the volume for. Optical and robotic sorting can make sense at scale in the right stream. At a small yard, the capital and maintenance can outweigh the labor saved. Run the numbers on your own tonnage before believing a vendor’s.

A quick way to compare opportunities

Opportunity	Typical value driver	Risk if wrong	Data needed	Usually worth starting?
Scale ticket and paperwork capture	Less double keying, fewer errors	Low (reviewed by a person)	Sample tickets and your forms	Yes
Settlement and invoice reconciliation	Recovered short pays, faster close	Low to medium	Contracts, tickets, settlements	Yes
Exception flagging	Catches leakage early	Low	Clean transaction history	Yes, often rules first
Vision-assisted grading support	Consistency on inbound	Medium	Labeled images from your streams	Maybe, as a pilot
Price forecasting	Speculative	High	Market data you cannot control	Rarely
Safety-critical detection	Compliance	High	Certified equipment	Not as an AI project

How do you decide what to try first?

Use a short filter.

1. Write down the current cost. Hours per week, errors per month, disputed dollars per quarter. If you cannot estimate it, you cannot judge a fix. Use your own numbers, not a vendor’s.
2. Check that the data exists and is reachable. Photos of tickets, a spreadsheet of past settlements, a PDF archive. If the inputs are locked in a system with no export, the first project is access, not AI.
3. Define “good enough” before building. What accuracy lets the clerk stop retyping? What rate of false flags is tolerable?
4. Keep a person in the loop at the start. Measure how often they change the model’s answer. When that rate is low and stable, loosen review on the easy cases.
5. Pilot on one stream or one location. Prove it on a real week of tickets before rolling it out.

Most of these are small projects, weeks rather than quarters, which is part of why they are good starting points.

What about the data underneath?

Every item above works better when tickets, contracts, inventory and sales live in systems that talk to each other. A lot of recycling businesses run on a scale system, an accounting package and a set of spreadsheets that connect through someone’s copy and paste. Putting the data in one place, with consistent material codes and supplier records, often delivers more than any model and is a prerequisite for the models that do help.

That is why we suggest looking at the data foundation and the AI opportunities in the same assessment, then ordering the work by payback.

What should you ask a vendor selling AI to a yard?

• What exact task does it do, and what does a person still do afterward?
• Can I see it run on my own tickets or my own images, not a demo set?
• How does it show uncertainty, and what happens to low-confidence results?
• Who owns the data and any models trained on it?
• What does it cost to run per month once the pilot ends?
• Can I get my data and configuration out if I leave?

Vague answers are an answer.