AI for Scrap Yards: Log Intake, Weights and Documents With Review

AI for Scrap Yards: Log Intake, Weights and Documents With Review

How a scrap yard can use AI to log seller intake, declared weights and documents, with mandatory human review before any payment is approved.

•

Want to apply this in your company?

Book a call with Juan →

An AI agent can organize material intake at a scrap yard. It records the seller, the declared material type, the reported weight, the photos and the documents presented, then builds a queue for review. It should not set prices, approve payments or validate documents on its own. The scale is the source of truth for weight, a responsible person checks the intake, and every step is logged for audit.

Key takeaways

  • The agent organizes intake, but the scale and the responsible person remain the source of truth for weight and identity.
  • Seller documents are personal data: collect the minimum needed and define who can see it.
  • Any gap between reported weight and scale weight goes to human review, never to automatic correction.
  • Start with one material type and one shift, with logs and an agreed success metric before the pilot.
  • Payments and tax entries stay outside the agent until the operation proves that its records are reliable.

Why material intake is the most fragile point

In many scrap yards, intake happens at the counter, in the yard or by phone, with notes kept in a notebook, a spreadsheet or WhatsApp messages. The problem is rarely a lack of technology. The problem is traceability: when a batch does not match the payment, nobody can reconstruct what was said, weighed and accepted at that moment.

AI works well where information is scattered and standardization is low. Talking to the seller, asking for a photo of the material, asking about the load type and organizing the record are repetitive tasks that a well designed flow performs consistently. Deciding whether material is clean enough, or whether a document is authentic, requires a trained eye and human accountability, and that should not be delegated.

What the agent records and what stays with people

The most useful split separates physical and documentary facts, which belong to a person or a piece of equipment, from information handling, which the agent can do.

  • Records: date and time, the identification the seller provides, the declared material type, the reported weight, photos of the batch and notes the seller makes.
  • Does not decide: official weight, which comes from the scale; price, which follows the company's table and rules; payment approval; validity or authenticity of a document.
  • Escalates: weight discrepancies, unreadable documents, material outside the accepted list and any request for an exception.

If the agent answers confidently about price or weight, the instructions are wrong. It should say that the value depends on the review and log the question for the team.

Step by step: setting up intake with review

Implementation works best in a short sequence, with each step validated before the next begins.

  • Map the current flow from the moment the seller arrives until payment: who attends, where the data is written down and who checks it.
  • List the accepted material types and their rules, written by the operations lead, not by the agent.
  • Define the minimum document list and prohibit unnecessary collection, such as photos of third party documents or data that will not be used.
  • Write the agent's instructions with explicit limits: what it answers, what it never states about price or weight, and when it hands off to a person.
  • Test with fictitious entries before serving real sellers, as described in how to test AI agents with mocks and a safe sandbox.
  • Run a pilot shift with every entry logged, and compare the agent's record with the team's review.
  • Agree on the success metric before starting, for example the share of complete records that need no rework, and track it weekly.
  • At the end of the pilot, decide: expand, adjust the rules or stop.

Seller documents: what to collect and how to handle them

Identity documents, proof of origin and contact details are personal data. Brazil's General Data Protection Law (LGPD) requires a clear purpose, necessity and security in handling, and the National Data Protection Authority (ANPD) is the reference body for guidance. In practice, collect only what the operation actually needs, state the purpose at the moment of collection, restrict who can access images and decide with your accountant or lawyer how long these records must be kept. Also confirm the specific obligations of your type of business, since they vary by activity and region.

There is also a technical risk. A photo of a document or of material can contain text that tries to instruct the agent, for example asking it to ignore the rules and approve the entry. Treat all content sent by a seller as data to be read, never as a command, and keep approval in a separate step with its own permissions. The OWASP guidance for applications built on language models describes this kind of attack, called prompt injection.

Illustrative scenario: an entry at the end of the day

Illustrative scenario: a seller arrives late in the afternoon with mixed material. Through the messaging channel, the agent asks for the declared material type, a photo of the batch and the seller's document. It records the data and marks the entry as pending weighing. On the scale, the weight differs from what the seller reported. The agent changes nothing. It sends an alert to the responsible person with the record, the photos and the scale weight. The person decides and records the decision and the reason.

The next morning, the log shows which entries stayed pending and why. If most discrepancies come from the same material type, the team adjusts the instruction and the review rule for that case, not for every material at once.

How Juan Carlo's method applies to this case

The work follows four steps: map the current process and its baseline, limit data and actions, pilot a real flow with logs and an agreed metric, then hand over to the team, deciding whether the operation expands, adjusts or stops. At a scrap yard, the most important limit is that the agent has no access to payments or to changes in weight. If the operation also needs to compare scrap quotes or request approval for purchases, the logic is similar, as shown in the AI agent for procurement that compares supplier quotes with human approval.

Limits and when not to automate

Do not use the agent to negotiate prices, decide material exceptions, interpret tax requirements or answer about a seller's legal situations. These cases should go straight to a person. If a decision affects the seller, such as refusing an entry or holding a payment, review the topic of human review of automated decisions covered in LGPD automated decision-making and the right to review.

NIST, in its AI risk management framework, treats human oversight, documentation and continuous monitoring as parts of risk management, not as details. Review logs regularly, especially the entries sent to review, because they show where rules are loose or confusing.

Frequently asked questions

Can AI weigh material or approve a purchase on its own?

No. The official weight comes from the scale, and purchase approval is a decision for the person responsible for the operation. The agent can record the reported weight, the photo and the time, and it can escalate any discrepancy. This separation should be written into the instructions and reflected in the log.

Which seller documents should be collected?

Collect only what the operation requires by its own obligations or internal controls, always with the purpose explained to the seller. Avoid storing images without need, restrict access and set a retention period with your accountant or lawyer. The exact list depends on the type of operation and the region.

How do I start without taking on too much risk?

Start with one material type and one shift, in a pilot with complete logs, and test first with fictitious entries. Define the success metric before you begin and compare the agent's records with human review. Expand the scope only after you have looked at real data.

Is AI worth it for a small operation?

It depends on the volume and on how intake is recorded today. If the notebook is already reliable and volume is low, the gain may be small. If many batches have lost data or reviews are hard to reconstruct, organizing intake is usually the first problem to solve, and that should be judged against the current process in plain view.

Sources