Beyond reading documents: AI for commodity operations
See how AI turns commodity trade documents into usable records, checks contract terms and routes exceptions for review after extraction.
At some point in most evaluations, someone asks: “Isn’t this just OCR with a nicer interface?” It’s a fair question. Plenty of tools can read a document and pull out the fields, and the way they’re described can sound a lot like what we do.
The difference becomes clearer when you look at what happens next. Once you have the vessel name, quantity and payment terms, someone still has to work out which trade they belong to, whether they match what was agreed, and what needs to happen. That’s the part we focus on.
Scanning has made a real difference. Pulling a table out of a messy PDF or reading a scanned bill of lading saves people time. But even when every field is read correctly, there can still be a lot of work left.
Take a trader’s recap. Reading it tells you what the email says. Turning it into a usable trade record means understanding how your desk books trades, what’s already been agreed with that counterparty, and which rules apply. Some of that information simply isn’t in the email.
So when we assess the output, we ask more than “Did it read this correctly?” We ask: “Is this the record your team actually needs?”
Five things that need to happen after extraction
1. Turn the information into a trade your desk can use
An email says “100,000 MT, five shipments.” That gives you a quantity and a shipment count. Your system may need five separate legs, each with its own tolerance, laycan and pricing window.
If that’s how your desk works, the agent needs to create that structure. Otherwise, someone still has to build the trade by hand.
2. Help fill in the gaps
Recaps often leave things out. People who trade together regularly don’t repeat every term in every email. If the Incoterm or payment terms are missing, an agent can look at previous trades with that counterparty and suggest what belongs there. It should also show where the suggestion came from, so your team can review it.
That’s a familiar process for an experienced execution person: check the history, make a judgment, and confirm anything uncertain.
3. Link the document to the right record
A bill of lading needs to belong somewhere: a particular contract, leg and shipment. Getting that link right makes the information useful. Without it, you may have extracted every field perfectly, but someone still has to find the trade and attach the document.
It’s a fairly unglamorous step, but much of what follows depends on it.
4. Check it against what was agreed
Once the trade is properly recorded, you can compare incoming documents with it. Does the counterparty’s confirmation match your recap? Is the B/L weight within the contractual tolerance? Has the laycan changed?
The rules matter here. A 2% quantity difference might be acceptable on one contract and need attention on another. The system needs to apply your agreed tolerance and bring the exceptions to the right person.
5. Carry the work forward and use corrections
With the record in place, the system can draft the contract, create the shipment, draw down the allocation or flag an exception for review.
And when someone corrects a field, that correction should help with future documents. Capturing why it was wrong gives the system something useful to apply next time, instead of leaving the team to fix the same issue repeatedly.
What that looks like on one trade
Say you’re buying material across five shipments. Here’s how the process could run.
Step | What happens |
|---|---|
1. The trader forwards the recap | It’s the email they’ve already sent to the counterparty. They don’t need to reformat it or fill in another template. |
2. The trade record is built | On this desk, shipments are represented as legs, so five legs are created. CFR is suggested from the trading history and clearly marked for review. The purchase is allocated against the open sale and routed for credit approval. The trader reviews the record instead of entering it from scratch. |
3. The commercial contract is drafted | Terms come from the trade record. Clauses come from the firm’s library, with an explanation for each choice. Anything unresolved is flagged. |
4. The counterparty’s confirmation arrives | It’s compared with the record. There’s one difference in the laycan, so that’s what the team is asked to review. |
5. The documents for shipment three arrive | Eleven weeks later, the B/L and mill certificate are linked to the right shipment. The quantity is within tolerance, so it doesn’t need attention. One assay element is outside specification, so an exception is raised with the relevant penalty clause attached. |
6. The invoice is reconciled | In this example, it’s checked against the B/L weight. The record passed to accounting agrees with the supporting documents. |
The team can focus on approving the trade and resolving the laycan and assay issues, without having to work through every field that already matches.
Three questions worth asking in a demo
“AI-powered” doesn’t tell you much about how a product will work in practice. These questions should give you a better sense of it. We’d encourage you to ask us the same ones.
Question | What to ask to see |
|---|---|
“Can you show me how you handle a missing term?” | For example, what happens when the recap doesn’t include an Incoterm? Can the system suggest one from relevant history, explain the suggestion and leave it for review? |
“What happens when the B/L disagrees with the contract?” | Ask to see the comparison, the tolerance rule being applied, and who gets notified if the difference needs attention. |
“What happens after I correct something?” | Ask how the correction affects the next document, and what your team has to do to make that happen. Does the system retain the correction and its context? Does it need a support ticket or a separate update? It helps to know that before you start. |
Related post
Questions & answers
FAQ
How is AI for commodity operations different from OCR?
Optical character recognition turns text in a scan into machine-readable text. Document extraction identifies fields such as quantity or vessel name. The workflow described here also builds a usable trade record, connects documents to the correct shipment and checks the terms before routing work for review.
Can AI handle a trade across several shipments?
Yes, when the system has the desk’s booking rules. A recap for 100,000 MT across five shipments can be represented as five legs if that is how the desk records it. Each leg can then carry its own tolerance, laycan and pricing window.
What happens if a recap leaves out an Incoterm or payment terms?
An agent can suggest a term from relevant previous trades and show the source. The suggestion needs to stay distinct from what the current recap actually says, so the team can confirm it before treating it as agreed.
How does AI check a bill of lading against a contract?
The B/L first needs to be linked to the correct contract, leg and shipment. Its weight can then be compared with the agreed quantity and tolerance. The reviewer should be able to see both figures, the rule applied and why a difference was flagged.
Where do people stay involved in the workflow?
People review proposed terms, approve trades and credit, and resolve exceptions such as a changed laycan or an assay result outside specification. The record and supporting evidence should be ready when a decision reaches them. Each firm’s approval rules determine which actions require review.
What happens after an operator corrects a field?
The correction should retain its reason and scope. A correction to how a supplier document is read may help with the next document; a commercial term negotiated for one trade should apply only where it was agreed. Ask what is retained and whether a separate update is needed.
Discover what CommodityAI can do for you
Reclaim your time, scale your operations, and unlock powerful insights with domain-specific AI.