Lessons in innovation: AI’s role in solving decades-old problems in Agri Trade

For decades, the agri-commodities industry has struggled with problems in trade execution that remained unsolved. Covantis was founded with the mission to digitize and bring efficiency to agri‑trade execution, so it was natural for us to take on this challenge. We felt it was our responsibility to push past technology limitations and try to deliver what the industry had long needed.

One of the biggest of those long-standing problems was the time-consuming manual checking of trade documents. For trade execution coordinators, reviewing contracts, Letters of Credit, and certificates is repetitive, stressful, and full of risk. The consequences are serious: errors in shipping documents under a Letter of Credit or Cash Against Documents (CAD) can delay payments by 7–14 days, raise working capital costs, and even trigger bank penalty fees.

Errors in draft documents slow approvals with buyers, often adding 2–3 days of back‑and‑forth revisions and disrupting cash flow. Manual preparation and verification of document sets can take hours per vessel, increasing costs and the chance of human error. This was a big, costly problem worth solving.

That’s why, even before the mass adoption of AI, we began working with machine learning (ML) to see if it could help us tackle this critical problem.

The early struggles

At first, we tried solving this problem with machine learning, but ML had serious limits:

• It needed massive amounts of training data, because ML systems learn by seeing hundreds if not thousands of examples before they can generalize. That meant every possible format, clause, or variation had to be manually collected, cleaned, and labeled by people. This is labor intensive and expensive.

• It struggled with unusual formats or rare discrepancies, because ML models tend to recognize only patterns they have seen during training. When a document looked slightly different (like a new layout, a different template, or a rarely used clause) the model often failed. Even small differences, such as values hidden under an ink stamp or presented in an uncommon format, could cause incorrect or missed results.

• Accuracy wasn’t high enough. Even 95–97% accuracy isn’t enough for users who require near-100%.

When we tested with early adopters, their feedback was blunt: “If I can’t fully trust the system, why not just do it myself?”, “Can an application really spot errors and discrepancies better than me?”, “What if I trust it and my bank rejects the documents as discrepant?”

And they were right. Imagine if 1 out of 100 transactions on your bank statement was wrong. You’d lose trust instantly.

We almost killed the project multiple times, pivoted again and again, and even tried what is known as a “human-in-the-loop” approach (HITL) – having people step in to manually check or correct the machine’s output – to patch ML’s weaknesses. Checking the document through HITL added time to the overall processing, because even if just one document in a shipment set required human review, the user still had to wait until the entire set was finished before moving forward.

The truth was simple: the technology wasn’t ready.

The breakthrough

Late in 2024, three things changed the game:

1. A new approach – We redesigned how we tackled the problem, rethinking the product’s value proposition and taking a different approach to solving the user’s problem.
2. Stronger AI – Today’s AI models are far more advanced. They generalize better and apply reasoning, which makes them capable of handling complex document sets, classifying and splitting documents into the right categories, and finding discrepancies with much higher accuracy.
3. Team creativity – Our product and engineering teams combined their knowledge of user needs with technical expertise to overcome AI’s weaknesses (accuracy issues, hallucinations, performance bottlenecks) and build a production-grade solution.

QuickDocs is born

That’s how QuickDocs came to life: our first production AI application. It’s now live and used daily by clients. It saves time, reduces errors, and helps them focus on the issues that matter most. On average, users report checking documents up to four times faster, experiencing 75% fewer errors in the document presentation process, and reducing working capital costs by eliminating delays in payments caused by errors in documents presented to the bank.

What are the benefits of QuickDocs?

Highly accurate – flags discrepancies down to a dot or comma.
Fast and smart – no manual entry of document instructions/check requirements. Splits large batch files into individual documents promptly, extracts data, and shows all checks in one dashboard.
Rule-aware – applies logic-based checks against contract, commodity, and country-specific terms.
Usability built for document management – structured interface, easy navigation, and clear dashboards instead of long, cluttered AI outputs.
Scalable – designed for agri-commodities but adaptable to any industry that needs document checking.

Why not use public AI tools or build in-house?

This is one of the most common questions we hear. Generic AI platforms and in-house tools promise speed and low cost, but they aren’t built for the realities of commodities trade. They miss details, struggle with large document sets and lack persistence or traceability. QuickDocs is different.

Here are just some of the reasons why QuickDocs outperforms generic AI platforms and in-house tools:

Accuracy you can trust – QuickDocs is trained on real trade documents (contracts, LCs, BLs, certificates), not internet text. It can flag discrepancies down to a dot or comma – critical in LC transactions.
Best of AI, not just one provider – Instead of locking you into a single model, QuickDocs routes tasks across multiple AI models, using each where it performs best.
Reliable on large sets – Public AI tools struggle with long or complex document sets, skipping checks or hallucinating.
Tailored usability – Designed for document checks, not chat replies. Coordinators get structured dashboards, clear navigation, and easy audit trails rather than long, cluttered text outputs.
Persistence and resolution – QuickDocs remembers your decisions across voyages. If you accept a recurring discrepancy once, the system applies it automatically in future, reducing repetitive work.
Enterprise-grade security – All data stays within Covantis’ secure cloud environment, never shared externally or used to train public models.

In short, purpose-built for commodities documentation, QuickDocs combines advanced AI, proprietary algorithms and Covantis’ scale in global agri-trade. It delivers accuracy, reliability and workflows that keep trade execution moving without the firefighting.

Lessons learned in AI innovation

From this journey, we learnt some lessons that make the difference between a proof of concept and a production grade solution. With QuickDocs, they meant the difference between repeated pivots and a product that clients now use daily:

Start with a real business problem – Anchoring QuickDocs in the painful, high ‑value challenge of document checking gave it purpose and ensured demand.
Continuous iteration with feedback – Each release of QuickDocs was refined through cycles of testing, user input, and updates.
Time your technology bets wisely – Our early ML experiments proved too costly and inaccurate, but we kept iterating with confidence that AI would overcome these limitations, and when stronger models arrived, we were ready to deliver real value.
Cross functional ‑ collaboration – Product, engineering, and commercial teams worked side by side with users, aligning technology choices with actual business value.
Integration into workflows – Adoption grew because QuickDocs fit seamlessly with existing Covantis tools like Voyages, avoiding duplicate data entry.
Enterprise‑grade security and compliance – Because clients share sensitive trade documents, keeping all data in Covantis’ secure cloud was non‑negotiable.

What’s next?

We’re proud of what QuickDocs has become, but we’re not finished. Our next challenge is making the app even more interactive and user-friendly. Trade execution coordinators handle documents worth millions of dollars; they need both trust and ease of use. Any false error or missed check could undermine that trust, and we take that seriously.

We’ll keep iterating, experimenting with the latest AI, and improving QuickDocs so it stays at the cutting edge. And we’ll continue looking for new use cases where AI can solve real business problems.

We don’t see ourselves as “an AI company.” We’re a solutions company using AI and other technologies to tackle real challenges in commodities world. QuickDocs is just the beginning.

Article by Petya Sechanova, CEO Covantis