Over the past few years, AI in trade finance has moved from a buzzword in trade finance to something banks are actively deploying. The conversation has shifted from ‘should we explore this?’ to ‘why isn’t our implementation delivering what we expected?’Â
That second question is the one I spend most of my time on. And it’s one I’ve looked at from both sides after more than a decade leading bank operations and now working with banks through full AI implementation cycles.Â
Across both perspectives, I’ve concluded that might seem obvious in hindsight: technology is rarely the limiting factor. Process is.Â
In practice, the implementations that succeed all align across three layers: process clarity, data availability, and system configuration.Â
If any one of those is weak, the outcome is the same: partial adoption, limited scale, and a system that never quite delivers on its promise. The banks that get this right don’t treat AI as a standalone solution; they treat it as something that has to be embedded across all three layers to work.Â
This piece is about what that actually means in practice and why getting it right is what separates implementations that compound in value over time from those that plateau at partial adoption.Â
The Scale of the Problem Being Solved
Let’s start with why this matters at the industry level. A 2018 ICC Banking Commission global survey estimated that four billion pages of documents circulate annually in documentary trade between the physical and financial supply chains of cross-border commerce.1 That number hasn’t fundamentally changed. Trade finance remains one of the most paper-intensive, manually processed workflows in global banking.Â
The consequences are well-documented: processing costs are high, turnaround times are measured in days rather than hours, and compliance checks, including sanctions screening, TBML red flag analysis, document examination against UCP 600 and ISBP, are fragmented across teams with limited shared visibility.Â
Compounding the problem, trade-based money laundering has become one of the most significant financial crime vulnerabilities in global trade. The Financial Action Task Force (FATF) and Egmont Group have documented how criminals exploit the complexity of international trade, including multiple parties, multiple jurisdictions, and the sheer volume of documentation, to disguise illicit funds and move them across borders.2 The scale and opacity of trade flows make detection genuinely difficult, even for well-resourced compliance teams.Â
In large financial institutions, compliance investigators may review thousands of alerts per month, with only a fraction resulting in confirmed matches. The majority require manual clearance, which is a significant drain on the very people whose judgment is most needed on genuine risk.Â
AI addresses all of these problems. The technology to do it exists and is mature. The challenge is not capability. The challenge is implementation. More specifically, it’s not whether AI works, it’s whether organizations are structured to actually use it.Â
The System-First Trap
When a bank purchases an AI-powered trade finance solution, the instinct is to focus on the system, specifically how it integrates, what it can process, and which screens the operations team will use. That’s understandable. Systems are tangible. You can demo them, test them, and go live on them.Â
But trade finance is not primarily a technology problem. It’s a process problem that technology can solve, but only if the underlying process is first mapped and understood.Â
Trade finance operations are document-heavy, multi-party, and governed by a dense web of rules: UCP 600, ISBP, URDG, local regulatory frameworks, and bank-specific risk policies layered on top. When you drop AI into that environment without doing the process work first, one of two things typically happens: the system gets underused, or it gets worked around.Â
Neither outcome delivers the value the bank was promised. And both lead to the same post-implementation frustration: adoption curves that plateau within months of go-live, with usage concentrated in a fraction of the workflow the system was designed to handle.Â
What a Process-First Approach Actually Looks Like
When I work with a new customer, the first question I ask is not ‘what do you want the system to do?’ It’s ‘walk me through how you handle a Letter of Credit today, from the moment documents arrive to the moment payment goes out.’Â
That conversation almost always surfaces the real friction. The pain points are predictable, even if they vary in detail across institutions:
- Manual document sorting that consumes hours before any substantive review even begins. Â
- Compliance checks running in parallel silos: trade operations, sanctions, and risk creating duplication and blind spots at the same time Â
- TBML assessments happening retrospectively, instead of being embedded at the point where decisions are actually made Â
- Sanctions screening generating high false-positive volumes, pulling experienced investigators into low-value work instead of real risk analysis Â
- No systematic vessel tracking or fair price validation, leaving shipment-level risk detection to inconsistent manual spot-checks Â
- None of these are edge cases. They are structural, and they show up in almost every implementation.Â
Once you see the process clearly, you can map AI to the specific friction points. The technology stops being a general-purpose tool and becomes a precise intervention: in the right place, at the right step, with the right configuration for that institution’s risk appetite and governance framework. This is where those three layers—process, data, and configuration—start to come together in a way that actually works in production.Â
This has a practical implication for implementation timelines. The process-mapping phase feels slow at the start. It requires cross-functional engagement, involving operations, compliance, risk, and IT in the same room, working through workflows that have often never been formally documented. It takes longer than a purely technical integration scoping exercise. Banks sometimes push back on this phase, particularly when there’s pressure to show go-live dates to senior stakeholders.Â
The banks that do this work are the ones whose adoption compounds. The banks that skip it are the ones asking difficult questions six months post-launch.Â
The Evidence Is in the Volume
I’m sharing this perspective because we’re seeing the results of this approach in our own numbers and they’re significant enough to be worth discussing openly.Â
Transaction volumes across our in-production customers grew 51% year-over-year in Q1 2026 versus Q1 2025. In Q1 2026 alone, we processed approximately 49% of our full-year 2025 throughput: with key customers still in active expansion phases.Â
It’s important to understand what those numbers represent. That growth is not driven by a surge in new customer acquisitions. It’s the result of existing customers scaling their usage, because the system is embedded in their actual process, not sitting alongside it as an optional tool that ops teams’ route around when they’re under pressure.Â
A system that’s only adopted at 20% of its potential is not a successful implementation, regardless of what the contract says or how strong the demo looked. A system that drives 51% transaction volume growth from in-production customers is one that banks have genuinely integrated into how they operate. There’s a meaningful difference between those two outcomes, and it lives almost entirely in the implementation approach, not in the technology.Â
The compounding effect matters too. Banks that reach deep integration don’t just maintain volume; they expand it as internal confidence grows, as more document types get added to automated workflows, and as compliance teams develop trust in the system’s outputs. That trust takes time to build, and it’s built through consistent performance on real transactions, not through proof-of-concept environments.Â
Building Successful AI Implementations: What Both Sides Need to Do
A process-first implementation requires honesty on both sides of the vendor-client relationship.Â
From the bank’s side, it means bringing operations, compliance, and risk into the implementation from the beginning, not just IT and procurement. The people who live inside the process every day are the ones who know where it actually breaks down: which document types cause the most delays, which sanction screening rules generate the most noise, where the TBML check is being done manually because the system doesn’t have the data it needs at the right point in the workflow.Â
Their input is not a nice-to-have. It’s the foundation of a configuration that will work at scale. Without it, you’re configuring a system against an imagined process rather than the real one. That gap is where adoption goes to die.Â
From the vendor’s side, it means resisting the temptation to optimise for fast go-live dates and early milestone metrics. A bank that goes live in six weeks with shallow integration is not a success story. A bank that takes twelve weeks, does the process work properly, and then shows consistent volume growth for the following eighteen months; that is the outcome worth building toward.Â
It also means being honest about what AI in trade finance can and can’t do. Vessel tracking, document examination, noun extraction for sanctions screening, fair price validation against market data — these are tractable, well-defined problems, and the technology handles them reliably. The AI is not the uncertain variable. The uncertain variable is always whether the bank’s internal process is configured to actually use the outputs.Â
That’s the vendor’s responsibility to solve alongside the customer, not to hand off as a post-go-live ‘change management’ problem.Â
Because in practice, implementation is the product.Â
Let’s TalkÂ
If you’re evaluating AI for trade finance, or you’re already mid-implementation and the adoption curve isn’t moving the way you expected, I’d genuinely like to hear what you’re seeing.Â
Not to run a sales process, but to understand where the friction actually is. In many cases, the diagnosis itself is where the real value sits. Â
Feel free to reach out to me directly. I’m always open to a conversation with teams working seriously on these problems.
ReferencesÂ
1.  International Chamber of Commerce (ICC) Banking Commission. Global Survey on Trade Finance, 2018. ICC estimated approximately four billion pages of trade documents circulate annually across documentary trade supply chains. iccwbo.orgÂ
2.  Financial Action Task Force (FATF) & Egmont Group. Trade-Based Money Laundering: Trends and Developments. December 2020. fatf-gafi.org/en/publications/Methodsandtrends/Trade-based-money-laundering-trends-and-developments.htmlÂ
