In April 1956, a converted tanker called the Ideal-X left Newark for Houston carrying fifty-eight trailer bodies on its specially fitted deck. Nothing about the cargo got simpler. What Malcolm McLean changed was the interface between the contents and everything that had to move them — not the contents. As Marc Levinson’s The Box (2006) tells it, conventional break-bulk shipping could leave vessels in port for days while cargo was loaded piece by piece, at enormous handling cost. Standardising the box made the handoff fast, cheap and mechanical between systems that didn’t need to understand each other’s contents.
FX is not shipping cargo, but its automation debate is repeating a mistake shipping solved seventy years ago. Paul Golden’s feature for this publication, “Structured FX Derivatives and the Challenges of Automation” (December 2025), reached a defensible conclusion: correlation baskets, barriers and target redemption structures resist full automation. Alexandre Gabovich of Makor Securities put it plainly — these trades “often require a level of human judgment, negotiation and documentation that current digital solutions cannot fully replicate.”
So why does one of the industry’s experienced infrastructure builders, in the same piece, locate the actual obstacle somewhere else entirely?
The workflow underneath every bank’s platform
Strip away the branding, and a structured FX trade at a major bank typically moves through a broadly recognisable sequence. A business-language hedging need becomes a candidate payoff — note, accumulator, barrier option — iterated on strike, tenor and notional; the exotics desk prices it against an internal vol surface using a model appropriate to the payoff; market and counterparty risks are assessed; legal and documentation requirements are addressed; the trade books and hedges; and the position becomes live — Greeks managed, barriers watched, correlation monitored — until it matures, knocks out or is restructured.
The precise workflow, controls and division of responsibility vary between institutions. But the underlying stages are not unique to any single bank. That is the shipping lane structured trades broadly travel, which is precisely where standardised interfaces can be inserted without needing to replicate what happens inside any individual bank’s hold.
The platform problem, not the complexity problem
Camus, Head of Hydra at Digital Vega, was blunt about where electronification has stalled: “if you have a product with Bank A you cannot passport it to bank B — they are not interoperable… it is not open architecture.” He credited the one condition that made much of the automation now possible — cloud infrastructure, and specifically Amazon Web Services: “even ten years ago it would have been impossible… the costs would not have stacked up.” Camus has identified an important bottleneck, and it is not simply intelligence. Much of the industry’s automation effort has historically been built around platforms — including single-dealer interfaces with their own connectivity, API contracts and onboarding. Interoperability between them remains limited.
A pricing capability embedded inside Bank A’s environment cannot automatically be picked up by Bank B’s crane. The opportunity is therefore to standardise the box at a layer where doing so does not require banks to expose their proprietary pricing models, inventory or client relationships: the interface through which pre-trade intelligence is requested and returned.
Why structuring itself cannot simply be automated away
Camus himself drew the line precisely: “correlation swaps will probably never be executed this way, but all the commonly traded products are going to be electronified increasingly quickly.” That boundary sits in a specific place, for three reasons.
First, structuring is partly a preference-elicitation problem, not simply an optimisation problem. A treasurer who says “protect my EUR receivables over the next eighteen months” has specified a business problem, not a derivative. Which barrier level is acceptable, whether a structured option or a forward strip is appropriate, and how hedge-accounting considerations under IFRS 9 or ASC 815 affect the choice depend on objectives, constraints and risk appetite that may only become clear as the client and structurer work through the trade-offs. Before those preferences are defined, there is no single objective function to hand to a machine.
Second, documentation and model governance introduce constraints that are not computational problems. An ambiguity in drafting is a legal liability, not a rounding error. Model-risk regimes reinforce the need for formal governance, clearly assigned responsibilities, independent challenge and effective oversight. In the United States, the interagency Model Risk Management guidance issued in April 2026 under Federal Reserve SR 26-2 replaced the earlier SR 11-7 guidance; in the UK, the PRA’s SS1/23 places responsibility for the overall model-risk-management framework on boards and senior management, with accountability allocated through the firm’s governance structure.
The history of structured FX supplies a useful warning. In October 2008, CITIC Pacific disclosed HK$807.7 million of realised losses associated with leveraged foreign-exchange contracts and a further HK$14.7 billion mark-to-market loss on its outstanding leveraged FX positions. The contracts included target redemption forwards, with limited profit potential but no equivalent knock-out feature for losses. The lesson is not that the mathematics was unavailable. It is that sophisticated payoff mathematics is no substitute for limits, governance, monitoring and effective escalation.
Third, the economics matter. A bespoke correlation structure may trade only occasionally at an individual institution, offering a very different return on automation investment from a vanilla flow business processing thousands of transactions. Camus is therefore describing a boundary that is economic and institutional as much as technical.
What is actually changing — not in theory, in the desk’s daily workflow
None of that means the rest of the process has to remain manual. Outside the elements that require genuinely bespoke structuring, negotiation and legal judgement, there is substantial scope for AI and machine learning to change the workflow.
Vol-surface construction and interpolation provide one example. Sparse tenor and strike combinations are an unavoidable problem in less liquid parts of the market. Parametric approaches such as SABR and SVI remain important, but machine-learning techniques can also be used as flexible approximation tools for fitting and interpolating complex surfaces. They do not remove extrapolation risk — no model does — but they add another set of tools for dealing with sparse and irregular data.
Natural-language systems can shorten the translation step between a business-language hedging requirement and a structured RFQ template. That is a real automation opportunity: reducing the mechanical burden of expressing the request, without pretending that the system has replaced the judgement needed to decide whether the resulting structure is appropriate.
Lifecycle monitoring is another. Systems can continuously watch barrier proximity, exposure and market conditions and escalate alerts before a trigger is reached. That does not make the decision for the risk manager. It changes how quickly the risk manager sees the information.
The most consequential shift may be in risk calculation. Large-scale CVA, XVA and derivatives-risk workloads have traditionally relied heavily on batch calculations because repeated simulation across complex books is computationally expensive. Brian Huge and Antoine Savine, then at Danske Bank’s Superfly Analytics group, introduced “Differential Machine Learning” in 2020: training neural networks not only on simulated prices but also on their derivatives with respect to model inputs. The resulting approximators are designed to return prices and risk sensitivities extremely quickly and can be applied to computationally intensive problems including XVA, FRTB and SIMM-MVA.
That does not mean every path-dependent book can simply be converted into a real-time calculation. For discontinuous payoffs such as digital and barrier options, later research has shown that naïve pathwise sensitivities can be biased and need to be handled carefully. But the direction is important: computationally expensive risk calculations can increasingly be approximated quickly enough to move some analysis from periodic batch processing toward real-time or near-real-time decision support.
Order-flow toxicity and regime classification belong here too — the layer I know best, because I build it. A Hidden Markov Model, in the tradition of the regime-switching work James Hamilton formalised in Econometrica in 1989, can infer an unobserved market state — calm, transitional, stressed — from the statistical behaviour of observable market variables.
David Easley, Marcos López de Prado and Maureen O’Hara’s 2012 Review of Financial Studies paper introduced VPIN as a way of estimating flow toxicity from volume imbalance and trade intensity. Flow is “toxic” in this framework when liquidity providers are at risk of being adversely selected. VPIN therefore provides a measurable market-state signal; it does not, by itself, establish why an individual trade moved the market or prove that informed flow caused a subsequent price move.
What ties this together is calibrated uncertainty. Anastasios Angelopoulos and Stephen Bates’ A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification describes a framework for constructing prediction sets or intervals with explicit statistical coverage guarantees under the relevant assumptions. Combined with out-of-distribution detection or abstention logic, that can support a useful engineering discipline: widening uncertainty — or declining to provide an estimate — when the system is operating outside the region in which its answer can be defended.
A surface model that extrapolates confidently into thin data may be more dangerous than a model that makes its uncertainty explicit, and the same principle applies to a lifecycle monitor or an XVA approximation.
An open question I was testing — and the answer
Highly bespoke and illiquid structured FX derivatives have received far less systematic microstructure analysis than spot FX or more standardised parts of the FX options market. That is understandable: transactions in a particular structure and tenor may be separated by weeks or months, leaving very little data on which to train a model directly.
The underlying spot market, however, trades continuously, and flow-toxicity measures can be calculated from much denser data. That raises an interesting question. Does spot-market toxicity contain incremental information about the market impact of an illiquid derivatives execution — or is apparent post-trade spread widening driven mainly by other factors such as dealer inventory, volatility-surface stress, trade size, market regime or the hedging flow generated by the derivative itself?
That is a causal question, not merely a correlational one. One way to investigate it is Double Machine Learning: using machine-learning methods to estimate nuisance relationships while constructing an orthogonalised estimate of the parameter of interest. In this case I used it to partial regime, size and session timing out of both the toxicity signal and the derivative-impact measure, then asked whether a relationship remained in the residuals.
The first test did not support the causal story I had hoped to find. On synthetic data built specifically to isolate the mechanism, the raw information coefficient between toxicity and impact was +0.23. Once regime, size and session timing were partialled out through the DML residualisation, it fell to +0.09 and lost statistical significance — approximately a 61% reduction.
That result needs to be interpreted for what it is. A synthetic experiment cannot establish that no causal relationship exists in the real FX market; its conclusion is conditional on the data-generating process used to construct the test. What it showed was that the causal claim I was considering was not robust even in a controlled framework designed to separate the proposed mechanism from obvious confounders.
I then put the same question directly to a senior FX structuring practitioner at a Tier 1 bank, independently of the model, in a private conversation rather than a published source, so I am not naming them here. Their answer — arrived at from desk experience rather than from seeing my analysis — pointed in the same direction. They argued that an apparent relationship between spot toxicity and options impact could instead arise because a single large participant is active in both markets simultaneously, rather than because spot conditions are themselves causally driving derivative impact.
Those were two very different pieces of evidence — neither sufficient on its own to establish a universal result — but both gave me the same reason to step away from the stronger causal claim. I shelved the product built around that claim the same week. The interface — a pre-trade layer that returns a bounded estimate or an explicit refusal to estimate — still works as engineering. What it would have been estimating turned out to be less defensible than the initial correlation suggested, and building further on that premise would have meant selling a story I no longer believed.
The abstention discipline this piece argues for is not just for the model. It applies to the person deciding whether to keep building.
A note on what this argument is not
Alan Dweck, COO buy-side solutions at SGX FX, offered the cleanest description of what remains irreducibly human — “a good way of looking at a trader’s job is that they manage house and market risk.” Nothing described above replaces that responsibility. It informs the person who carries it, at more points in the workflow than it did five years ago. Technology can increasingly assist with elements of product discovery, structuring, pricing, monitoring and risk analysis, but it does not remove the judgement involved in genuinely bespoke structuring, legal negotiation or the management of house risk.
A system of this kind is only as trustworthy as its abstention discipline, and only as accurate as the data against which it has been calibrated. Extending such systems toward real dealer execution history remains an important part of the work in implementations including my own. A system that starts guessing on thin data rather than exposing uncertainty or declining to answer becomes exactly the quiet risk this approach exists to avoid.
Conclusion: The box, not the trade
A correlation swap will not suddenly become simple to structure because the interface around it improves. Its complexity is mathematical, economic, legal and institutional. Standardising the box never made the cargo simpler — it made the handoff cheap enough that the complexity of the cargo stopped being an excuse for friction everywhere it didn’t need to be.
That is the more useful way to think about what AI can standardise in structured FX.
What is changing across the automatable parts of the workflow is not the underlying trade. It is the ability of someone, cheaply and from systems they already use, to ask what condition the market is in, what the risk might be, how uncertain that estimate is and whether the system has enough evidence to answer at all.
Some of those questions once required a platform login, a spreadsheet or a phone call. In the segment the industry itself concedes is tractable, an increasing number can now be exposed through a standardised interface.
The box is getting simpler.
The cargo does not have to.
