Liquid Supply Chains

A container of goods crossing the Pacific is, economically, a bundle of near-certain future cash flows. The goods exist, their movement is tracked in real time, the buyer is often an investment-grade retailer, and the historical loss rate on financing such flows is a fraction of one percent. Yet for most of the firms that ship them, those cash flows are frozen: the working capital embedded in goods in transit and invoices awaiting payment cannot be borrowed against, sold, or pledged at reasonable cost. A typical small exporter pays its supplier at or before shipment, waits roughly 45 days for the goods to move from factory gate to destination port on the main east-west lanes, delivers, and is then paid 60 to 90 days after delivery on standard open-account terms. It is out of pocket for roughly 100 to 135 days or more on every order, and its growth is capped by cash rather than by demand.²

The central claim of this paper can be stated in one sentence: if verification costs fall while error rates do not rise, the minimum viable ticket size of trade finance falls, and the trade finance gap becomes experimentally testable as a unit-cost problem rather than a risk problem. Everything that follows unpacks that sentence: the evidence that motivates it, the mechanism that would carry it, the failure modes that could defeat it, and the measurements that would decide it.

We use the term liquid supply chain to describe the opposite state: one in which every eligible, verifiable shipment and invoice can be financed at the moment it is created, collateralized through enforceable digital documents, and distributed to capital markets in standardized form, much as a mortgage or a credit card receivable can be today. This paper asks what would happen to firms, to trade, and to the financial system if supply chains became liquid, and why the question has only recently become more than hypothetical.

The argument proceeds as follows. Section 2 documents a puzzle: an asset class with exceptionally low realized losses that capital nonetheless refuses to fund at scale. We state the puzzle together with its principal rival explanation, adverse selection, and explain why the two cannot be separated on existing data. Section 3 identifies the three converging shifts, technological, legal and informational, that change the underlying cost structure. Section 4 sets out the economic mechanism: the conversion of operational risk into priceable credit risk, and the analogy, with its important limits, to the standardization episodes that liquefied mortgages and containerized freight. Section 5 traces the consequences through firm, trade and capital-market channels, and Section 6 confronts the risks, including the one recent, directly relevant failure: Greensill. Throughout, we treat the claim that AI changes trade finance as a hypothesis to be tested, and we specify the measurements that would falsify it. Section 7 closes with the research agenda and an explicit invitation to researchers to design those tests before the data exist.

The Asian Development Bank estimates the global trade finance gap, defined as demand for trade financing that financial institutions decline to meet, at US$2.5 trillion per year, unchanged from 2023 and up from about US$1.5 trillion in 2015.³ The gap is approximately 10 percent of global merchandise trade,⁴ and it is concentrated where it does the most damage: 41 percent of trade-finance applications from small and medium-sized enterprises are rejected,⁵even though SMEs constitute roughly 90 percent of the world’s businesses and more than half of global employment.⁶The most recent survey reports, for the first time, SME rejection rates close to those of large corporates, a convergence the ADB cautions is provisional; even taken at face value, it does not dissolve the puzzle, because the gap’s absolute incidence remains concentrated among SMEs and emerging-market firms, and because the argument of this paper concerns the minimum viable ticket size rather than differential rejection rates as such.

What makes the gap remarkable is the risk profile of the asset being rationed. The ICC Trade Register, the most comprehensive dataset on trade-finance performance, covering more than US$25 trillion of exposures, reports default rates persistently below 0.3 percent across major products, with exposure-weighted loss rates on letters of credit of the order of ten basis points or less.⁷ Trade finance is short-tenor, self-liquidating and secured by goods with observable market value. By the standards of corporate or consumer credit, it is among the safest lending banks do.

Three caveats must discipline this evidence before anything is built on it. First, and most importantly, the Trade Register’s loss rates are conditional on approval. Low realized losses on the approved book are exactly what an effective screening process would produce if rejected applicants were unobservably riskier than approved ones; the data are therefore consistent both with the unit-cost explanation advanced here and with a classical adverse-selection account in which the rejected 41 percent is rejected because it deserves to be. The two hypotheses cannot be separated on the existing public data. They can, however, be separated by experiment, because they make opposite predictions about the same observable: if the gap is a unit-cost phenomenon, segments that become bankable when verification costs fall should perform at loss rates comparable to the incumbent book; if it is adverse selection, losses in the newly approved segments should rise sharply toward the level that justified their exclusion. Section 7 makes this the central discriminating measurement of the research agenda. Second, the lowest headline loss figures belong to letters of credit, which are bank-intermediated and doubly secured; the unmet demand is concentrated in open-account receivables and pre-shipment finance, products with thinner documentation and a different fraud profile, and the inference from one product’s loss history to another’s must be made with care. Third, the gap estimate itself is a survey measure of self-reported rejected and discouraged demand, with no price attached; some fraction of any such gap is demand that is unviable at any compliant price, and the US$2.5 trillion figure should be read as an upper bound on the addressable problem rather than a market size.⁸

With those caveats stated, the unit-cost hypothesis runs as follows. Originating a compliant trade-finance exposure requires document verification, sanctions and trade-based money-laundering screening, counterparty due diligence, collateral perfection and ongoing monitoring. These costs are largely fixed per transaction: in ADB’s 2016 survey, 90 percent of banks cited anti-money-laundering and know-your-customer compliance as an impediment to expanding trade finance, and the constraint has persisted, with more than 70 percent of banks in the 2021 survey still ranking AML/KYC requirements as the single biggest hindrance; the 2025 survey reports funding and macroeconomic constraints overtaking KYC at the top of the ranking, but compliance concerns recur survey after survey and reducing compliance costs remains among the measures banks would most prioritize to narrow the gap.⁹ When the fixed cost of safe origination exceeds the fee income available on a US$50,000 shipment, declining that shipment is the rational decision, whatever its credit quality. The gap, on this reading, is not mispriced risk; it is a minimum viable ticket size, and the SMEs below it are excluded by arithmetic rather than by riskiness.

The same cost structure would explain a second anomaly: the near-total absence of capital markets from working-capital finance. Mortgages, which are long dated, complex and historically loss prone, are funded substantially through securitization, while trade receivables and in-transit inventory, despite shorter duration and lower losses, remain overwhelmingly on bank balance sheets.¹⁰ Global factoring turnover reached €4.0 trillion in 2025, but it is dominated by bank-affiliated factors in Europe; North America accounts for roughly 4 percent of world volume.¹¹ Assets cannot be distributed to investors who cannot verify them, and verification has been too expensive to perform at the level of the individual small exposure. Heterogeneous, paper-based, operationally opaque assets do not securitize.

3.1 Technology: the marginal cost of underwriting falls

The diligence that prices small-ticket finance out of existence consists of reading and cross-checking bills of lading, invoices and packing lists, screening counterparties, detecting trade-based money laundering typologies, assessing credit, and assembling an audit trail. This is structured, document-heavy, rule-governed work of the kind that contemporary agentic AI systems perform well, with human review reserved for exceptions. The economic significance is not that machines do this work faster, but that they convert a fixed labor cost per transaction into a much lower, scalable marginal cost. If the cost of compliant origination falls by an order of magnitude, the minimum viable ticket size falls with it, and a large fraction of the rejected 41 percent becomes commercially financeable at unchanged risk appetite.

An important qualification narrows this claim. The binding constraint banks report is not only the labor cost of compliance but the liability attached to compliance failure: sanctions and money-laundering regimes impose penalties that are largely invariant to how cheaply the screening was performed, and a bank declining a small shipment is often pricing the tail cost of a compliance error rather than the hours of an analyst. Automation therefore relaxes the constraint only if it reduces error rates, not merely labor input. The proposition that machine verification can match or exceed human screening accuracy at scale is plausible, the closest empirical analogue being the fintech-lending literature, where underwriting on non-traditional digital data has been shown to predict default as well as or better than traditional credit information,¹²an analogue that bears on the information content of non-traditional data rather than directly on the error rates of automated versus human compliance screening, but it has not been demonstrated in audited trade-finance production, and we do not assume it. Practitioner expectation points the same way: in the ADB’s 2025 survey, over 85 percent of banks saw potential for AI in fraud prevention and risk management and a majority were assessing AI as a means of increasing financing capacity, though this measures expectation rather than demonstrated performance. It is the first item on the measurement agenda of Section 7. Stated precisely, the technological claim of this paper is conditional: if automated origination can achieve compliance and fraud-detection performance at parity with incumbent processes at an order of magnitude lower cost, then the minimum viable ticket size collapses. The conditional is testable within months on live transaction flow.

3.2 Law: digital documents become enforceable collateral

Cheap underwriting is worthless if the resulting claim cannot be enforced. Until recently, the legal architecture of trade ran on paper: possession of a paper bill of lading was constructive possession of the goods, and electronic substitutes had uncertain status. That has changed since 2021 across the major common-law trading jurisdictions, with the European Union’s contribution phasing in on a longer timetable. The United Kingdom’s Electronic Trade Documents Act 2023 gives electronic bills of lading and bills of exchange the same legal effect as paper,¹³ implementing the principles of the UNCITRAL Model Law on Electronic Transferable Records that Singapore and a growing list of jurisdictions have also adopted.¹⁴ In the United States, control of an electronic bill of lading has been recognized since the 2003 revision of UCC Article 7 (§ 7-106), and the 2022 amendments to the Uniform Commercial Code, now enacted in 33 jurisdictions including the District of Columbia, refine that control framework and extend perfection by control to a wider class of digital assets,¹⁵ while UCC § 9-313 permits perfection by possession through a bailee, so that the party controlling the bill of lading holds collateral that physically travels with the shipment.¹⁶ In the European Union, the ViDA package adopted in March 2025 will make structured e-invoicing and near-real-time digital reporting mandatory for intra-EU B2B transactions from July 2030, turning the invoice itself into a structured data object reported to tax authorities in near-real time, on a known statutory timetable.¹⁷ Collectively these reforms do for trade claims what land registries did for mortgages: they make the collateral legible, transferable and enforceable at low cost. The UK and Singapore reforms are in force today; the direction of travel elsewhere is dated and legislated rather than speculative.

3.3 Data: underwriting on observed flows rather than reported financials

The third shift is informational. Transport and freight management systems now generate continuous, third-party-verifiable telemetry on the physical movement of goods. A lender that underwrites on shipment data observes the trade itself, from booking and vessel to milestone events and proof of delivery, rather than relying on the borrower’s self-reported financial statements. This matters most for fraud: the classic failure mode of receivables finance is the fictitious invoice, a claim on a trade that never happened.¹⁸ An invoice anchored to an observed, independently documented shipment is structurally harder to fabricate. The precedent for underwriting on observed behavior rather than reported financials is by now well established empirically: digital-footprint and cash-flow data have been shown to predict consumer and small-business default at least as well as bureau data.¹⁹ Logistics data extends the same principle to trade credit, with the added advantage that the underlying events are physical and externally observable.

Why would these shifts liquefy supply chains rather than merely cheapen bank lending? The answer lies in decomposing the risk a financier of small-ticket trade actually bears. Only part of it is credit risk, the probability-weighted loss from non-payment, which the ICC data show to be small and which investors can price. The remainder is operational tail risk: fraud, document failure, legal unenforceability, servicing error. This component is not merely large; it is unquantifiable, because it depends on the integrity of processes the investor cannot observe. A first-loss investor cannot price what it cannot quantify, and so declines, whatever the headline default statistics say.

Automation and digital collateral attack exactly this component. When every document is machine-verified at source, every decision is logged in an auditable trail, every exception is escalated to human review under a defined protocol, and the financier controls an enforceable electronic bill of lading so that cargo releases only on repayment, the operational tail is engineered down rather than priced in. It is not engineered away; Section 6 discusses the adversarial response. What remains, if the engineering works, is predominantly quantifiable credit risk on a short-duration, self-liquidating asset, which is the raw material of securitization. The sequence is mechanical: standardized origination produces a seasoned performance record; the record supports forward-flow purchases by institutional investors; sufficient seasoning supports rated asset-backed issuance. Trade exposures then trade the way mortgage exposures trade.

It is worth being concrete about the legal architecture the mechanism presupposes, because the claim that operational risk can be engineered down stands or falls on enforceability. A reference implementation in the United States, where the deepest pool of institutional credit capital sits, runs as follows. A bankruptcy-remote special-purpose vehicle finances the shipment or purchases the receivable; the electronic bill of lading governing the goods is held to the order of that vehicle as an electronic document of title under UCC § 7-106, control of which perfects a security interest in both the document and the goods it covers (§§ 9-312(c), 9-314), with a bailee’s authenticated acknowledgment under § 9-313 available where goods are held outside a negotiable document (nn. 15 and 16 above), so that the financier’s interest is perfected by control or possession rather than by filing alone and the collateral travels with the shipment; release of the goods at destination is conditioned on payment or acceptance; the originator retains a first-loss position; and decision-level audit trails are available to funders. Because § 7-106 dates from the 2003 revision of Article 7 and is enacted in virtually every state, the structure does not depend on the still-incomplete rollout of the 2022 amendments. Functionally equivalent structures are available today under the UK Electronic Trade Documents Act and Singapore’s implementation of the Model Law, where the electronic document itself is now capable of possession. The point of specifying the structure is not novelty, since each element is settled law or settled securitization practice, but executability: the combination has only recently become available at the level of an individual small shipment, which is the unit-cost claim of Section 3 restated in legal form.

Two historical episodes suggest the magnitude of what standardization of this kind can do, and each carries a warning as well as a promise. The first is the standardization of the American mortgage from the 1970s, which converted a local, heterogeneous bank asset into a globally funded one and permanently changed the cost and availability of housing finance. The analogy must be handled honestly: mortgage liquidity was built not on standardization alone but on agency guarantees that absorbed credit risk at taxpayer expense, and no public guarantor stands behind trade receivables. Private substitutes, retained first-loss positions, decision-level transparency to investors, and the discipline of short tenor, must do the work that the government-sponsored enterprises did for mortgages, which is one reason the screening problem discussed in Section 6 is first-order rather than incidental. The second episode is physical rather than financial: containerization, which by standardizing the unit of freight collapsed handling costs and is estimated to have done more for trade among industrialized countries than trade agreements did over the same period.²⁰The liquid supply chain is, on this reading, the completion of the container revolution: the same standardization, applied to the financing of the box rather than to the box itself. The comparison is not ours alone: the ADB describes the digitalization of trade documents as potentially “transformational on a scale not seen since the size of container ships was standardized” (n. 3 above, p. 13).

5.1 Firms: growth uncapped from cash

The first-order effect operates on credit-constrained firms. The best causal evidence on what happens when constrained firms receive credit comes from directed-lending settings: Banerjee and Duflo show that Indian firms gaining access to expanded directed credit, in their setting working-capital credit lines extended to formally registered manufacturers well above microenterprise scale, used it to expand production rather than to substitute for other borrowing. Sales and costs rose nearly one-for-one with the credit, and the implied return on a marginal rupee of lending was roughly 100 percent before interest, several multiples of even the 30 to 60 percent informal market rates they document, which is the signature of binding constraints.²¹ Trade-exposed SMEs are constrained in a specific, mechanical way: a cash conversion cycle in the region of 100 to 135 days means each order locks up capital for a quarter or more,²² so a fixed bank line caps the number of orders a firm can run concurrently regardless of demand. Financing that scales with each shipment converts the constraint from a stock (the credit line) into a flow (the trade itself). For a firm earning double-digit gross margins per order cycle against financing costs of 1 to 2 percent per month, the arithmetic of relaxation is large.

5.2 Trade: the credit-supply channel runs in both directions

A substantial empirical literature establishes that trade finance is not a veil: shocks to its supply move real trade flows, and the choice of financing instrument itself responds to risk and enforcement environments, exactly the margins that digital documents and observed flows shift.²³Amiti and Weinstein show that deteriorating bank health transmitted directly to the exports of the firms those banks financed, and hit exports far harder than the same firms’ domestic sales;²⁴ Paravisini and co-authors, using matched credit-export data from Peru, estimate the elasticity of exports to bank credit supply;²⁵ and Chor and Manova show that tighter credit conditions significantly deepened the trade collapse of 2008 and 2009.²⁶Manova’s work further demonstrates that credit constraints distort not just the volume but the composition of trade, suppressing entry by smaller firms into finance-intensive sectors.²⁷

If contractions in trade-finance supply contract trade, an expansion that closes part of a US$2.5 trillion gap should expand it. A deliberately rough calculation indicates the order of magnitude at stake, and it rests on four assumptions we state explicitly. First, the financeable share: suppose only one-fifth of the measured gap proved commercially financeable once unit costs fell, roughly US$500 billion of additional annual working-capital supply. Second, the elasticity: the matched-data estimates of the intensive-margin elasticity of exports to credit are in the region of 0.2.²⁸ Third, the margins: that elasticity captures only the intensive margin; effects at the extensive margin, entry by previously excluded firms, operate in addition and are precisely where a reduction in minimum ticket size bites, although the evidence on the relative size of the two margins is mixed. Fourth, symmetry: the elasticity is identified from a credit contraction, and applying it to an expansion assumes the response is symmetric, an extrapolation rather than a finding, with the easing counterfactual in Chor and Manova the closest direct evidence in the expansion direction. On those assumptions, the implied trade effects are measured in the hundreds of billions of dollars annually, concentrated among the small firms and emerging-market exporters where rejection currently falls; relax any one of them and the number moves accordingly. We present this as an order of magnitude, not a forecast; turning it into an estimate is part of the research agenda. The ADB expects demand for trade finance to rise as trade diversification and supply-chain reconfiguration accelerate, and warns that if supply does not increase commensurately, the resulting gaps will impede the opportunities arising from that reorganization, making the supply response more, not less, consequential.²⁹

5.3 Capital markets: a new short-duration asset class

For investors, liquid supply chains create something scarce: a large-volume, short-duration, self-liquidating asset class whose performance is anchored in observable flows of goods and short maturities, even though it remains exposed to trade and macroeconomic cycles, with historical credit losses below one percent.³⁰ The macro-financial consequences of opening such a channel are familiar from the mortgage precedent: capital deepening, spread compression as investor competition replaces bank-balance-sheet pricing, and a partial decoupling of trade credit supply from the health of any individual banking system. The last point is worth dwelling on. When banks contracted in 2008 and 2009, trade finance contracted with them.³¹ A distribution channel into diversified institutional capital would make working-capital supply more resilient to banking-sector stress, although, as Section 6 discusses, it imports risks of its own.

5.4 Distribution: who gains

Because the financing gap is concentrated among SMEs and in developing regions, the incidence of closing it is progressive across firms. SMEs employ the majority of the world’s workers, and World Bank evidence indicates that financing-induced job creation is strongest among precisely the small, high-growth firms most likely to be rejected today.³² The mechanism also has a leveling property. It extends to small firms the embedded, transaction-level financing that large firms already enjoy through captive finance arms and supply-chain finance programs. Liquidity, in this sense, is not a new privilege but the generalization of an existing one.

We have stated the strongest version of the case so far. Five failure modes deserve equal prominence, and one recent collapse binds several of them together.

The case for liquid supply chains rests on a single economic claim: that the trade finance gap is substantially a unit-cost phenomenon, and that the unit cost is now falling for technological, legal and informational reasons that are observable and dated. If the claim is right, the consequences run from the cash position of a small exporter to the composition of global capital markets. If it is wrong, the gap will persist even as underwriting costs fall, and the explanation must lie elsewhere, in risk, in regulation or in demand. Section 2 set out why the existing public data cannot decide the question; only measurement on live origination can.

The research design. The claim is testable, and the ADB itself has called for exactly this measurement.³⁶A credible pilot evaluation would track, against matched controls: underwriting and compliance cost per shipment; compliance and fraud-detection error rates relative to incumbent human processes, the condition on which the entire technological argument of Section 3.1 rests; time from application to decision; approval rates for previously rejected SME segments; realized credit losses in newly approved segments against both the ICC Trade Register benchmarks and the originator’s incumbent book, which is the discriminating statistic between the unit-cost and adverse-selection explanations of the gap; the willingness of unaffiliated institutional investors to fund the resulting assets, and at what spread; and downstream firm outcomes such as order volume, employment and survival. Each is observable within the life of a single financed exposure, whose tenor runs from roughly 45 days for in-transit finance to 60 to 90 days for post-shipment receivables, which makes trade finance an unusually fast laboratory for a question that extends well beyond it: whether AI expands access to scarce institutional capability, or merely makes existing institutions cheaper to run. That question is now being asked, and funded, by the institutions building the underlying systems, and trade finance offers their measurement programs something most settings cannot: the output of the automated work is a priced, short-maturity, independently auditable financial asset, so the value of the automation can be read off realized losses and market spreads rather than inferred from surveys or task-level benchmarks.

The offer to researchers: the design window. The author’s firm is at the beginning of origination and has, as yet, no performance record; we state this plainly, because it defines what can honestly be offered. What a new channel offers that an established one cannot is the design window: almost every evaluation in empirical finance is constrained by decisions taken before the researchers arrived, variables never logged, declined applicants never tracked, rollouts never randomized, and every one of those decisions becomes irreversible here at launch. The engagement sought is specific: a small number of structured research partnerships, with priority for proposals arriving with independent third-party funding, whether from research councils, multilateral institutions or the economic-research programs of major AI laboratories, since external funding both preserves independence and ensures the work is resourced to completion.

Governance commitments. To such partnerships the firm commits the following, under agreements that protect counterparty confidentiality while guaranteeing independence of analysis and publication, including of unfavorable results, with no right of approval over findings. The pipeline will be instrumented for research from the first transaction: decision-level logs, shipment-level telemetry, and retention of declined as well as approved applications, without which the selection question at the heart of Section 2 cannot be answered. Where operationally feasible, rollout will be structured to support causal identification; any randomization would apply in one direction only, extending credit to applicants whose counterfactual is rejection and never withholding it from those who would otherwise be approved. Pre-registration of the evaluation design before the data exist, including the sample and seasoning thresholds at which performance results are first reported, is welcomed.

Because individual exposures mature in roughly 45 to 90 days, a usable panel forms at pilot volumes and within quarters, on a timetable compatible with grant cycles; the bet a researcher makes is not on the firm reaching scale, but on it reaching its first cohorts of transactions. An evaluation designed before the first transaction, by a funded examiner with no stake in the result, is worth more than any retrospective dataset: to the literature, to the institutional investors who will in time be asked to fund the resulting assets, and to the discipline of the firm itself. Researchers and research funders interested in the design window while it remains open are invited to make contact.