Sector note · Logistics and ports

Ports lose more money to late paperwork than to slow cranes

Attention in this sector goes to the quay: berth allocation, yard optimisation, predictive arrival. The controllable loss sits indoors, in a shared mailbox, priced at so many dollars per container per day.

Alphaweb sector note

The money is not on the quay

With volumes flat and berths congested, margin comes from the cost per file, not from moving more boxes.

Ask a terminal operator or a freight forwarder where artificial intelligence belongs and the answer usually points outward: berth allocation, yard optimisation, predictive arrival times, autonomous straddle carriers. Those are real programmes with real returns. They are also expensive, slow, and tied to capital equipment on a replacement cycle measured in decades. Meanwhile the loss you can actually control this quarter sits indoors, in a shared mailbox, in a folder of PDFs, in the spreadsheet where a coordinator tracks which containers are approaching their last free day.

The macro picture sharpens this rather than softening it. UNCTAD's Review of Maritime Transport 2025 reported that seaborne trade "is expected to stall in 2025, with volumes barely rising (+0.5%)" while "ports are under strain from disruptions, leading to congestion and longer waiting times". Flat volumes and congested berths mean margin stops coming from moving more boxes. It comes instead from the cost of handling each file, and from the charges you stop paying.

Demurrage is the clearest case. A container accruing demurrage is usually not waiting for a truck. It is waiting for a customs release, a corrected bill of lading, a delivery order, a proof of payment, or a reply from someone who is on annual leave. The charge is denominated in dollars per container per day, but the cause is documentary. That is the counter-intuitive part of this sector: the highest-value software does clerical work.

Demurrage is rarely a shipping failure; it is a document that arrived late, priced by the day.

Why this work suits software that acts

Document reconciliation has a checkable right answer and regulated deadlines, which suits an agent.

Every process named here has the same shape. It is a reconciliation between documents that are supposed to agree and routinely do not: the delivery note against the booking and the signed proof of delivery; the commercial invoice against the packing list and the declaration; the carrier's demurrage invoice against the gate-in and gate-out records and the free time actually contracted. There is a checkable right answer and an authoritative reference to check it against. That is a far better fit for an agent than anything open-ended.

The work is also deadline-shaped, and the deadlines are written down. The US Federal Maritime Commission's billing rule requires vessel-operating carriers and marine terminal operators to issue detention and demurrage invoices within 30 calendar days of charges last being incurred, gives billed parties at least 30 calendar days to request mitigation, refund or waiver, and states plainly that "failing to include any of the required information in a detention or demurrage invoice eliminates any obligation of the billed party to pay". Those are clocks. Human teams miss clocks at volume, particularly at month-end. Software does not.

The inputs are documents rather than feeds, which is usually treated as a problem and is in fact the opening. The Digital Container Shipping Association put electronic issuance at roughly 11% of bills of lading in 2025 — "a significant step forward compared to other digital trade efforts, but still far from the level needed to reach full scale" — against a collective industry commitment to reach 100% by 2030. Nine in ten still arrive as a scan attached to an email. That is a poor basis for classical EDI integration and a workable basis for software that reads.

None of this is speculative. McKinsey's account of AI in freight logistics reports one transportation company that deployed 50 AI agents and automated 60% of check calls, 73% of order acceptances and 80% of paper invoice payments, saving tens of thousands of hours of labour; a separate transportation company's AI-enabled supply chain platform, applied to “pricing, capacity sourcing, freight tracking, and document handling”, has boosted productivity by more than 40% since 2022. Note what those numbers attach to. Invoices, orders, documents. Not cranes.

Finally, the economics are cumulative rather than individual. A forty-pound discrepancy on one delivery note does not justify an hour of a person's time. Forty thousand of them justify a department — which is precisely why those departments exist, and why they are always behind.

Three that go first, and one that does not

Delivery notes, then document chasing, then demurrage recovery; customs file assembly waits its turn.

Sequence matters more than ambition. Begin where ground truth is unambiguous, where a mistake surfaces within a day, and where nothing the software does constitutes a legal filing. That rules in delivery note matching, document chasing and demurrage recovery, in that order, and it rules out the process most people want to start with.

Customs file assembly is deliberately not on the list, and any vendor who puts it first is selling. The declarant carries liability for what is filed. An agent can gather the commercial invoice, packing list, certificates of origin, licences and transport documents, reconcile values and quantities across them, flag the classification questions it cannot settle, and present a complete, evidenced file. A named, licensed person still reviews it and files it. Build that in month four, once the first three have earned the organisation's trust, and build it as assembly with a human declarant rather than as autonomous filing.

What actually stops this

Access, ownership and undocumented practice stop this; only decaying reference data is a technical problem.

The first obstacle is not the model. It is that the data lives in other people's systems: the terminal operating system, each carrier's portal, the national customs platform, the forwarder's TMS, the customer's ERP. In a pilot, nobody grants you API access, and the procurement cycle to get it outlasts the pilot. The workable path is the ugly one — the shared mailbox, portal credentials the customer already holds, and scheduled file drops. Plan for it, price for it, and do not let a demo built on clean API data set the expectation.

The second is that nobody owns demurrage. It is incurred by operations, disputed by customer service, paid by finance and recharged by commercial. An agent that produces forty well-evidenced mitigation requests a week produces forty tasks that no one's objectives reward, and they will sit. Name the owner and put the recovered amount into their numbers before a line of code is written.

The third is that the documented process is a fiction. The routine cases follow it; the rest live in the head of a coordinator who has sat at that desk for fourteen years and knows which carrier's free time excludes weekends at that terminal, which consignee always sends the wrong HS code, and which agent will release against a scanned guarantee. If you do not buy two days a week of that person's time for the build, you will automate the fiction and find out in production.

The fourth is genuinely technical, and it is the one that never goes away. These agents depend on reference data that decays: tariff schedules, free-time terms by carrier and terminal and contract, port and customs holiday calendars, incoterm-driven cost allocation. An agent working from a stale free-time tariff does not fail loudly. It files confident, wrong claims and burns the relationship with the carrier's customer service desk. Maintaining that layer is a standing cost — a part-time person and a monthly review, not a one-off data load. Budget for the second year, because the second year is not free.

What the auditor will want to see

Keep the source document, rule version, approver and timestamp for the audit period, not the logging default.

The regulatory question in this sector is not whether software did the work. It is whether you can reproduce the decision eighteen months later. European customs is moving in a direction that makes this sharper. In March 2026 the Council and Parliament agreed the customs reform, under which "businesses importing to and exporting from the EU will only need to submit customs information once to this single portal" — the EU Customs Data Hub, operational for e-commerce goods from 1 July 2028 and for all goods movements from 1 March 2034 — alongside an EU Customs Authority in Lille that will analyse continuously updated import and export data to identify the riskiest cargo for inspection.

Read that as a supervisor would. When declarations are visible in aggregate and analysed continuously, a systematic classification error introduced by your software is not a small mistake repeated many times. It is a pattern, and it is detectable from the outside before you notice it from the inside. Retain the evidence accordingly: the source document as received, the version of the rule applied, the identity of the human who approved, and the timestamp — kept for the audit period rather than for whatever your logging platform defaults to.

Ninety days, and what it can honestly settle

Ninety days tests the method on one lane with one named owner; it does not transform an operation.

Ninety days is enough to prove or disprove the thesis on one trade lane. It is not enough to transform an operation, and a plan that promises otherwise should be read as a sales document. The sequence below assumes a single lane, a single customer or business unit, and a named operational owner with the recovered amount in their targets. If you cannot secure those three things, the problem is not ready, and no amount of engineering will make it ready.

What it is worth

Measure demurrage paid, chasing hours and amended entries first; a single ROI figure means nobody looked.

We will not give you a percentage. The honest range is two orders of magnitude wide, because demurrage exposure depends on free time negotiated, lane mix, inland infrastructure and how disciplined your document flow already is. A forwarder with generous contractual free time, steady volumes and good pre-alert discipline will find little here. One moving project cargo through a congested port on spot terms will find a great deal. Anyone quoting you a single figure across that spread has not looked at your files.

So measure three numbers before anyone builds anything. How much did you pay in demurrage and detention over the last twelve months, and what share of it did you accept without dispute. How many hours a week does your team spend chasing documents that should have arrived unprompted. How many customs entries did you amend after filing, and for what reason. If nobody in the business can produce those figures within a fortnight, that is itself the first finding, and it says more about where the money goes than any pilot will.

Software that carries a process end to end is worth what the process costs you today, less what it costs to run the software and keep its reference data honest. Establish the first number properly and the second stops being a leap of faith.

A ninety-day sequence that survives contact

1. Days 1-15 — Baseline the three numbers and pick one lanePull twelve months of demurrage and detention paid, split into disputed, recovered and silently accepted. Count hours spent chasing documents and the number of customs entries amended after filing. Then choose one trade lane and one business unit, and get a named operational owner who carries the recovery figure in their targets. Resist the urge to pick the worst lane; pick the one with the cleanest records, because you are testing the method, not the crisis.
2. Days 16-30 — Sit with the coordinator and map the real processTwo days a week of the person who actually does the work, not a process map from the quality system. Capture the exceptions by name: the carriers whose free time excludes weekends, the consignees who send the wrong classification, the terminals whose gate records lag by a day. In parallel, secure the unglamorous access — mailbox, portal credentials, file drops — because that negotiation takes longer than the build and will not be hurried later.
3. Days 31-50 — Run delivery note matching in shadowBuild the matcher and run it read-only against the last quarter's completed files, where the answers are already known. Measure the false-positive rate on claims, not the accuracy rate on extraction, because a claim raised in error costs a relationship and a missed one costs only what it was worth. Tune until the operations owner agrees they would have sent every claim it raises.
4. Days 51-70 — Take document chasing live, demurrage in shadowChasing goes into production with human approval on every outbound message for the first fortnight, then approval by exception. At the same time, rebuild the free-time clock in shadow and reconcile it against the carrier invoices you have already paid. The gap between your reconstruction and the carrier's is the honest measure of what recovery is available, and it is usually smaller than the vendor slide and larger than finance expects.
5. Days 71-90 — File real claims, publish the number, then decide on customsA named human sends agent-prepared mitigation requests inside the regulatory window. Publish the count filed, the amount conceded and the amount refused with reasons, so the second phase is argued from evidence rather than from enthusiasm. Only then decide whether to start customs file assembly, and scope it as assembly reviewed by a licensed declarant rather than as autonomous filing.

What we read

The documents behind this note. Each entry says what it is, what it found, and why it should change what you do — then the link to the original.