AI for freight forwarders, customs brokers and 3PLs.

Customs assessment stopped being a conversation at a counter. CBIC Circular 40/2020-Customs, F. No. 450/26/2019-Cus IV(Pt.), dated 4 September 2020, set out the all-India roll-out of Faceless Assessment in five phases, reaching complete coverage on 31 October 2020. Assessment moved to Faceless Assessment Groups coordinated by National Assessment Centres, which the circular records as organised commodity-wise according to the First Schedule to the Customs Tariff Act, 1975. The officer reading your bill of entry is not at your port.

Circular 45/2020-Customs of 12 October 2020 did the same to the documents. It states that every relevant document is submitted only electronically via e-Sanchit, either by the beneficiary or by the Participating Government Agency, and that verification should proceed on what was uploaded rather than requiring physical submission. TaxGuru’s account of the later Anonymised Escalation Mechanism describes an expedited-clearance request that an importer or broker can register through ICEGATE or the Turant Suvidha Kendra where a bill of entry has been pending beyond a working day, subject to a twenty-four-hour gap after filing, routed automatically to the concerned Faceless Assessment Group.

Read that as an engineering brief. The assessing officer has exactly what you uploaded and nothing else. A field that is wrong on upload is not corrected by explanation; it is corrected by a query, a re-submission and a delay, and the delay is the expensive part.

Domestically the clock is tighter. NIC’s e-way bill release notes record validity set at 200 km per day from 30 December 2020. TaxGuru’s summary of the GSTN advisories for FY 2025-26 lists two restrictions effective 1 January 2025: generation restricted to documents dated within 180 days, and extension limited to 360 days from original generation. From 1 April 2025, taxpayers with turnover of ten crore rupees or more cannot report invoices older than thirty days to the invoice registration portal. Deadlines that used to be argued are now enforced in software.

Global Trade Review reported on 9 February 2026 that the commerce and industry ministry had published draft legislation on electronic trade documents, quoting the ministry that the existing legal framework does not fully recognise them or provide a comprehensive basis for cross-border digital trust services, with feedback accepted until 11 March 2026. Trade Finance Global, writing on 2 April 2026, identifies it as the Digital Trade Facilitation Bill, 2026, records that the Directorate General of Foreign Trade ran the consultation, and notes that Section 6 requires information equivalence, reliable methods and an audit trail.

We take no view on the drafting. The operational point is the transition. Trade Finance Global notes Singapore enacted most of the UNCITRAL Model Law on Electronic Transferable Records in 2021, with modifications, and paper did not stop arriving there. Counterparties migrate on their own schedule, and the mixed period runs for years. A pipeline that reads the messy input carries you across it, and it degrades gracefully: when a genuine electronic document turns up, extraction is skipped and validation still runs.

Vendors quote a document-level accuracy figure because it is one number and it sounds good. It is close to meaningless. A commercial invoice carries a shipper address, a description of goods, a quantity, a unit price, an invoice value, an incoterm and a classification. Getting the address line slightly wrong costs nothing, because nothing downstream reads it. Getting the invoice value wrong flows into duty. Getting the classification wrong is a query, possibly a penalty, and at the wrong moment a demurrage clock.

So the first artefact is not a model choice. It is a table of fields with a cost against each, agreed with the people who currently carry the consequence: what it costs when this field is wrong and nobody notices, and what it costs when it is flagged and a human checks. Those two numbers set the thresholds, and they are why two firms handling identical document types end up with different systems.

The asymmetry does the design work. A system that is uniformly ninety-six per cent accurate on every field is worse than one that is ninety-nine on six fields and refuses on the seventh, because the second can be trusted without supervision and the first cannot.

The usual order is to build the pipeline, demonstrate it on a handful of documents, deploy, and discover the failure distribution in production. Whatever examples the engineer happened to look at become the implicit test set, so the pipeline is tuned to those and there is no way to tell whether a change made things better or merely different.

We invert it. A sample from the real inbox, not the clean folder somebody prepared. Labelling by people who know what the fields mean, with the image retained and each field located on the page, so a failure can be classified as never legible, legible and read wrong, or read right and put in the wrong field. Three different bugs, three different fixes, and one accuracy score conflates all of them.

Labelling is also where the awkward discovery happens, early enough to be cheap. Two experienced people in the same operation will disagree about which of the three dates on a bill of lading is the one you mean by shipped. Every extraction project meets this. Meeting it during labelling costs a week; meeting it after go-live costs a re-processing exercise and the confidence of the team using the tool. After that the harness is a gate: it runs per field before every release, and a regression on a high-cost field stops the release whatever happened to the average. That is ordinary data engineering discipline applied to documents.

Demonstrations run on born-digital PDFs. Operations run on a photograph taken on a warehouse floor, a scan of a scan, a page skewed so the last column is missing, and a bill of lading with a carrier stamp printed over the consignee block. Each breaks something different, which is why they belong in the evaluation set from the first week rather than in a hardening phase later.

Language is its own layer. One packing list can carry English headers, a Hindi or Gujarati supplier name, a Chinese chop and transliterated place names spelled three ways by three people. Systems built on one language per document degrade quietly: extraction succeeds, returns something plausible, and is wrong in a way no confidence score reports.

Handwriting is the case people most want to wish away. Margin annotations are frequently the operative instruction, because that is where amendments live: a quantity struck through and rewritten, a revised delivery date, an initialled correction. A pipeline reading only printed text will return the superseded value with high confidence, which is worse than returning nothing. The correct behaviour is to detect that a region has been annotated and route the document, whether or not the handwriting can be transcribed. Detecting that something was written matters more than reading it. Practically this is computer vision work before it is a language problem, and the model at the end is the least interesting component.

Most projects treat the residual as an implementation detail and lose their value there. Exceptions are not evenly distributed. They cluster on the hardest documents, from the newest suppliers, at the times of highest volume, and on exactly the high-cost fields where the threshold was set to refuse. The exception queue is the operational system; extraction is a component of it.

So we design the queue first. Who owns it by name, not by department. What is shown: the image with the field highlighted in place, the candidate value, the reason for the flag, one decision, so the reviewer never leaves the screen. How long before escalation, tied to the real clock, which for a bill of entry is the query cycle and for a domestic movement is the e-way bill validity window. What happens when the reviewer is on leave, and when the queue is longer at six in the evening than anyone will clear.

Two rules earn their place. Every exception becomes a labelled example fed back into the evaluation set, so the queue shrinks instead of becoming a permanent staffing line. And silent auto-correction is not allowed: if the system fixes a value by inference from another document, that is a recorded decision with its reasoning, not an edit nobody can find when a shipment is disputed six months on. Wiring that routing into the systems you already run is workflow automation, and it is usually the larger half of the build.

Field-level accuracy is an engineering diagnostic. Straight-through rate is the commercial one, because it is proportional to cost. A document needing one glance still consumes an attention switch and a queue slot, so the saving is a step function: a document either costs nothing or costs a person.

Three things keep it honest. It is measured per document type and per counterparty, because a blended rate hides that four suppliers generate most of the exceptions, which is often the cheapest problem in the system to fix and has nothing to do with AI. It is measured net of rework, so a document that went straight through and was corrected downstream counts against the rate. And it is reported at the thresholds you agreed, because a rate quoted at a threshold nobody would accept is a marketing number.

What we will not do: promise a rate before seeing your documents, scope from a folder of tidy PDFs, or put a model in front of an unresolved definition. Where the real difficulty is that three systems hold three versions of the same shipment, we say so first, because no extraction quality fixes that. Work runs in fixed-scope phases, and the harness, the labelled set and the infrastructure definitions are handover artefacts rather than things you ask for later. More on phases in engagement models, and on jurisdiction and data residency on our India page.

Which documents do you actually work on?

Bills of lading, commercial invoices, packing lists, certificates of origin, e-way bills, delivery orders and the annexures that travel with them. The document type matters less than the field set. What changes the engineering is which fields carry consequence and how many issuers produce the document: an invoice from forty suppliers is harder than a bill of lading from four carriers, whatever the volumes suggest.

Our documents arrive as WhatsApp photographs of printouts. Is that workable?

Yes, and it is the normal case rather than the degraded one. A photograph taken at an angle on a warehouse floor, with a stamp over the consignee block and a correction pencilled in the margin, is what the pipeline is for. What we will not do is build against clean PDF samples and meet the real distribution at go-live. We ask for a sample from the actual inbox before we quote.

What straight-through processing rate should we expect?

We will not give you a number before reading your documents, and you should be wary of anyone who does. The rate is a property of your document mix, your field set and your tolerance, not of the model. What we commit to is the measurement: a labelled set from your own traffic, a per-field cost you set, and a rate reported at thresholds agreed before anything is built.

Who reviews the exceptions, and what do they see?

Your people, in a queue we build for the purpose. The design question is not whether there will be exceptions but who owns them and how long they have. A reviewer should see the document image with the field located on it, the extracted value, the reason for the flag, and one decision to make. If they have to open the original elsewhere, the queue costs more than the manual process it replaced.

Should we wait for the Digital Trade Facilitation Bill before investing in extraction?

No. Global Trade Review reported on 9 February 2026 that the commerce and industry ministry had published draft legislation, with feedback open until 11 March 2026, on the ground that the existing framework does not fully recognise electronic trade documents. That is a statute, not a migration. Trade Finance Global notes Singapore enacted most of the UNCITRAL Model Law on Electronic Transferable Records in 2021, and paper did not stop arriving.

Do you integrate with ICEGATE, the e-way bill system and our TMS?

We build against published interfaces and whatever your existing systems expose, and we scope that work explicitly rather than assuming it. The part worth raising early is authentication and rate limits, because those reshape a design late. Multi-factor authentication on the e-way bill portal became mandatory for all users on 1 April 2025, per TaxGuru’s summary of the GSTN advisories, and an automation built around one shared login has to be rebuilt.

What does a first engagement look like, and do you subcontract?

No subcontracting: the people who scope your engagement build it. A free scoping call first, then a fixed-scope diagnostic with defined exit criteria before any pipeline is written. The diagnostic produces the labelled evaluation set, a per-field cost model and an honest projection of what can be automated. Diagnostics start in the low lakhs. We are at B-204, Kanakia Wall Street, Andheri - Kurla Road, Chakala, Andheri East, Mumbai, Maharashtra, 400093, India, Mon–Fri, 10:00–19:00 IST.