By Artem Korsakov ·
A theoretical example. An agent in Ningbo sends a freight forwarder an email in English: two 40HQ containers, GW 18,600 kg, 56 CBM, vessel departure moved from 14 to 21 November. Attached are an invoice and a packing list. The logistics coordinator copies the data into the shipment card and into the request for the rail leg from Vladivostok. The packing list shows a gross weight of 18,060 kg; the email says 18,600. Nobody cross-checks, and the discrepancy surfaces at the station when the weight does not match the waybill.
An AI assistant parses such emails and attachments into the card's fields, shows where each value came from and immediately flags discrepancies between documents. The coordinator reviews them and decides what to clarify with the agent.
In short: the first product parses similar agent emails, cross-checks details between documents and prepares questions. The customs commodity code (HS code; in Russia, the TN VED code), sanctions restrictions, permits and confirmations to the client stay with specialists.
Why free text is harder than a regular form
A client may write “same as last time”, send dimensions without units, give a date without a year or attach an old packing list. Information from several emails is not always consistent. If you simply ask a model to “make a request”, it may produce neat text in which assumptions look like confirmed data.
That is why the product needs structure before the model is connected. For the chosen service, define the required fields, the allowed formats and the signs of uncertainty. The source of each value must be available to the reviewer. If the information is missing, the field stays empty with an explanation. You must not fill in the usual country, currency or unit just because that was the case in most past requests.
Scenario 1. A draft request from an email
Start with one incoming channel and one type of service. The assistant extracts the sender, pickup and destination points, dates, number of packages, weight, dimensions and contact. The exact set is defined by the logistics coordinator. Special requirements and questions that could not be parsed unambiguously go into a separate block.
Next to each field there should be a link to the email, the page of the attachment or the text fragment. The reviewer quickly compares the value with the original and does not have to search for it in a long thread. Based on the result, they can accept the field, correct it or request clarification. Only after review does the draft become a working request.
A hypothetical example: the email says “12 boxes, 480 kg”, while the attached file lists 14 packages. The assistant shows both values and raises a question about the number of packages. It does not automatically pick the latest figure and does not tell the client that the request has been accepted with specific parameters.
Training example: the email and the attachment give different numbers of packages. The AI shows both values; the decision is made by the logistics coordinator.
Scenario 2. Cross-checking details between documents
AI-assisted document checking is useful for finding candidate errors. You can match invoice numbers, party names, number of packages and shipment identifiers. But first you need to agree which fields must match and which source is considered primary. A difference does not by itself mean a violation.
Some checks are better done with ordinary rules: number format, whether a field is required, matching identifiers, an allowed unit of measure. The model helps where non-standard text has to be read or the corresponding field found. Combining the two approaches makes the result explainable: a formal mismatch and a model's assumption are shown separately.
Net weight must not be compared with gross weight as if they were the same quantity. The consignee name and the delivery address may belong to different roles. A translated description is not always literal. In the interface, the disputed spot is flagged for a person, and the source document is not changed. The specialist's confirmation and the reason for the decision are kept in the history.
Scenario 3. Questions for the client and a short summary
After parsing the request, the assistant can prepare a short list of clarifications: what the pickup address is, when the cargo is ready, whether the packaging and specifications are confirmed. The questions are based on missing data; they contain no invented requirements. Before sending, an employee checks the recipient and the text.
For a long thread, a summary is useful: what has been agreed, what has changed and which questions are open. Separate confirmed terms from preliminary discussions. An agent's phrase “we'll try to make it” does not become a promised deadline. Every significant agreement needs a link to the source and a date, especially if a later email changed it.
At the first stage, a “prepare draft reply” button is enough. Automatic sending is better left out of the basic pilot. It requires separate control of recipients, authority and errors. Attachments and changes to payment details go through the company's established verification process. A convincing tone in an email does not replace it.
Languages, units and time need separate checks
International requests contain different date notations, decimal separators and measurement systems. “03/04” without context is ambiguous. “Delivery Friday” does not give a time zone or an exact date. Inches and centimetres give very different volumes. That is why, next to the normalised value, you keep the original entry and an explanation of the conversion.
Keep company names and addresses in the original language. For search, you can add a translated variant, but do not replace the original details with it. Abbreviations and item codes require caution: a similar spelling does not always mean the same product. Poor scan quality, handwritten corrections and illegible stamps should lead to declining to fill in the field with confidence.
The time of an event and the moment a message was received also differ. The assistant can extract an ETA from an email, but it remains a forecast reported by a specific source. It must not be turned into a confirmed arrival just because the stated date has come. The card stores the date type, the time zone and the update time.
Training example: three date entries the assistant must not complete on its own. Each entry has its own ambiguity. The assistant shows the original and asks a question; it does not substitute a guessed date.
General statistics on average delivery times are also unsuitable for answering a specific client. A working answer needs confirmed information from the carrier, the leg, the update time and the current ETA. The assistant can find them in permitted messages and explain discrepancies to the coordinator.
What not to include in a small AI pilot
Do not delegate to the model the final classification of goods under the customs commodity nomenclature (HS / TN VED), sanctions screening, export control or required permits. Such decisions depend on current rules, jurisdictions, product characteristics and the parties involved. An error has consequences that a convenient interface cannot compensate for. If necessary, the product can pass information to a specialist and save the confirmed result.
Do not promise automatic selection of a “legal route”, a guaranteed date or a complete counterparty check. Nor should legal obligations be determined from a single Incoterms term. Within the first implementation, the assistant gathers information for an authorised employee. Expanding its authority requires a separate project and control.
Instructions contained inside a received document must not control the system's actions. An attachment is a data source. It does not become an order to disable checks or send files. Allowed actions are defined in advance. For a pilot, you can give the assistant no rights at all to send, change reference data or access unrelated shipments.
A separate queue for international exceptions
Do not merge “could not read” and “needs a specialist's decision” into one error status. An illegible invoice requires a new copy. A mismatch in package counts requires checking which document version is current. An unclear delivery term requires clarifying the agreements. The owner and the next action differ in these cases.
In the first product, a few routes are enough: clarify the source, check the details, agree on the terms, hand over to a specialist. The model suggests a category and a short explanation with the source. The employee confirms it, after which ordinary logic creates a task in the existing records. This way the assistant sorts the incoming flow instead of adding a long answer to every document.
How to organise quality checks
Select an anonymised sample of typical requests: plain text, several attachments, a corrected version, a poor scan, different languages and contradictory information. An expert labels the required fields and expected questions in advance. In addition to the share of correctly extracted values, check the cases where the model should leave a field empty.
Count missed discrepancies and false alarms separately. If the tool creates dozens of unnecessary questions every time, the employee will stop using it. Check the full working time: upload, waiting, review and correction. Nice recognition of a single file does not prove savings on the real flow.
To accept a pilot, agree in advance on acceptable errors and mandatory manual checks. These thresholds apply to your process; they are not universal standards. After changing the model, template or recognition method, rerun the control sample. Keep decisions made and configuration versions so you can analyse disputed results.
Commercial data and access rights
Before passing documents to an external service, define which information is allowed, and the provider's conditions for processing, storing and using data. Limit access to a specific team's shipments, provide for deletion and an activity log. Questions of applicable requirements for personal and commercial data are resolved separately with the responsible specialists.
The first version can be built as a separate queue of drafts without changing the main system. Then add an agreed export or integration. This lets you test usefulness and quality before widening access. The employee should see the original, the result and the available action on one screen, and be able to finish the work manually when the model is unavailable.
How to fit the assistant into your existing set of services
If the company already works in a TMS, 1C (a widely used Russian business accounting platform) or spreadsheets, we first check the built-in templates, recognition and import. If the needed step is already there, it makes sense to configure it.
For a custom product, it is enough to cover the missing piece: email or an approved folder, an extraction queue, review by an employee, an agreed handover into the existing records. If no exchange is available, you can start with a file export. The customs broker's system, electronic document exchange and carrier portals continue to do their jobs. They are not replaced with a new portal for its own sake.
Next step
To begin, we choose one type of request and anonymised examples. Processing international freight requests becomes the first scenario: we define which fields to extract, what to check with rules and where a human decision is mandatory.
Related reading: Documents for international shipping: completeness and versions; Implementing AI in business: how to test one task in a pilot.
Частые вопросы ### Can the assistant be connected to any documents right away?
Technically uploading a file does not yet mean reliable processing. Start with a limited set of formats and scenarios, then expand it after testing on examples.
Will the assistant replace an international logistics specialist?
In the product described, it prepares data and highlights questions. The specialist checks the information, makes decisions and confirms actions within their authority.