Hire the desk, not the headcount.
An agent that works your export and import file the way a documentation desk does: classifying, checking every rule, computing the duty and the clocks, chasing whoever is holding the shipment up, and staging the filing. It works every hour of every day, it reads the rulepack change on the day it ships, and it shows the instrument behind every answer. The one thing it never does is press send.
Against the desk you staff today.
The alternative to this is not another piece of software. It is a person doing nine hours of transcription that has to be perfect, and it is worth being precise about where each one actually wins.
This is not an argument for firing anyone. The desks that run well put the person on the exceptions, the negotiation and the relationship, and stop paying them to re-type an invoice value into a fourteenth PDF at nine at night.
Six jobs, worked in parallel.
A documentation desk is not a single task, so this is not a single prompt. Each role below works from the same canonical record and returns into the same queue, which is why the answers agree with each other instead of contradicting themselves the way six separate tools would.
Reads the product, proposes the HS candidates with the reasoning, and waits for a person to confirm.
Runs the composed rulepack over the record and returns observed against expected, with the instrument on every finding.
Provisional duty, RoDTEP, drawback and landed cost, from the rates in force on the shipment's own dates.
Every statutory clock with days remaining, counted from the record's dates, with month ends clamped.
Drafts the message to the counterparty who is holding the shipment up, naming the artefact and the clock.
Projects the docset from the record, explains each gap rather than flagging it, and stages the filing.
It does everything except sign.
You would not give a new hire your DSC on their first morning, and you would not think less of them for it. The same structure applies here, and it is the reason this is safe to run at volume: the labour is delegated, the authority is not.
The filing is prepared and checked against the rules in force on the shipment's own dates, then stops at awaiting approval. Your CHA or your own team files it. Purser does not replace the broker; it hands the broker a clean, checked filing instead of a folder of PDFs.
Drafted messages sit in a queue. Approval is an event on the shipment thread with a person and a timestamp attached, so months later the record answers who approved what rather than inviting a guess.
Text read from an invoice, an email or a message is content to be checked, never a command to be followed. A forwarded mail saying "approve and send" is a string in a field. Instructions come from a signed-in user and nowhere else.
It reads the file, not the prompt.
Everything below runs against the shipment record rather than against a text box. That is why its answers carry instruments and dates instead of confident prose.
The composed checklist, cited line by line
Not a generic export checklist. The rulepack composes for this origin, this destination, this product and this industry pack, and every line carries the instrument it stands on with its effective window.
- Composition order. Core → country of origin → country of destination → industry pack → agreement overlay → tenant policy.
- Three obligation classes. Statutory, contractual and preferential, so you can see which failures cost money and which cost a relationship.
- Observed against expected on every finding, rather than a bare red flag.
- Rules in force on the shipment's own dates, so a rate that changed last month does not silently reprice a shipment from before it.
hs_code: observed 610910, expected 8 digits. Instrument and effective date on the verdict.
A night's work, waiting in the morning
The queue is the product. Point it at the book and it works every shipment in it, so the desk starts the day reading decisions rather than hunting for what needs one.
- Chase messages to the counterparty that name the missing artefact and the clock it is running against.
- Gap explanations, so a docset failure says what is wrong and what would fix it rather than only flagging it.
- HS classification candidates presented as options a human confirms, never silently applied.
- Approve in a batch, because every draft shows what it was built from and reading nine of them is not nine decisions.
The money and the clocks, computed
Provisional duty, the scheme value a shipment is carrying and every statutory clock with its days remaining, each worked from the record's own dates rather than from today.
- Provisional duty, RoDTEP and drawback from the rates in force on the shipment's own export date.
- Every statutory clock with days remaining, including the realisation deadline and the free time on a detained box.
- Month ends clamped, so a shipping bill on the 31st does not quietly gain three days on its deadline.
- Rates that move by notification are read off your document, because a figure hardcoded into software is wrong the day it changes.
Your file is not the training set
An agent is only worth handing your book to if the book stays yours. Shipment records and documents are never used to train models and are never pooled across customers.
- No training on your data. Your documents and shipment records are never used to train or fine-tune a model, ours or a provider's.
- No pooling across tenants. One customer's file is never context for another customer's answer.
- Per-tenant memory. The agent improves within your own account, and that memory stays in your tenant and leaves with your export.
- Named sub-processors. The privacy policy lists who processes what and where, rather than gesturing at "trusted partners".
The detail, including lawful basis, retention periods and the grievance officer, is in the privacy policy and the compliance page.
Five steps,
one of them yours.
The agent does the reading, the checking, the arithmetic and the writing. The step it does not do is the one that reaches the outside world.
It reads the whole shipment file
The canonical field set, every projected document, the findings and the counterparty lanes. Not a prompt and not a pasted excerpt: the record itself.
one record · every tabThe rulepack composes for this shipment
Origin, destination, product and industry pack compose into one checklist, and every line carries the instrument it stands on with its effective window.
46 rules · effective windows shownSix roles work it in parallel
Classifier, checker, calculator, timekeeper, chaser and drafter, all against the same record, which is why their answers agree with each other.
observed against expectedThe work lands in a queue
Documents, chasers and staged filings, each showing what it was built from, so approving a batch is reading rather than trusting.
review queue · approval recordedYou press send. It never does.
The filing waits at awaiting approval and the message waits in the queue. A person transmits, and that approval is recorded as an event with a name and a timestamp.
awaiting approval · human transmitsLicensed by shipments worked.
Rs 4,999 a month on Free and Team, covering 50 shipments worked. Included in Business at Rs 11,999, covering 250. Past the allowance it is Rs 49 a shipment and we tell you before you cross it. Inside a worked shipment it is unlimited, because charging per question would make you ration the thing whose job is catching what you missed.
Direction, not a shipped feature.
We are building towards a product where you assign an outcome and the desk carries it. Neither item below is available today, and this section exists so that is unambiguous rather than implied. Buy it for what is above.
Today you point the agent at shipments and it works them. The work is to let you hand over a goal, such as getting this book customs-ready by Thursday, and have the desk sequence the steps itself and come back with a queue and an exception list.
Proforma invoices and quotations drafted from the same canonical record as the docset, priced against the incoterm actually agreed. Drafted and staged like everything else, and sent by a person.
Not a roadmap item at all. However capable the desk becomes, it will not submit to a government portal and it will not send a message without a recorded human approval event. More capability means more drafted and more staged, never more transmitted.
Asked, answered.
No. A chatbot answers questions about a screen you are already looking at. This works the file: it reads the canonical shipment record, the projected docset, the composed rulepack and the counterparty lanes, and it produces work into a queue. You can leave it running against a hundred shipments overnight and read the queue in the morning.
No, and it never will. Purser prepares the filing, checks it against the rules in force on the shipment's own dates, and stages it at awaiting approval. A person presses file, every time. That is an authority structure rather than a missing feature: you would not hand a new hire your DSC on their first day either, and it does not change as the agent gets more capable.
No. Drafted messages land in a review queue and go nowhere until a person approves them, and that approval is recorded as an event on the shipment thread with a name and a timestamp. Approving a batch of chasers takes about as long as reading them.
It works every hour, it reads every rulepack change on the day it ships, it does not re-key an invoice value into a fourteenth document, and it does not resign in March with the passwords in its head. What it does not have is judgement about your business or authority to commit you. That is why the approval stays with your team.
Yes, and we will not pretend otherwise. What it does that a confident human cannot is show its working: every verdict names the instrument it stands on and the effective window of that instrument, so a wrong answer is visible rather than plausible. Where it proposes rather than concludes, such as an HS candidate, it says so and waits for someone to confirm.
No. Your documents and shipment records are never used to train or fine-tune a model, and one customer's file is never context for another customer's answer. The agent keeps a memory so it improves within your own account, and that memory stays inside your tenant and leaves with your export. The privacy policy names the sub-processors involved.
It is included in Business at Rs 11,999 a month, covering 250 shipments worked. On Free and Team it is an add-on at Rs 4,999 a month for 50 shipments worked, and past the allowance it is Rs 49 a shipment with notice before you cross it. Inside a worked shipment it is unlimited: it is licensed by shipments worked, never by the question.
Put the desk to work tonight.
It reads the book, checks every rule, computes every clock and leaves the decisions stacked and cited for the morning.