For freight forwarders, customs brokers, 3PLs and the ops desk that lives in email

AI agents for logistics and freight forwarders: every document checked before it is filed, every exception caught before the customer calls.

An agent that reads the commercial invoice, packing list and bill of lading as they arrive, cross-checks them against each other and the booking, flags the mismatches that would have bounced at customs, chases the shipper for what is missing, suggests the classification for the broker to confirm, and watches each shipment for the exceptions that eat an ops team’s day. The licensed broker still files. The customer still gets a person. The desk stops re-keying.

Thirty minutes, free. You leave knowing which workflow to start with, what a pilot covers and what it costs.

PilotAI Agent Pilot: one workflow, 2–3 weeks, $4,500 (₹3,50,000) fixed. What the pilot includes.

Who this is for: Freight forwarders, customs brokers, NVOCCs and 3PLs with 5 to 100 staff, a few hundred to a few thousand shipments a month, documents arriving by email and WhatsApp as PDFs and photos, a forwarding system or a very full shared drive, and an operations desk whose day is reading documents, re-keying them and chasing people.

The week this fixes

If this is your Monday, keep reading

The mismatch is found at filing, not at booking

The packing list says 48 cartons and the invoice says 46. The consignee’s address on the bill of lading has a different pin code from the one on the invoice. Nobody noticed until the entry was keyed and queried, and now the container sits while someone emails the shipper, who is asleep in another time zone.

Every shipment is forty emails

Documents arrive one at a time, in three threads, with the wrong subject lines. The desk asks for the missing certificate of origin twice, the shipper sends the wrong version, and the rate request from another customer waits in the same inbox for a day.

The customer learns about the delay from the tracking page

A vessel rolls, a transhipment slips, a truck is stuck at a checkpost. The ops desk knows at 10 am and tells the customer at 4 pm, after the customer has already called twice. The relationship is decided by the exceptions, and the exceptions are handled last.

One workflow, before and after

From documents in the inbox to a filing-ready set, with the shipper chased for the rest

Trigger: An email from a shipper at 6:10 pm with three attachments: "PFA docs for the Nhava Sheva shipment, please file tomorrow."

Today

  1. ShipperSends the commercial invoice, the packing list and a draft bill of lading. The certificate of origin is not attached. The invoice is a photo of a printout.
  2. Ops deskOpens the email at 9:20 am. Re-keys the invoice lines into the forwarding system. Notices at 11 am that the carton count differs between the two documents. Emails the shipper.
  3. ShipperReplies at 3 pm with a corrected packing list. Still no certificate of origin; nobody asked for it yet.
  4. BrokerStarts classification at 4 pm from the invoice description "machine parts". Asks the desk for the technical specification. The desk asks the shipper. The filing moves to the next day.
  5. CustomerCalls at 5 pm to ask whether it was filed. It was not.

With the agent

  1. AgentReads the three attachments at 6:11 pm, including the photographed invoice, and extracts shipper, consignee, line items, quantities, weights, values, Incoterms and ports into the shipment record as a draft.
  2. AgentCross-checks: 48 cartons on the packing list against 46 on the invoice; consignee pin codes; gross weight against the booking; invoice total against the sum of the lines. Flags two discrepancies with the exact fields and pages.
  3. AgentChecks the document set against the lane’s checklist (certificate of origin required for this origin and product) and emails the shipper at 6:14 pm: the two discrepancies, the missing certificate, and a one-line list of what to send back. Follows up at 9 am if nothing has arrived.
  4. AgentSuggests a classification for each line from the description and the firm’s own history of similar shipments, with the reasoning and a confidence note, as a draft for the broker. Nothing is filed.
  5. Broker and deskAt 9:30 am: a filing-ready set with the corrected packing list, the certificate received overnight, the discrepancies closed, and a classification to confirm or change. The broker files before lunch. The customer gets "filed, here is the number" before they thought to call.

What stays with a person

  • The filing, the classification decision and the valuation: the licensed broker signs, and the agent’s suggestions are drafts with reasons attached.
  • Anything with a regulatory edge: restricted goods, licences, a declaration that looks wrong in a way the rules do not cover. Escalated with the documents, not improvised.
  • The customer conversation about a delay or a dispute. The agent drafts the first notice; a person sends it and owns what follows.
What it saves

Documentation hours, filing delays and customer calls: a worked example

MeasureTodayWith the agent
Desk hours on reading and re-keying, per month~165 hours~35 hours
Desk hours on chasing documents, per month~100 hours~25 hours
Discrepancies foundA day after arrival, often at filingWithin minutes of arrival, before anyone keys an entry
Shipper chased for missing documentsWhen the desk gets to itWithin minutes, with a follow-up the next morning
Customer told about an exceptionWhen they callDrafted the moment the carrier status changes, sent by a person

Worked example. The assumptions:

  • A forwarder handling about 400 shipments a month, each arriving with three to five documents by email or WhatsApp
  • About 25 minutes of desk time per shipment on reading, re-keying and cross-checking by hand, and about 15 minutes more on chasing missing or corrected documents
  • About a quarter of shipments have at least one discrepancy or missing document today, found on average a day after the documents arrived
  • The agent extracts and cross-checks about 85% of document sets without a person touching them; the rest are photographed or unusual and reach the desk with the fields it could read filled in

The hours are real and the pilot measures them. The larger number, which we will not estimate here, is the cost of a filing that bounced: demurrage, detention, a re-file, a customer who moves their next shipment elsewhere. A pilot counts discrepancies caught before filing against the previous quarter, and days from documents-received to filed, rather than quoting a saving on penalties.

Put your own numbers in the automation savings calculator, or get a price range from the AI agent cost estimator.

The agents

Usually one to start, three when it is working

Each agent has a small set of tools it is allowed to call and nothing else. That list is the guardrail, and it is written down here so you can see it.

Documentation check agent

Watches the ops inbox and WhatsApp for shipment documents, extracts the fields from PDFs and photographs into a draft shipment record, cross-checks invoice, packing list, bill of lading and booking against each other and against the lane’s document checklist, and produces a discrepancy list with the exact fields. Chases the shipper for corrections and missing documents on a schedule the desk sets, and stops chasing the moment a person replies.

  • read_inbox
  • extract_document_fields
  • cross_check_documents
  • check_lane_checklist
  • request_from_shipper
  • update_shipment_draft
  • escalate_to_desk

Classification assistant

For each line item, suggests the classification from the description, the technical specification where one exists, and the firm’s own history of similar shipments, with the reasoning written out and a confidence note. Everything is a draft for the licensed broker; the agent has no tool that files, and says so when asked.

  • lookup_similar_shipments
  • suggest_classification
  • request_specification
  • draft_for_broker

Shipment exception agent

Reads carrier and tracking status for every open shipment, recognises the exceptions that matter (rolled, transhipment missed, held, documents queried, detention clock started), drafts the customer notice in the firm’s template with the new ETA and the next step, and puts it in a person’s queue to send. Produces the morning list of shipments that need a decision today.

  • read_tracking_status
  • classify_exception
  • draft_customer_notice
  • queue_for_approval
  • morning_exceptions_list

Guardrails

  • The agent never files, never signs and never submits a declaration: there is no tool for it. Classification, valuation and the declaration itself are the licensed broker’s, and every suggestion carries its reasoning so it can be checked in seconds.
  • Extracted fields are drafts until a person confirms them; a discrepancy the agent cannot resolve from the documents is escalated with the pages, not guessed.
  • Restricted or licensed goods, dangerous goods declarations and anything the lane checklist marks as regulated go to a person before any shipper is contacted.
  • Customer notices about delays and disputes are drafted, not sent: a person sends them and owns the thread. Shipper chasers for missing documents are sent by the agent within rules the desk sets, with a cap on reminders.
  • Commercial data (prices, consignees, lanes) stays in the forwarder’s own systems and accounts; the model sees the shipment it is working on, not the archive, and nothing is used to train anything.
  • Every extraction, check, message and suggestion is logged against the shipment, so a query from customs or a customer has a complete trail.

Systems it talks to

  • Email (Gmail or Microsoft 365) and WhatsApp Business API, where the documents actually arrive
  • Forwarding and customs software: CargoWise, Logi-Sys, Softlink, Freightify, or the shared drive and spreadsheet the firm actually runs on
  • Carrier and tracking feeds: the lines’ and airlines’ tracking pages or APIs, port community systems where available
  • The firm’s own history: past shipments and classifications, as the reference the suggestions draw on; a shared sheet for the lane checklists and the exceptions list
Proof, and its limits

What we have built, and what we have not measured yet

The document chaser on the demos page is this design with bank statements swapped for shipping documents: it reads what arrived, checks it against a list, and asks for the rest. The CA-firm reference build shows the chase-and-check loop with its tool list and guardrails. We have not yet published a forwarder result and will not quote one until a firm agrees to publish theirs; the pilot measures discrepancies caught before filing and days to filed on your own shipments.

FAQ

Questions logistics and freight forwarders ask us

Can it read photographed invoices and scanned packing lists?

Yes, including photos of printouts taken on a phone, within limits: a blurred total or a cropped page is flagged as unreadable rather than guessed, and the shipper is asked for a clean copy. The pilot measures the share of documents extracted cleanly on your own inbox in week one, before anything is built on it.

Does the agent classify goods or file the entry?

It suggests classifications with its reasoning for the broker to confirm or change, and it has no tool that files. Classification and the declaration are the licensed broker’s decisions and liability, and the agent is built so it cannot be talked into taking them. What it removes is the hour of reading and re-keying before the broker can start.

Which forwarding systems does it work with?

Systems with an API or an import format (CargoWise, Logi-Sys, Softlink, Freightify) can take the draft shipment record directly; a firm on a shared drive and a spreadsheet starts with the sheet in the middle, which is often the quickest pilot. Email and WhatsApp are where the documents arrive, and those are connected in every case.

What about customer data and confidentiality?

Documents stay in your email, your drive and your forwarding system; the agent reads the shipment it is working on and writes a draft record, and nothing is used to train a model. Where a customer requires it, the agent can be limited to their lanes or excluded from their shipments. The data-flow diagram is written in week one of the pilot and is yours to show a customer.

What does it cost?

The AI Agent Pilot covers the documentation check agent on your inbox for one lane or one customer for a fixed fee in two to three weeks, measured on discrepancies caught before filing and desk hours against your last quarter. The classification assistant and the exception agent are scoped after. The cost estimator gives a range for your volume.

Next step

Tell us about your shipments

Roughly how many a month, which lanes, where the documents arrive, which forwarding system. We reply with what a pilot on your documentation desk would cover.

Or start with a number: the AI agent readiness quiz takes three minutes and gives you a score and a 30-day plan.

Or send the details

Name and email are all we need. The rest helps us reply with something specific.