What does AI actually do in freight forwarding?

Start with the work, not the word.

In a forwarding company the work arrives as messages. An enquiry. A rate reply from an agent. A question about where the cargo is. A carrier bill. Somebody opens each one, works out what it is, finds the information and writes the answer. AI takes some of those steps.

Five workflows are where it does real work today. A product can be strong at one and absent at the next.

1. Reading the enquiry

An enquiry rarely arrives as a form. It is three lines of email with a packing list attached, or a photograph of a handwritten note. The job is to turn that into shipment facts and notice what is absent. No dimensions means no chargeable weight, which means no price, which means the enquiry sits.

Worked example: “Spots what is missing in an enquiry and drafts the email asking for it.”

2. Pricing from your own rate sources

A price comes from one of three places. Your own rate cards. A carrier account read through your own credentials. Or another person, which is the next workflow. For the first two, software can do the whole retrieval: find the card, apply the break, add the surcharges, check it against your margin rule. What comes out is a priced draft, not a sent quote.

3. Chasing the agents you have to ask

Every forwarder has lanes it does not hold. Somebody writes to four agents, chases two on Thursday, forgets the third and quotes late. Nobody can make an agent answer faster. But writing the request, sending it to all of them at once, and chasing whoever goes quiet is mechanical work.

Worked example: “Emails all your agents for a lane in one go, with the route and charge lines filled in, then chases the ones who do not reply.” Software controls the asking and the chasing, never the answering.

4. Answering “where is my cargo?”

The highest-volume, lowest-value email in the business. The answer already exists in the carrier’s system, and the cost is that somebody leaves the thread to fetch it. FreighAI states the mechanism this way. Answers “where is my cargo?” on the customer’s own thread, from carrier status, for approval. If the carrier has not scanned the shipment, no software knows where it is.

5. Matching the carrier bill

A carrier invoice arrives with many airway bills on it. Every line has to find its job and be checked against what was agreed. By hand this is slow, and it is the first thing dropped in a busy month. Dropping it is how a rate difference becomes a permanent loss.

Worked example: “Reads a carrier bill and matches every airway bill back to its job, checking line by line before it is booked.” The checking is mechanical. Deciding whether to dispute a charge with a carrier you depend on for space is not.

The five workflows are easier to judge when you follow one enquiry through them. Take the ocean enquiry above.

• 01It arrives at 09:14, three lines long, weights missing. It is read: two high cubes, Nhava Sheva to Rotterdam, door to port.
• 02The gap is spotted and the email asking for weights is drafted.
• 03The rate card is pulled, and the two agents covering the inland leg are written to at once.
• 04A quote is drafted, margin rule applied, last quarter’s price beside it.
• 05Anita in pricing reads the draft next to the thread and approves. It sends in her name.
• 06The agent who has not replied is chased without anybody remembering to.

Only one step involves a person, and it is the step where money is committed. The rest is what a good coordinator does, without anybody holding it in their head.

A demonstration shows this running cleanly. Ask what happens when the weights never arrive, and when an agent quotes in another currency.

It overlaps, and the difference is worth being precise about.

Rule-based automation does exactly what you told it to do. If the subject line contains RFQ, create a record. It is fast, cheap to check and completely predictable, which is a real virtue. Its weakness is that freight email does not arrive in the shape your rules expect, so somebody spends their life extending the rules.

Software that reads works the other way around. It is given the message and works out what it is, including messages nobody wrote a rule for. That is why it copes with the forwarded thread and the photographed note. Its weakness is the mirror image. It is not fully predictable, so it needs a person where a mistake would cost something. Most products that work are a mix of both, so ask which half a vendor means.

A second question hides underneath: where the software sits. Some products live inside the mailbox and help one person answer the message they opened. Others sit in front of it and turn every message into a task before anyone opens it. Both are legitimate, and different purchases. One product states its position as: “FreighAI is an AI-first freight operations platform that can run standalone or work alongside an existing TMS or ERP. It does not require a forced replacement.”

This section matters more than the three before it. A claim is only useful once you know its edge.

It does not file your customs entries

Customs filing is a regulated act performed by a party the authority recognises. US Customs and Border Protection issues customs broker licences to “persons (i.e., individuals, corporations, partnerships, and associations)” to conduct customs business on behalf of others, under section 641 of the Tariff Act of 1930. The licence attaches to a party, not to software.

The declaration is a demanding document. HMRC’s guidance on a full import declaration lists the customs procedure code, the commodity code, the declaration unique consignment reference, the consignee and consignor, the transport methods and costs, the valuation methods, and the certificates and licences. Software can pre-fill much of that from documents it has already read. Choosing the commodity code and standing behind the entry belongs to the licensed party.

It does not keep your statutory books

Checking a carrier invoice and keeping a statutory ledger are not the same job. Bank reconciliation, accruals, provisions and anything an auditor signs stay in your accounting system. Telling you a bill does not match the job is useful. It is not bookkeeping.

It does not make the judgement calls

Some decisions look like arithmetic and are not. Whether to go below your floor to hold a customer through a bad quarter. Whether to trust the agent who was late twice. Whether to tell a customer their cargo has rolled. These belong to your people. And nothing here touches the cargo itself. It does not load a container or run your warehouse floor. It works on the messages and the records about the cargo.

Three failure modes are worth understanding before you buy anything.

It needs a person at the point of consequence

This is not a weakness better models will remove. It is how AI systems are expected to be built. The European Commission’s summary of the AI Act says high-risk systems must carry “appropriate human oversight measures”, “logging of activity to ensure traceability of results” and a “high level of robustness, cybersecurity and accuracy”. A freight quoting workflow is almost certainly not high-risk, and this guide does not claim it is. The point is the shape of the expectation.

Stated as a product principle: “Your team approves. They do not compose, forward, file or chase. The reply goes out in their name.”

It needs something to read

Every capability here begins with data. Take the data away and the capability goes with it. If your rate cards live in one person’s spreadsheet with merged cells and a colour code only they understand, nothing can price from them. If an agent replies “as per last time”, there is no rate to read.

The industry knows this. IATA describes e-freight as an initiative that “aims to build an end-to-end paperless transportation process for air cargo through a regulatory framework, electronic messages and high data quality”. Format range is fair to test: one product states its range as “Reads PDF, Excel, Word, forwarded email, WhatsApp messages, photos and scans.” Bring your worst scan.

It is weakest where your business is hardest

Common lanes and enquiries shaped like the last thousand are handled well. Accuracy drops on the unusual case: the out-of-gauge piece, the shipment with two consignees, the dangerous goods line that changes the handling. Those are also the shipments that carry your margin.

NIST’s AI Risk Management Framework is “intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems”. Measure is one of its four functions, and it sits in the middle for a reason.

Two questions cut through most of it. Carry them into every demonstration, and ask them of every product, including this one.

Does your team still have to open the mailbox?

This separates two categories that use the same words. If yes, you are being shown an assistant. It helps the person who already opened the message, so the saving is capped by how many messages one person can get through in a day. If the answer is no, what reaches your team is already prepared and assigned, which changes the ceiling rather than the speed. Neither is wrong. A vendor who cannot answer clearly has not decided what they are selling.

What runs by itself?

Most of the cost in a forwarding office is not the work. It is remembering the work. The agent who has not replied since Tuesday. The quote that expired quietly. The invoice never matched. Ask for the list of what happens without a person remembering. As an example: “Chasing, tracking checks, quote expiry and follow-ups run on their own. Nobody has to remember.” The test is whether somebody can show you the chaser going out.

The OECD’s AI Principles are a fair frame. Three of the five are “Transparency and explainability”, “Robustness, security and safety” and “Accountability”. In a demonstration: show me why it did that, show me a bad day, and tell me who is answerable.

What to put in front of it

Only your own material tells you anything. Bring four things from last week: a clean enquiry, a difficult one where the details were contradictory, the rate card you would really price from in the state it is really in, and a carrier invoice with several airway bills on it.

Then watch the last two, not the first. Two products that both say “quotation” can mean completely different work.