AI agents can do a defined part of freight forwarding work today: the document reading, field assembly, and drafting behind order entry, ISF preparation, vendor invoice posting, and party maintenance. They cannot do the parts that rest on a relationship or a commercial judgment, and no agent should submit anything to CBP or write to your TMS without a person approving it. The dividing line is testable rather than a matter of opinion. Work transfers to an agent when it repeats dozens of times a week, when its answer already sits in documents somebody sent you, and when a reviewer can confirm the finished output in seconds. At one small US import forwarder running these agents in production, the jobs that pass those three tests accounted for about 17.5 hours a week of keying: roughly 30 minutes per lot across 25 or more lots, and 15 minutes per ISF across about 20 filings. That is one operation, not a benchmark, and the tests below are what generalize.
A forwarding day is built out of a hundred four-minute jobs arriving out of order: a lot to create, an ISF due with its documents split across three emails, an air arrival, a load confirmation, six vendor invoices to check against expected cost, a new shipper to add as a party, a consignee address to correct in four places, a customer name now spelled two ways. Each one is small. None can be skipped, batched, or handed off without a check.
What counts as a back-office job function?
A back-office job function is one repeating unit of shipment paperwork that starts with a document arriving and ends with a record existing in the TMS. Order entry is one. ISF preparation is another. Vendor invoice posting, party creation, address maintenance, charge reconciliation, and duty classification review are each their own job, with their own trigger documents and their own definition of finished.
The distinction matters because “automate the back office” is not a project anybody can finish. A hundred distinct jobs behave as a hundred separate problems that happen to share a desk, and each needs its own documents understood before anything can be done with it. Software that promises the whole desk at once tends to deliver a document scanner and a screen of fields you still key in yourself.
Why does the day disappear if every task is small?
Because each four-minute task carries a context switch on either side of it. You hold a shipment in your head for ninety seconds, drop it, and load an unrelated one. Only the four minutes are visible on a timesheet.
The compounding cost is who is doing it. The people keying are the ones who know the accounts, and customer service and sales are the only two activities that grow a forwarding business. Hiring does not fix the ratio: another operator resets the ceiling but the keying still rises in a straight line with volume, and a new hire needs months before they know your accounts.
Which jobs transfer to an agent and which do not?
Score each job against three tests. It has to pass all three, and the third does most of the filtering.
- Frequency. Does it repeat at least a few dozen times a week? A job that runs twice a month will never repay the setup.
- Source of truth. Does the answer exist in a document somebody sent you, or in a colleague’s head? Reading a bill of lading is tractable. Knowing what your client will tolerate is not.
- Checkability. Can a reviewer put the finished output next to its source documents and know in seconds whether it is right? This is the test vendors skip.
Run those three across the back office of a small US import forwarder and the ranking comes out consistently. A shop whose volume sits in quoting or export will score its own list differently, which is the point of scoring rather than copying:
| Job | Frequency | Answer in the documents | Checkable in seconds | Verdict |
|---|---|---|---|---|
| Order entry / lot creation | 25+ per week | Yes, four documents per lot | Yes, field against source | Automate first |
| ISF preparation | ~20 per week | Yes, worksheet plus attachments | Yes, ten elements against source | Automate second |
| Vendor invoice posting | Every job, every vendor | Yes, invoice plus job charges | Yes, expected against billed | Automate third |
| Party and address maintenance | Continuous | Yes, from the bill itself | Yes, name against registry | Automate alongside |
| Rate negotiation | Per account, per season | No, it is a relationship | No | Keep with people |
| Demurrage write-off calls | Occasional | No, it is a client judgment | No | Keep with people |
What does each job actually involve?
| Job | Documents read | Output produced | By hand |
|---|---|---|---|
| Order entry | Bill of lading, commercial invoice, packing list, origin worksheet | TMS lot with parties, containers, PO lines and charge lines filled in | ~30 min per lot, ~12.5 hrs/week at 25 lots |
| ISF preparation | Origin worksheet plus attachments | Drafted ISF-10: seller, buyer, importer of record, consignee, manufacturer, ship-to party, country of origin, HTS-6, container stuffing location, consolidator | ~15 min per filing, ~5 hrs/week at 20 filings |
| Vendor invoice posting | Vendor invoice against the job’s expected charges | Posted charge line, plus a flag on every invoice that does not match | Continuous, usually deferred to month end |
| Party and address maintenance | The bill of lading itself | New party created or existing party corrected, duplicates folded together | Minutes each, unbounded over time |
| Duty classification review | Your own filed HTS codes against current tariff treatment | A flag on a classification that looks wrong or a duty rate that changed | Nobody schedules it |
Three of those deserve a note the table cannot carry.
ISF is the one with a clock. The filing is due 24 hours before cargo is laden aboard the vessel at the foreign port under 19 CFR Part 149, with eight of the ten required by that deadline and two, the container stuffing location and the consolidator, filable as late as 24 hours before arrival. The window is what makes it worth automating. When a worksheet lands late or an element is wrong, the rework falls on your team on the day the vessel is loading, and that is time nobody planned for.
Vendor invoices are the job most forwarders underestimate, because the value is not the posting, it is the mismatch. An invoice posted to the wrong shipment is money missing from a margin report that surfaces two months later.
Party maintenance is the quiet compounding one. Every duplicate party you prevent is a future reconciliation you never have to run, and a customer file your team still trusts in two years.
Air import and trucking sit behind these in the queue rather than beside them. Air paperwork behaves differently and runs on a tighter clock, and trucking dispatch carries less document structure to read. Both are real jobs worth automating and both are later work rather than shipped capability. A vendor telling you every lane is solved today is selling ahead of what the technology does.
Where does the human stay in the loop?
A person reviews and submits every external write, on every job in the list. The agent reads, assembles, drafts, and flags what looks wrong, then hands the work to someone with the source documents attached.
That boundary exists because the filing and the record carry your team’s name, and a wrong field costs more time to unwind than it cost to enter. So the agent runs fast where speed is safe, which is reading and assembling and drafting, and stops where a mistake would have to be chased back through a client. The reviewer’s job changes from producing the work to checking it, and checking a filled form against its source takes a fraction of the time building it did.
What do the reclaimed hours actually buy?
At the one forwarder we measure, the keying handed over came to about 17.5 hours a week, close to two working days, with the team still reviewing and approving every record. Those are first-party figures from a single live operation, consented for use here, not an industry estimate.
What that time converts into is the point. The operator who spent the morning creating lots spends it on the customer whose container is stuck and the account that has been waiting a week for a quote. Peak season stops meaning that everyone stays late and the growth work stops entirely.
TIO runs these as named agents against the TMS you already use rather than replacing it: an order entry agent that builds the lot, an ISF agent that drafts the filing, a finance agent that reconciles vendor invoices against job charges, and an HTS agent that checks classifications against current duty treatment. Party and address maintenance is not a separate agent; it runs inside the order entry and ISF agents, because that is where the names arrive. Each agent reads your own inbox and stages a draft for a person to approve. The fuller roster and the figures behind it are in our guide to the AI agent workforce for freight forwarders, and the job-by-job breakdown sits on our freight back-office automation page.
If your ranking comes out different from this one, we would like to hear why. Book a demo and we will walk your list, not ours.
Frequently asked questions
Can AI agents do freight forwarding work?
They can do a defined part of it. Document reading, field assembly, and drafting transfer well, which covers order entry, ISF preparation, vendor invoice posting, and party maintenance. Rate negotiation and client judgment calls do not transfer, because the answer lives in a relationship rather than in a document. At one small US import forwarder running these agents in production, the transferable jobs accounted for about 17.5 hours a week of keying. A person still reviews and submits every record.
Which back-office job should a freight forwarder automate first?
Order entry, in almost every case. It is the highest-volume keying job in a forwarding back office and the easiest to check. At one small US import forwarder running these agents in production, one ocean import lot takes about 30 minutes to build by hand across 25 or more lots a week, which is roughly 12.5 hours before any other job is counted. A reviewer can confirm a staged lot against its source documents in a fraction of that.
How do you decide if a freight forwarding task is worth automating?
Three tests, and a task has to pass all three. It repeats at least a few dozen times a week. The answer already exists in documents somebody sent you, rather than in a colleague's head. A reviewer can look at the finished output and know in seconds whether it is right. Rate negotiation and demurrage judgment calls fail the third test and should stay with people.
How many hours a week does back-office keying actually take?
About 17.5 hours a week at the US import forwarder we measure, close to two working days. That is 25 or more lots at about 30 minutes each plus about 20 ISF filings at about 15 minutes each. These are first-party figures from a live operation, consented for publication, not an industry estimate.
When is the ISF-10 due?
The filing is due 24 hours before the cargo is laden aboard the vessel at the foreign port, under 19 CFR Part 149. Eight of the ten elements are required by that deadline. The other two, the container stuffing location and the consolidator, can be filed as late as 24 hours before the vessel arrives. The tight window is why ISF prep is worth automating: there is no room for a second pass when a document arrives late.
Does an AI agent file the ISF or write to the TMS on its own?
No. Every agent stops before the external write. It reads, assembles, drafts, and flags what looks wrong, then hands the work to a person with the source documents attached. That person reviews and submits. Your team stays the filer and the approver on every record, so the review step is a design decision rather than a setting you can turn off.