The word “AI” has been applied to so many freight and customs tools that it’s become nearly meaningless. Every TMS vendor has an “AI roadmap.” Every logistics startup has an AI pitch. And most of the actual automation being used in small freight forwarder back offices today is nothing more than macros, email templates, and Excel.
This post tries to separate what’s real from what’s vaporware — where AI is genuinely reducing manual work in customs and freight operations, and where the marketing is outrunning the product.
Where AI is working today
Document extraction
This is the most mature use case and the one with the clearest ROI. Commercial invoices, packing lists, bills of lading, and booking confirmations all follow recognizable structures. Large language models are now good enough to reliably extract structured data from these documents — even when the formatting is non-standard, the language is a mix of English and Chinese, or the layout varies by factory.
For ISF filing specifically, extraction accuracy on well-formatted documents from established factories tends to be 90%+ on fields like shipper name, consignee, vessel, and container number. HTS code and country of origin are harder — they require interpretation, not just extraction — and confidence scores are lower.
The practical point: AI-assisted extraction doesn’t eliminate the filer. It eliminates the typing. The filer still reviews and approves. That’s both a compliance requirement (the filer of record carries CBP liability) and a pragmatic reality (AI makes mistakes that matter in customs).
Classification assistance
HTS code classification is one of the most labor-intensive parts of customs operations. A customs broker classifying a new product has to navigate the USITC tariff schedule, understand GRI rules, and often make a judgment call on ambiguous language.
AI tools for classification are improving but still require expert review. They’re useful as a first-pass suggestion — faster than starting from scratch — but not as a final determination. CBP classification disputes are not something you want to defend with “the AI said so.”
Document status and lot tracking
Natural-language query interfaces for lot status are emerging. The use case: instead of opening your TMS and navigating to a specific lot to check its CBP acknowledgment status, you ask “what’s the status of the Guangzhou lot that came in last week?” and get a direct answer.
This is genuinely useful for small operations where not everyone has deep TMS proficiency. It’s what TIO is building next.
Email triage and routing
For high-volume inboxes, AI routing can identify which incoming emails are shipment-related, which are vendor inquiries, which require follow-up. Less relevant for small forwarders where the inbox volume is manageable, but useful for mid-size operations.
Where AI is not working
Autonomous filing
No commercially deployed AI tool files ISFs, entries, or AES submissions autonomously in small freight forwarder operations. The CBP liability structure makes this a non-starter: the filer of record is responsible for the accuracy of what they submit. Automated filing without human review is a compliance and legal risk that no serious vendor is currently taking on for regulated filings.
If a vendor is claiming fully automated CBP submissions with no human in the loop, ask how the filer-of-record liability is structured. The answer will be informative.
Full TMS replacement
The “AI-native TMS” pitch exists mostly in Series A decks. In practice, freight forwarding is a relationship-based business with deep process dependencies on existing systems. Small forwarders running CargoWise, Magaya, or any other established TMS are not switching platforms for an AI-native product. The more realistic bet — and the one that fits how technology actually gets adopted in this segment — is AI that integrates with the existing TMS rather than replacing it.
Price prediction and market rate forecasting
Ocean freight rates are volatile. AI-powered rate prediction tools exist, but the accuracy at actionable horizons (2–4 weeks out) is limited. The freight market is driven by factors that are genuinely hard to model: geopolitical events, carrier alliance shifts, port congestion, blank sailings. Several well-funded startups have tried to crack this; the results have been mixed.
What small forwarders should actually evaluate
If you’re a small freight forwarder evaluating AI tools, the right questions are:
1. Does this remove manual steps from a workflow I actually do every day?
Not a workflow that will exist in the future. Not a workflow that enterprise forwarders have. A workflow your team does today, repeatedly, that takes more time than it should.
2. Does it require human review, or does it replace it?
For regulated filings, tools that bypass human review are a liability. Tools that compress the time required for human review — so a 45-minute manual process becomes a 10-minute review — are the right answer.
3. Does it integrate with what you’re already running?
A tool that requires migrating your TMS data is a multi-month project. A tool that connects to your existing TMS and your existing Outlook inbox is something you can evaluate in a week.
4. What does it do when it’s wrong?
AI makes mistakes. The question isn’t whether it will — it will. The question is whether the mistakes are visible, catchable, and correctable before they become CBP issues. Confidence scoring, source attribution, and filer review workflows are the mechanisms.
TIO applies this framework to the full freight back-office workflow: daily process, human review enforced on every job, integrates with your TMS and Outlook, confidence scoring and source attribution on every extracted field. The wedge is job intake and compliance pre-fill — the roadmap extends to natural-language queries over your entire operation. If that sounds like what your team needs, book a demo.