Building effective AI agents for freight
The logistics teams that use AI successfully tend to share one habit: they start with simple, easy-to-understand workflows instead of heavy frameworks. There is a small set of these patterns, and each one fits freight neatly. Here is how a forwarder or importer can use every one of them to build a custom workflow inside their own operation.
- Know the split: a workflow runs the AI through fixed steps you set; an agent lets the AI pick its own steps. In freight, most wins are workflows.
- Start with the simplest version on your busiest task, and keep a person approving the final result.
- Only let the AI run free when the steps truly cannot be planned ahead. It costs more, and small mistakes can pile up.
- The patterns: augmented model, prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer, and autonomous agents.
Workflow or agent: know the difference
One difference matters more than any other. A workflow runs the AI through a fixed set of steps you define. It is predictable, cheap, and easy to check. An agent lets the AI decide its own steps as it goes. It is more flexible, but it costs more, runs slower, and small mistakes can pile up.
In freight that maps cleanly. Quoting, invoice audit, and RFQ comparison happen the same way every time and the stakes are high, so they fit workflows with a person approving the result. A true agent earns its place only where you cannot map the steps in advance, like chasing down a shipment that has gone sideways. The rule: build the simplest thing that works, and add complexity only when it clearly makes things better.
The patterns at a glance
Seven patterns, from the simplest setup to a full agent. Most freight teams only need the first few. The rest are there for when a job really calls for them.
| Pattern | Type | What it is | Freight example |
|---|---|---|---|
| Augmented model | Building block | An AI hooked up to your data, systems, and past jobs. | Drafts a quote from an email; answers cost questions from your own data. |
| Prompt chaining | Workflow | A fixed set of steps, each one checked before the next. | Quote build; invoice audit. |
| Routing | Workflow | Sort each request, then send it to the right handler. | Inbox triage; sort RFQ responses by mode. |
| Parallelization | Workflow | Run many checks at once, or run one check several times and take the majority. | Price 50 lanes at once; double-check HS codes. |
| Orchestrator-workers | Workflow | A lead AI splits the job on the fly; helpers do each part. | Multi-leg door-to-door quote; multi-forwarder tender. |
| Evaluator-optimizer | Workflow | One AI drafts, a second one checks it, and they go back and forth. | Quote or claim letter; customs valuation memo. |
| Autonomous agent | Agent | An AI that keeps working step by step until the job is done. | Exceptions and re-routing; landed-cost recovery. |
The building block: the augmented model
Every pattern sits on one foundation: an AI hooked up to your data (rate sheets, contracts, past shipments), your systems (your TMS, carrier and forwarder portals, tracking and customs), and a memory of past jobs (a customer's usual lanes, a supplier's quirks). On its own, this already does real work.
It reads an inbound quote email, retrieves the matching rate sheet and surcharges, and drafts the quote, remembering that this customer usually ships DDP on Shanghai to Los Angeles.
Connected to your PO and ERP data plus forwarder rate cards, it answers "what did we pay DSV on Ningbo to LA last quarter, and how does that compare to this tender" by retrieving and computing, not guessing.
Prompt chaining: quote build and invoice audit
Chaining breaks a job into a fixed set of steps, each one easier than the whole, with a check after each step that stops things if something is missing. Best when the job splits cleanly.
Step one pulls origin, destination, commodity, weight, and Incoterm from the request. A check confirms every field is there, and asks if one is missing. Step two pulls the right rate and surcharges. Step three works out the all-in price and margin. Step four writes a customer-ready quote. Each step is simple and easy to check.
Read the forwarder invoice, match it to the booked quote or contract, check whether it matched, flag every line that is over the agreed rate, then draft the dispute note. The same shape that powers our landed cost calculator.
Routing: triage the inbox
Routing sorts each request and sends it to the right handler, so each one can be sharp instead of trying to do everything at once.
Sort each inbound email into spot quote, booking, document request, claim, or arrival notice, and send it to the right workflow. Route simple quotes to a cheaper, faster AI and the messy, complex ones to a stronger one.
Route forwarder messages by type (rate update, invoice, arrival notice, exception), or sort RFQ responses by mode (air, FCL, LCL) into the right comparison logic.
Parallelization: many lanes, many lines, many checks
Run several pieces of AI work at the same time and combine the results. Two versions: split the work into independent parts, or take a vote by running the same check several times and going with the majority.
Splitting the work
An RFQ across 50 lanes prices every lane at the same time. A document pack (bill of lading, packing list, commercial invoice, certificate of origin) gets a separate checker per document, all running together.
A 200-line invoice is checked by splitting the lines across several checkers at once. Landed cost is worked out across hundreds of SKUs in one pass.
Taking a vote
An HS code is classified a few times over, and you go with the majority and flag any disagreement. The wrong code means penalties, so the extra checks are worth it.
Compliance screening (denied party, dual-use, whether a shipment is in or out of an IEEPA or Section 301 scope) run as several checks, with a vote that balances false alarms against misses.
Orchestrator-workers: multi-leg quotes and tenders
A lead AI breaks a job into parts, hands each one to a helper, and pulls the results back together. Unlike the version above, the parts are not fixed in advance; the lead decides them based on the request. Use it when you cannot predict the steps.
A door-to-door, multi-leg quote: the lead reads the request, decides which legs exist (origin haulage, export clearance, main freight, destination clearance, last mile), sends a helper to price each, and puts together one all-in quote. The number and type of legs depend on the shipment.
An RFQ or tender: the lead reads the scope, decides which forwarders to ask and which lanes to price, sends a helper to review each forwarder's response, and builds the bid-comparison table with a pick based on lowest landed cost, not lowest headline rate.
Evaluator-optimizer: drafts that critique themselves
One AI writes a draft, a second one reviews it and gives feedback, and they go back and forth. Use it when you have clear standards and a review genuinely improves the result.
Draft a customer quote or a claim letter, then have a second AI check it against the contract terms, completeness, and tone, going back and forth until it passes.
A customs valuation memo or a negotiation email: draft it, check it against your policy and the facts, then improve it. Or a routing search where the reviewer decides whether a cheaper, compliant option is worth another look.
Autonomous agents: exceptions and recovery
An agent is an AI that works step by step using your systems, checks what actually happened at each step, pauses for a person at key moments, and stops when the job is done or it hits a set limit. Save it for open-ended, multi-step jobs in a setting you trust.
An exceptions agent watches track-and-trace, spots a rollover or delay, looks into it using the carrier's system and other sailings, proposes or books a re-route, and hands off to a person before anything is committed.
A landed-cost-recovery agent keeps auditing every invoice across all your forwarders, opens disputes, and tracks what comes back. The IEEPA refund hunt is exactly this shape: scan entries, pull out the refundable duties, assemble the claim, file, and track, with a person approving before anything is submitted. See our tariff refund calculator for the starting point.
What matters most in freight
Three things run through all of these patterns: keep it simple, keep it transparent, and give the AI good tools to work with. Two of them matter even more in freight.
Transparency. Show the steps and the source of every number: which rate sheet, which contract clause, which invoice line. Freight runs on trust and a paper trail, and a number you cannot trace is a number you cannot bill on.
Good tools. Your rate store, TMS, and carrier systems are the tools the AI uses. Describe each one as clearly as you would for a new ops hire, and set the inputs up so the AI cannot quietly get it wrong: make it use a full lane code, a valid Incoterm, an exact reference. In practice, teams spend more time getting these tools right than wording the request, and the same will be true for you.
The build order
Most freight teams do not need a big, fully self-driving agent. They need the simple setup plus one or two step-by-step workflows on their busiest task, with a person approving the result. That is quoting for most forwarders and invoice audit for most importers. Ship that, see how it does, and move up to the fancier patterns only where the steps truly cannot be planned. Simple first, complexity only when it pays for itself.
Frequently asked questions
What is an AI agent in freight forwarding?
An AI connected to your systems (rate sheets, TMS, portals, the inbox) that can do a job like building a quote or checking an invoice using those systems. In practice, most useful freight setups are step-by-step workflows with checks, not fully self-driving agents.
What is the difference between a workflow and an agent?
A workflow follows fixed steps you set, so it is predictable and easy to check. An agent decides its own steps, which is more flexible but costs more and makes more mistakes. Start with workflows; save agents for open-ended jobs.
Where should a forwarder or importer start?
Your busiest task: quoting for forwarders, invoice audit for importers. Use a simple AI setup plus a step-by-step workflow, with a person approving the result, then build out from there.
Do I need an AI framework to build this?
No. Start simple, since most of these patterns are small. Heavy frameworks hide what the AI is actually doing, which makes freight workflows harder to fix and trust.
Is it safe to let AI quote or file customs on its own?
Keep a person on any step that commits you (sending a quote, booking, filing an entry). Let the AI gather, draft, and check; let a person approve. You keep the speed without the costly mistakes.
TallyHaul partners with a handful of forwarders and importers at a time, maps where your quotes, rate sheets, and invoices live, and layers these exact patterns on top. No rip and replace.