How do you catch an AI agent's freight mistake before the customer does?
How do you catch an AI agent's freight mistake before the customer does?
Picture the third week after a food-equipment manufacturer starts letting an AI agent book its outbound LTL. An order comes in for four crated units to a restaurant group's new location. The customer master still holds the old ship-to address for that account, from a location that closed in the spring. The agent reads the order, requests quotes, books the lowest approved carrier and sends the BOL. Nothing about the booking is wrong given the data it had. The mistake is in the record it was handed. If nobody looks, the customer finds it when the crates arrive at an empty building.
There are three places that mistake could have been caught: when the order arrived (the ship-to address didn't match the delivery address in the PO email), at the approval step (a person reviewing the first loads to a new location), or at pickup (the carrier's dispatcher asking about a closed facility). The customer finding it is the fourth place, and the most expensive one.
You catch an AI agent's freight mistakes the same way you catch a coordinator's: by putting checks where mistakes become costly, sending anything outside the rules to a person, keeping every action on a record your team can read, and auditing the invoice before you pay. The difference is that an agent does the checks every time, at the same steps, and leaves a record of each one. What you have to decide is which checks matter, which actions need a person's approval, and who receives the exceptions.
Why "will it make mistakes?" is the wrong first question
Every freight operation makes mistakes. Coordinators copy the wrong address, forget an accessorial, book a carrier that isn't approved for hazmat, or miss a reply from a carrier that canceled. The useful question is where mistakes come from in your operation today, which ones reach the customer, and whether the new setup catches more of them earlier. An AI agent that follows the same rules on every load removes some kinds of mistakes, such as a forgotten follow-up, and leaves others, such as a wrong address in the ERP, exactly where they were.
Three things to settle before the first load
- What the agent may do without approval. Requesting quotes is low risk. Booking a carrier, changing a delivery date or paying an invoice above the rate are higher risk.
- Who receives exceptions. Each load needs an owner, and each exception type needs a destination: the load owner, a manager, customer service or finance.
- What "done right" looks like. Agree on the checks for the first lanes so the team can tell, after two or three weeks, whether the work matches what it would have done.
Let's review the checks you'd want on the first lanes. You don't need to prepare anything: Cartage maps the workflow with you on the call.
See it on your own loads. Wilson, by Cartage, is an AI logistics agent that quotes, books, tracks, audits and pays for freight with your approved carriers.
Book a demo →What kinds of freight mistakes can an AI agent make?
An AI agent's freight mistakes fall into five groups: wrong inputs, wrong interpretation, carrier-side changes it didn't catch, timing misses and money errors. Knowing the group tells you where the check belongs and who should see the exception.
The five groups
| Group | Example | Where it starts | Who usually sees it first without a check |
|---|---|---|---|
| Wrong inputs | Old ship-to address in the customer master; wrong pallet count on the order | ERP, order email, warehouse | The carrier at pickup, or the customer at delivery |
| Wrong interpretation | An email says "ship Tuesday" and means the ready date, not the delivery date | Free-text instructions | The customer, when the freight is late |
| Carrier-side changes | The carrier accepted, then dropped the load; the driver changed the BOL count | The carrier or the dock | The receiver, or finance when the invoice arrives |
| Timing misses | An appointment wasn't booked; a pickup window was missed | Receiver rules, carrier dispatch | The customer's receiving team |
| Money errors | A reclass charge paid without checking; a duplicate invoice | Carrier billing | Finance, months later, or nobody |
Input mistakes are the largest group
Most problems on a load trace back to data that was wrong before anyone booked anything: an address, a weight, a dock hour, a freight class, a receiver who changed its appointment rules. An agent that reads the order faithfully will book faithfully on bad data. That's why the most valuable checks run before quotes go out: does the ship-to match the PO, does the weight make sense for the pallet count, does the receiver's record say it needs an appointment, does the order have every field the mode needs.
Interpretation mistakes need a rule for ambiguity
When an instruction can be read two ways, the safe behavior is to ask, not to pick one. "Deliver by Friday" from a customer whose dock closes at noon on Fridays is a different load from "deliver Friday". An agent should flag ambiguity to the person who owns the order and wait for an answer before booking, and the rule for which fields require confirmation should be written down.
What an AI "making something up" would look like in freight
The specific risk people worry about with AI is that it fills a gap with something plausible but false: an address that looks right, a carrier name that sounds real, a rate nobody quoted. In freight execution, three design rules make that kind of mistake hard to make and easy to catch:
- Addresses and references come from the order, the ERP or a person, never from the agent's own guess. If a field is missing, the agent asks.
- Carriers come from the approved list. A carrier that isn't on the list can't be booked, however it got into the conversation.
- Prices come from quotes and rate confirmations on the record. A rate that has no quote behind it is an exception, not a booking.
The article on what a freight AI can do before it asks for approval goes deeper on these limits.
Carrier-side changes are where silence costs the most
Many expensive mistakes aren't made by whoever booked the load. They happen after: a carrier drops the load the afternoon before pickup, the driver writes a different piece count on the BOL, the delivery slips a day. If nobody notices, the customer does. An agent that follows up on every load and compares each carrier message against the booking turns these into exceptions the team sees first.
Which checks belong before booking, before pickup and before payment?
Each stage of a load has its own checks, and each check has a clear owner when it fails. The goal is not to check everything twice. It's to check the few things at each stage that would be expensive to discover later.
A checkpoint map
| Stage | What to check | What it prevents | If the check fails |
|---|---|---|---|
| Order received | Ship-to matches PO and customer record; required fields present for the mode | Wrong destination, requotes | Ask the order owner before quoting |
| Quotes requested | Same details sent to every carrier; only approved carriers asked | Quotes that can't be compared; unapproved carriers in the mix | Fix the request; drop the carrier |
| Booking | Carrier approved for this lane and freight; rate within limits; rate confirmation matches the quote | Wrong carrier, overpaying | Wait for a person's approval |
| Before pickup | Carrier confirmed the truck; pickup appointment set; BOL matches the order | Missed pickups, dry runs | Escalate to the load owner the day before |
| At pickup | BOL count and weight match; PRO number recorded | Reclass and reweigh surprises | Route the change to the load owner |
| In transit | Tracking updates against the delivery appointment | Late deliveries the customer discovers | Tell the team, then the customer, with a new ETA |
| Delivery | POD received, signed, with exceptions noted | Claims without evidence | Chase the POD; flag damage notes |
| Invoice | Invoice matches rate confirmation, BOL and POD | Paying for charges never agreed | Hold the invoice and ask the carrier why |
The checks that matter most at the start
If you only set up four checks in the first month, choose these:
- Destination check before quoting. The ship-to on the order against the PO and the customer record. This catches the most damaging single mistake: freight going to the wrong place.
- Approved carrier and limit check before booking. The carrier must be approved for the lane and the freight, and the rate inside the approval limit. Anything else waits for a person.
- Carrier confirmation the day before pickup. A confirmed truck, not just an accepted tender. Most same-day drops show up here first.
- Invoice against the load before payment. The amount, the accessorials, and any reweigh or reclass against the rate confirmation and the documents.
BOL changes deserve their own rule
When the driver changes the pallet count, weight, freight class, address, carrier or price on the BOL, the change should go to the load owner immediately, not sit in a scanned document. A changed count at pickup may mean a short shipment the customer will notice; a changed weight may mean a reweigh charge later. The rule is simple: any change to those fields after booking is an exception with an owner.
Don't add checks that no one reads
Every check needs a person who will act when it fails. A daily report of 200 "possible issues" will be ignored by Wednesday. Start with the checks above, send each failure to one named owner, and add more only when a mistake gets through that a new check would have caught.
What should go to a person, and what can run on rules?
In the first weeks, a person approves every booking and every action outside the rules. As the team sees the same kind of action done right, that action moves to rules, one type at a time. That's assisted execution: the agent does the work, a person approves the decisions that carry risk, and the line between the two moves only when the record shows it should.
A starting split
| Action | First weeks | Once proven |
|---|---|---|
| Reading orders and asking for missing details | Runs on its own | Runs on its own |
| Requesting quotes from approved carriers | Runs on its own | Runs on its own |
| Booking within the approval limit on a familiar lane | Person approves | Runs on rules |
| Booking above the limit or on a new lane | Person approves | Person approves |
| Rebooking after a carrier drops a load | Person approves | Runs on rules within the limit |
| Changing a delivery date or appointment | Person approves | Runs on rules if the customer's rules allow it |
| Telling the customer about a delay | Person sends or approves | Depends on the company's choice |
| Paying an invoice that matches the load | Person approves | Runs on rules |
| Paying an invoice over the rate | Person approves | Person approves |
Approval limits are the main control
An approval limit says how far the agent may go without asking: a maximum rate per lane, a maximum difference from the target rate, a maximum freight value, a list of lanes or customers that always need a person. Limits catch two kinds of mistake at once: a decision that's wrong, and a decision that's right but expensive enough that the company wants a person to own it. The guide to configurable approvals for an AI logistics agent covers how to set them.
Exceptions go to the load owner, not to a queue
An exception that sits in a shared queue is an exception the customer will find. Each load should have an owner, and exceptions on that load go to that person first, by the channel they already use. If the owner doesn't respond in a set time, the exception goes to a backup. The article on who should approve the next move when a shipment goes wrong covers the escalation order.
Customer messages are a separate decision
Some companies want the agent to tell customers about delays directly; others want every customer message to go through customer service. Neither is wrong. What matters is that the team knows before the customer does, so the message, whoever sends it, comes with a new delivery time and a plan.
How do you review an AI agent's work without redoing it?
Review the shipment record, not the work itself: what the agent saw, what it did, what it asked and who approved it. If reviewing a load means rebuilding it from email threads, the review will stop after the first week. A record with change history makes it a few minutes per load.
What a reviewable record contains
- The order as received, with the source (ERP, file, email).
- Every question asked about missing or unclear details, and the answer.
- The quotes requested, from whom, and what came back.
- The carrier booked, the reason, and who approved it if approval was needed.
- The rate confirmation, BOL, tracking updates, POD and invoice.
- Every change after booking, with who or what made it and when.
A review routine for the first lanes
| Frequency | What to review | Who | What you're looking for |
|---|---|---|---|
| Daily | Open exceptions and loads waiting for approval | Load owner | Nothing waiting longer than it should |
| Daily, first two weeks | Every booked load on the first lanes | Logistics lead | Would we have booked the same carrier at the same price? |
| Weekly | A sample of loads that ran without approval | Logistics lead | Rules applied as intended |
| Weekly | Invoices held and why | Finance or logistics | Real overcharges vs. rules that are too tight |
| Monthly | Mistakes found, by group and by who found them | Logistics manager | Are more mistakes caught before the customer? |
Measure who found the mistake
The single most useful number is the share of mistakes found by the team or the agent before the customer, the carrier or finance found them. Log each mistake with three facts: what went wrong, which group it belongs to (input, interpretation, carrier, timing, money) and who caught it. After a month, the log shows which checks are working and where to add one.
When a mistake gets through
Treat it like any operational miss. Read the record, find the step where it could have been caught, and decide whether the fix is a data correction (update the customer master), a rule (require confirmation for new ship-to addresses), an approval change (lower the limit on a lane) or a new check. Then record the fix so the same mistake is caught next time. The team should be able to see and edit the rules the agent follows; a rule nobody can read is one nobody can fix.
Widening scope
Add lanes, customers or action types once the review shows the current scope runs cleanly: few exceptions that turn out to be the agent's fault, no mistakes that reached a customer, and invoices that match. Each widening should start again with closer review for the new part.
How does Wilson, by Cartage, catch freight mistakes before the customer does?
Wilson, by Cartage, is an AI logistics agent that executes freight inside the company's rules and approval limits, asks before guessing, sends exceptions to the load owner and keeps every action on the shipment record. Traditional logistics software gives teams information. Wilson executes the work required to move shipments forward.
Wilson works as a free logistics department for manufacturers and distributors: what big shippers build over years, Wilson does from day one. Wilson quotes, books and tracks freight, and pays carriers. Each of its four parts contributes a different check.
The labor
The labor is where most checks run. Orders reach Wilson by email, through an ERP connection or as CSV or XLSX files. Wilson checks each one against what the mode needs. If a detail is missing or a date is impossible, it doesn't book: it holds the order as pending, asks the customer for the missing detail and tells the team. It never invents the missing value. It books only carriers the company has approved, within approval limits, and sends everything outside those limits to a person. It follows up with carriers to the final POD by phone and email until it gets an answer, and works with the team in the channels it uses, including Microsoft Teams, Slack, WhatsApp and text. When a carrier drops a load, a pickup is at risk or a driver changes the BOL, the load owner hears about it first, and the customer is told only if the company has configured that. The team doesn't chase carriers.
The data
The data makes review possible. Wilson keeps every order, load, quote, rate, document and contact on record as it works, connected to the company's ERP. Each load carries its change history: what came in, what Wilson asked, what it booked, who approved and what changed after. Wilson also keeps rules and memories, such as "a new ship-to address needs a person to confirm it" or "this receiver only takes appointments from 7 to 11 a.m.", added by the team, and the team can edit them. When the team keeps making the same change on single loads, Wilson asks whether it should become a rule; it never adopts one silently. Changes to address, freight class, carrier or price on the BOL are routed to the load owner. Wilson replaces the TMS for freight execution and connects to the company's ERP. Wilson works with SAP, NetSuite, Microsoft Dynamics 365, Oracle, Infor, Epicor, Sage, Acumatica and other ERP environments, with the exact connection depending on the customer's data flow and scope.
The network
The network adds options when a carrier fails. The shipper's volume counts with every other Wilson customer's, which can bring better freight rates on its lanes, and the network includes carriers and brokers. The company still approves who hauls its freight, so a rebooking after a dropped load stays inside the approved list.
The payments
The payments are the last check before money moves. Wilson audits every freight invoice against the load, the rate confirmation, the BOL and the POD, including reweigh and reclass lines, and pays carriers from the Wilson Wallet. When an invoice is over the rate, Wilson asks the carrier why before paying. Invoices over the rate or outside the rules wait for a person.
Where each mistake group is caught
| Mistake group | Where Wilson checks | Who sees the exception |
|---|---|---|
| Wrong inputs | When the order arrives, before quotes | The order owner |
| Wrong interpretation | When an instruction is ambiguous | The order owner, who confirms |
| Carrier-side changes | Carrier follow-ups, BOL and tracking against the booking | The load owner |
| Timing misses | Appointment and pickup confirmation before the day | The load owner |
| Money errors | Invoice audit before payment | The person who approves held invoices |
Wilson starts with assisted execution. The company sets approval limits, and routine actions can run automatically within those rules as the workflow is proven. It handles LTL, FTL, ocean, drayage, parcel and cross-border freight. It is not a freight broker or a customs broker.
Pricing
Wilson is free on Wilson's network: the shipper pays the freight rate and nothing on top. When a shipper keeps its own carriers, the fee is 2% to 0.5% of what Wilson pays them, falling with volume. There are no seat fees, no setup cost and no contract. Timing depends on scope: most teams start on a few lanes and expand from there.
FAQs
How would we catch a Wilson mistake before our customer does?
Wilson, by Cartage, an AI logistics agent, works inside the company's approval limits and starts with assisted execution, so a person approves bookings and anything outside the rules in the first weeks. Exceptions such as missing details, dropped loads, BOL changes and late pickups go to the load owner as they happen, every action is kept on the shipment record with change history, and every invoice is audited against the load before payment.
What is the most common source of freight mistakes?
Wrong or missing inputs: an outdated address, a pallet count or weight that changed, missing accessorials, or receiver rules nobody recorded. These cause problems whether a person or software books the load, so the most valuable checks run before quotes go out.
Should a person approve every load an AI agent books?
At the start, yes, on the first lanes. Once the team has reviewed enough loads to trust a type of action, that action can run on rules within approval limits. High-value loads, new lanes and rates above the limit can stay with a person permanently.
Can an AI agent invent a carrier or an address?
It shouldn't be able to. Addresses and references should come from the order, the ERP or a person; carriers from the approved list; prices from quotes on the record. A setup that enforces those rules turns a guess into an exception. Wilson never invents a missing value: it holds the order as pending and asks.
What should happen when a driver changes the BOL at pickup?
The change should go to the load owner right away, with the original and new counts or weights. Then the invoice should be checked against the booking and the signed documents before any reweigh or reclass charge is paid.
How much time does reviewing an AI agent's work take?
With a shipment record that shows every action and change, a load review takes minutes. Plan for closer daily review in the first two weeks on the first lanes, then a weekly sample.
What happens to an invoice that doesn't match the load?
It should be held before payment. The invoice is compared with the rate confirmation, the BOL and the POD, and any charge that isn't backed by them goes back to the carrier with a question. Invoices over the rate or outside the rules wait for a person to approve them.
When should an AI agent's scope be widened?
When the record shows the current lanes run cleanly: few exceptions caused by the agent, no mistakes that reached a customer and invoices that match. Widen one lane, customer or action type at a time, with closer review at the start.
Who should receive exceptions?
The person who owns the load, through the channel they already use, with a backup if they don't respond in a set time. Shared queues with no owner let exceptions age until the customer calls.
Sources
- Cartage, Wilson product FAQ and product scope, October 2026: https://cartage.ai
- Cartage, Wilson product pages (labor, data, network, payments), November 2026: https://cartage.ai/data
Let's review the checks you'd want on the first lanes. You don't need to prepare anything: Cartage maps the workflow with you on the call.
