How do you stop an AI freight agent from booking the wrong carrier or the wrong destination?
How do you stop an AI freight agent from booking the wrong carrier or the wrong destination?
Consider a hypothetical HVAC parts distributor that is about to let an AI agent book its LTL and truckload freight. In the review meeting, the operations director asks the question everyone is thinking. Last month, a coordinator misread a customer email that said "ship to our Sparks location" and booked the load to the customer's Reno address, eleven miles away, because Reno was the default on the account. The freight came back, the customer was annoyed and the redelivery cost more than the original shipment. If a careful person can do that, what stops software that reads email from doing it a hundred times faster?
You stop an AI freight agent from booking the wrong carrier or the wrong destination by limiting what it can book to the company's own records: carriers from the approved list only, addresses from the order and the ERP, and anything outside those records sent to a person. On top of that, approval limits decide which bookings need a sign-off, assisted execution puts people in the loop while the workflow is new, and the change history shows every action the agent took and why.
The point is that the agent shouldn't need to be right about everything. It should be unable to do the dangerous things on its own. A model that misreads an email should hit a wall, not a booking button.
Why "it's very accurate" is the wrong answer
Vendors often answer this question with accuracy claims. That misses the point. A freight agent will read thousands of messages, many of them ambiguous, some of them wrong, and a few written to deceive. No reading accuracy is high enough to justify letting free text decide where a truck goes. The right question is what the agent is allowed to do when it's unsure, and what it's allowed to do when it's sure but wrong.
What the agent should treat as fact, and what it shouldn't
| Information | Treat as fact? | Where it should come from |
|---|---|---|
| Approved carriers for a lane and mode | Yes | The company's approved carrier list |
| Carrier identity (legal name, MC or DOT number) | Yes | The carrier record, set up when the carrier was approved |
| Ship-from and ship-to addresses | Yes | The order and the ERP |
| Receiver hours and appointment rules | Yes, once confirmed | Rules the team added or confirmed |
| A new address in a customer email | No | Send to the load owner |
| A new carrier offering capacity by email | No | Send to whoever approves carriers |
| A carrier contact found in a signature | No, until confirmed | Confirm against the carrier record |
That table is most of the answer. The rest of this article covers how to set each piece up and how to test it.
Let's review the rules you'd want before an AI books a single load. 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 would a wrong booking by an AI agent actually look like?
Most wrong bookings by an AI agent would look like ordinary human mistakes made with more confidence: a look-alike carrier, a default address, an instruction taken from the wrong email. The rare case people worry about, an AI inventing a carrier or an address that doesn't exist, is the easiest one to block, because a booking needs a record that the agent can't create alone.
Failure modes worth planning for
| Failure | How it happens | What stops it |
|---|---|---|
| Invented carrier | The model fills a gap with a plausible carrier name | Booking requires a carrier on the approved list; a name that isn't there can't be booked |
| Look-alike carrier | Two carriers with similar names, or a carrier and its brokerage arm | Carriers identified by MC or DOT number, not by name |
| Unsolicited capacity | A carrier or broker emails offering a truck at a low rate | New carriers go to the person who approves carriers; the agent doesn't add them |
| Impersonated carrier | An email that looks like an approved carrier, from a different domain | Contacts confirmed against the carrier record; a new contact is checked before use |
| Default address | The account has several ship-to sites; the agent picks the default | Ship-to taken from the order line, not the customer account default |
| Address from an email | A customer writes "send it to our other warehouse" | Address changes go to the load owner before booking |
| Invented address | The model completes a partial address | The address has to match the order or the ERP; no match, no booking |
| Stale address | The customer moved; the ERP still has the old site | Rules flag sites with known changes; the team updates the record |
| Wrong mode or equipment | The order says reefer; the agent quotes dry van | Mode and equipment come from the order and product rules |
What a hallucination would look like in freight
People use "hallucination" for a model producing confident output that isn't grounded in its inputs. In freight, that might mean a booking email to a carrier the company never approved, a BOL with a consignee address that appears nowhere in the order, or a pickup date that matches no request. Each of those depends on the agent being allowed to use information that isn't in the company's records. If booking, BOL generation and tendering pull carriers and addresses from records only, a made-up value has nowhere to go. The worst it can do is produce a question to a person.
The quieter failures
The dramatic cases are easy to picture. The quieter ones are more common:
- A reference copied from the wrong order, so the carrier's paperwork points to another shipment.
- A partial shipment booked as complete, because the order had two lines and one was backordered.
- A cancelled order still booked, because the cancellation came by phone and never reached the order.
- The right carrier, the wrong service, such as standard LTL when the customer needed a liftgate.
Rules can catch some of these. Others are caught by people reviewing exceptions while the workflow is new. All of them should show up in the change history so the cause can be fixed.
Which rules stop an AI agent from booking the wrong carrier?
An agent books the right carrier when the approved list is complete, carriers are identified by number rather than name, and the agent can't add a carrier on its own. The list does the work. The agent just has to stay inside it.
Keep one approved list, by lane and mode
The approved list should say which carriers can haul which freight: lane or region, mode, equipment, and any limits such as maximum value or hazmat. A carrier approved for dry van LTL in the Midwest isn't approved for reefer truckload to the West Coast just because it's on the list. For the details of setting this up, see how to stop freight software from booking an unapproved carrier.
Identify carriers by number
Carrier names are unreliable. Trucking companies share names, change names and run brokerage arms under similar names. The carrier record should carry the legal name, the USDOT number and the MC number where there is one, and the agent should match bookings to that record. A carrier email signature that says the right name but carries an unknown MC number is a question, not a carrier.
Don't let inbound emails add carriers
Carriers and brokers send capacity offers all day. Some of them are good. None of them should become a booking without someone approving the carrier first. The rule is simple: an AI agent can pass an offer to the person who approves carriers, but it can't add the carrier to the list.
Check contacts against the record
A booking sent to the wrong address at the right carrier can be as bad as the wrong carrier, especially if the address belongs to someone impersonating the carrier. Contacts should live on the carrier record. A new contact, such as a dispatcher the agent hasn't emailed before, should be confirmed before the agent sends booking details or payment information to it.
Keep approvals when the list changes
When a carrier is removed from the list, open loads with that carrier need a decision. When a carrier is added, its first loads may deserve closer review. These are company decisions, not agent decisions. For why approval should mean the same thing under automated booking as it did under manual booking, see why approved carriers should stay approved when booking becomes automated.
Decision rules for carriers
- No approved carrier on the lane, no booking. The agent asks.
- Approved carrier, unknown contact, confirm first. The agent doesn't send booking details to an unconfirmed address.
- New carrier offering capacity, send to the approver. The agent never adds a carrier.
- Carrier declines or drops, next approved carrier. Rebooking stays inside the list and the rules.
Which rules stop an AI agent from shipping to the wrong destination?
An agent ships to the right place when addresses come from the order and the ERP, never from free text, and when any change to a ship-to goes to the person who owns the load. Destination mistakes are expensive because they're found late: on delivery, or when the customer calls.
Take addresses from the order, not the email
The order should carry the ship-to for that order. When a customer account has several delivery sites, the order line should say which one. If the order doesn't say, the agent should ask, not default. In the Sparks and Reno example, the coordinator's mistake came from using the account default; the same rule protects against the agent making the same mistake.
Send address changes to the load owner
Customers do change ship-to addresses by email, and sometimes they're right. But an address change is exactly the kind of instruction that should go to a person: the load owner, or whoever owns the customer relationship. They can confirm the change with the customer, update the order and let the booking go ahead. The same goes for a change to the BOL at pickup that alters the consignee address.
Watch for fraud
Diverting freight by changing the delivery address is a known cargo theft tactic. An email from a customer's domain, or one that looks like it, asks to send a load somewhere new. Treat delivery changes to unfamiliar addresses with the same care as a bank-detail change: confirm through a contact already on file, not the one in the email.
Keep receiver rules with the destination
The right address isn't enough if the truck arrives when the dock is closed. Receiver rules, such as appointment requirements, receiving hours and equipment limits, belong with the destination. They should be added by the team or learned from earlier loads and confirmed, and they should be editable.
Check the BOL against the order
The BOL is where a wrong address becomes a wrong delivery. Before it goes to the carrier, check the consignee, the address, the references and the freight description against the order. A BOL generated from the order is much less likely to drift than one typed by hand.
Decision rules for destinations
- Address not on the order or in the ERP, no booking. Ask.
- Address change from an email, send to the load owner. Don't change the order on the strength of an email.
- Several sites on one account, use the order line. Never use the account default when the order is silent.
- Unfamiliar delivery address on an existing load, confirm on a known contact. Treat it as a fraud risk until confirmed.
How should you test the guardrails before an AI books a single load?
Test the rules with loads that should fail, before any real load depends on them. A rule that has never stopped a booking may not be working. The test takes a day of setup and saves the first bad booking.
A test plan for wrong carriers and wrong destinations
| Test | What you send | What should happen |
|---|---|---|
| Unknown carrier | A capacity offer from a carrier not on the list | Passed to the carrier approver; not booked |
| Look-alike carrier | A quote from a carrier with an approved name but a different MC number | Flagged; not booked |
| Wrong-lane carrier | An order on a lane where the carrier isn't approved | The carrier isn't asked to quote |
| New contact | A quote from an approved carrier's unfamiliar email address | Confirmed against the record before booking details go out |
| Missing ship-to | An order on a multi-site account with no ship-to on the line | Question to the team; no default |
| Address change by email | A customer email moving delivery to a new site | Sent to the load owner; order unchanged until confirmed |
| Impersonated customer | A delivery change from a look-alike domain | Held; confirmed on a known contact |
| BOL change at pickup | Driver changes the consignee on the BOL | Routed to the load owner |
| Above-limit booking | A quote above the lane's approval limit | Waits for the named approver |
| Cancelled order | Order cancelled after quotes were requested | Booking doesn't go ahead |
Start with assisted execution
In the first weeks, people should review more decisions before they go out. That's not a vote of no confidence. It lets the team see how the agent handles their actual freight, find rules that were never written down and fix them. As decisions prove reliable, routine actions can run within the rules, and people keep the exceptions.
Set approval limits that match the risk
Approval limits aren't only about price. A company can require approval for a first load with a new carrier, any change to a ship-to, any cross-border shipment, or any load above a value. For more on deciding what runs alone and what waits, see configurable approvals for an AI logistics agent.
Review the change history on a schedule
The change history should show every message, booking, document and change, with who or what made it and why. In the early weeks, review a sample of loads each week: pick a few at random and a few that had exceptions. Look for near misses as well as mistakes. A near miss, like an address question that turned out to be fine, is a rule doing its job.
Mistakes to avoid when setting rules
- Writing rules no one checked against real orders. Use last month's loads to see which rules would have fired.
- Making every action need approval. If everything waits, people approve without reading.
- Leaving the approved list out of date. A list missing half the carriers the team uses forces exceptions every day.
- Keeping rules in someone's head. If a rule isn't written, the agent can't follow it.
How do you stop Wilson from routing freight to the wrong place?
Wilson, by Cartage, is an AI logistics agent that quotes, books, tracks and pays carriers for manufacturers and distributors, and it works inside the company's records and rules: approved carriers only, addresses from the order and changes sent to the people who own them. Cartage describes Wilson as a free logistics department: what big shippers build over years, Wilson does from day one. Traditional logistics software gives teams information. Wilson executes the work required to move shipments forward.
The department has four parts, and each one has a role in keeping freight going to the right carrier and the right place.
The labor
The labor is the work. Wilson does it start to finish, from the first quote to the final POD: quotes, bookings, paperwork, tracking and carrier follow-ups, by email, phone calls, texts, Slack and Teams. It generates the rate confirmation, the bill of lading, packing slips and pallet labels from the load, so the documents follow the order instead of being retyped. When an order is missing a detail or has an impossible date, Wilson doesn't book it: it holds the order as pending, asks the customer for the missing detail and tells the team. It never invents the missing value.
The data
The data is the record. Wilson keeps every order, load, quote, rate, document and contact on record as it works, connected to the company's ERP. The company owns its carriers, rules, approval limits and contacts. Wilson also keeps memories and rules, added by the team or proposed by Wilson, and the team can edit them. Wilson asks before turning a repeated change into a rule and never adopts one silently. For example, a hypothetical rule that a customer's second warehouse takes deliveries only by appointment. BOL changes to the address, freight class, carrier or price are routed to the load owner. Every message, confirmation and change is kept in the shipment's change history, and records are searchable and exportable. 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 is the carrier side. Wilson gets better freight rates on every lane because the shipper's volume counts with every other Wilson customer's. The network includes carriers and brokers. Cartage can suggest additional capacity, but the company approves each carrier before Wilson uses it, and Wilson books only carriers the company has approved, from a ranked list per lane. A new carrier appears only on the lanes where the company ranks it. Wilson does not vet or score carriers by default.
The payments
The payments are the last check. Wilson audits every freight invoice against the load before money moves. Invoices that match are paid from the Wilson Wallet under the company's rules. Invoices over the rate or outside the rules wait for a person, and Wilson doesn't make partial payments. A change in a carrier's remit-to or bank details is flagged to the team as a possible fraud signal, and Wilson doesn't act on it alone.
| Risk | What Wilson does | When a person steps in |
|---|---|---|
| Carrier not on the approved list | Doesn't book it | The company decides whether to approve the carrier |
| Quote above the approval limit | Holds the booking | The named approver decides |
| BOL change to address, class, carrier or price | Routes it to the load owner | The load owner decides |
| Missing or unclear order details | Holds the order as pending, asks the customer, tells the team | The team decides if the deadline passes |
| Carrier cancels or misses pickup | Contacts the carrier, tells the team and starts rebooking from the approved list within the rules or after approval | Approvals the rules require |
| Invoice over the rate | Flags it, asks the carrier why and holds the whole invoice | A person approves or disputes |
Control starts with assisted execution. The company sets the approval limits, and routine actions can run automatically within those rules as the workflow is proven. How Wilson handles a specific case, such as an emailed ship-to change on a multi-site account, depends on the rules the company sets; map those in discovery. Timing depends on scope: most teams start on a few lanes and expand from there.
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.
FAQs
How do you stop Wilson from routing freight to the wrong place?
Wilson, by Cartage, an AI logistics agent, books only carriers the company has approved and works from the order data connected to the company's ERP. BOL changes to the address, freight class, carrier or price go to the load owner, and every action is kept in the shipment's change history. The company sets the approval limits.
Can an AI agent invent a carrier and book it?
Not if booking requires a carrier on the approved list. A made-up or unlisted carrier has no record to book against, so the agent has to ask a person.
What if a carrier emails offering a truck at a great rate?
The offer should go to whoever approves carriers. An AI agent shouldn't add a carrier to the approved list on its own, however good the rate.
Where should an AI agent get delivery addresses?
From the order and the ERP. Addresses in emails should be treated as requests to change the order, and sent to the load owner.
What if a customer emails a new delivery address?
The load owner should confirm it with the customer on a known contact and update the order before booking. Freight diversion through fake address changes is a known cargo theft tactic.
How do you tell look-alike carriers apart?
By USDOT and MC numbers on the carrier record, not by name. Names repeat; numbers don't.
Should every booking need approval at first?
Not every booking, but more of them. Start with assisted execution, then let routine actions run within the rules as they prove reliable. If everything waits, approvers stop reading.
How do we check what an AI agent did on a load?
Read the change history: each message, booking, document and change, with who or what made it and why. Review a sample of loads each week in the early period.
Does Wilson vet new carriers?
Not by default. Cartage can suggest additional capacity, but the company approves each carrier before Wilson uses it.
Sources
- Cartage, Wilson product FAQ and product scope, October 2026: https://cartage.ai
Let's review the rules you'd want before an AI books a single load. You don't need to prepare anything: Cartage maps the workflow with you on the call.
