Image: Southeastern Freight Lines truck by Indy beetle, licensed under CC0 1.0.
You approve a lane strategy on Tuesday, and by Tuesday afternoon an agent is on the phone with four carriers, naming numbers. The question worth asking is not whether it sounds human. It is what happens when a dispatcher says yes to something you would have refused.
A voice agent that negotiates is a voice agent that can commit you. Rate, equipment, pickup window, detention trigger: each one is a promise that becomes a rate confirmation, then an invoice, then a dispute six weeks later. The safeguards matter more than the voice does.
Whatever the agent says, you own
The legal question has already been tested, in a different industry and for a small sum. In Moffatt v. Air Canada, 2024 BCCRT 149, the airline argued it could not be held responsible for wrong information its website chatbot had given a customer about bereavement fares. British Columbia's Civil Resolution Tribunal called the suggestion that the chatbot was a separate legal entity a remarkable submission, held the airline responsible for everything on its own website, and ordered it to pay $650.88 in damages.
That is a consumer tribunal, not a freight court, and the decision binds nobody. The principle is still the one your general counsel will reach for: a company owns what its software tells the other party. In procurement the other party is a carrier with a rate confirmation in hand, and the sums are not $650.
So the design question is narrow. What, specifically, prevents the agent from agreeing to something you would not have agreed to? Four things, running at four different moments.
Layer one: the call that never gets placed
Before Miles dials, the carrier has to clear. Operating authority active, insurance on file and in force for the commodity, the number being dialled matching the carrier record rather than a recently changed contact, no open fraud flag. If any of that fails, there is no call at all. The refusal happens before the first word, not as an alert somebody reads on Thursday.
This is the layer people underrate, because a blocked call looks like nothing happening. Verisk CargoNet's analysis of 2025 put cargo theft and fraud losses across the United States and Canada at close to $725 million, roughly 60 percent above 2024, with confirmed theft incidents up 18 percent from 2,243 to 2,646 and the average value per theft reaching $273,990. The American Transportation Research Institute has tracked the method shifting underneath those totals: strategic theft, which works through fraud, identity deception and fictitious pickup rather than bolt cutters, grew from 2 percent of incidents in 2018 to 25 percent in 2023.
Strategic theft runs on a convincing conversation. An agent that dials quickly, negotiates well and never gets tired is precisely the surface that method wants to find. Vetting the carrier is not a compliance box ahead of the call. It is the first safeguard against the call.
Layer two: equipment settles before price does
The second block sits between qualification and the quote. Miles cannot record a number until equipment and dock time are agreed, and cannot accept equipment that fails to match the load. A 53-foot dry van requirement means a 53-foot dry van. A 48-foot trailer, a reefer running dry, or a flatbed the dispatcher plans to tarp is a different conversation, not a near miss.
The cost of getting this wrong is visible in the spot market. DAT's June 2026 figures put all-in dry van spot rates at $3.00 per mile and refrigerated at $3.39. On an 800-mile lane that 39-cent gap is around $312, which is what you overpay when a reefer is quoted against a van requirement. Run the mismatch the other way and the loss is worse: the carrier discovers at the dock that the load needs temperature control, the load does not move, and you rebook at whatever the market charges for same-day coverage.
Equipment belongs in the hard constraints, not in the set of things the model weighs against price. Miles asks about the trailer and the dock time in the opening line, before any figure has been said out loud.
Layer three: acceptance criteria, not just a rate ceiling
A price ceiling is the guardrail everyone builds first and the least interesting of the four. Miles works against a target, a soft ceiling and a hard ceiling, floored by a zero-base cost calculation, and a quote above the hard ceiling cannot be accepted no matter how the call is going.
Rate is not the only thing a phone call commits, though. Pickup time, the detention trigger, quick pay terms, the accessorial schedule and the second-stop assumption are all live on that call, and all of them reach accounts payable eventually. Acceptance criteria covers the whole set: every term the agent is allowed to concede carries a boundary, and the acceptance step does not fire if any one of them sits outside it. A carrier who takes your number but will not take detention starting at two hours has not accepted your load.
That detention clause is the one that comes back. What gets agreed in half a sentence on the phone is the standard Penny measures the invoice against months later, which is the whole argument in our post on what a detention charge has to prove. Terms conceded casually on the phone are terms you lose quietly on the invoice.
The award itself is atomic. Four carriers are called in parallel, one lock is taken, and the other three are closed out professionally. Two carriers can never both be told yes.
Layer four: reading the transcript back
Three gates that run before a commitment can only block what somebody anticipated. Natural language leaks around them. An agent that says "we can usually work with that" has not recorded anything, and a dispatcher can reasonably hear a commitment in it.
So the transcript gets audited once the call ends, against the award that came out of it. The audit hunts for what was spoken but never structured: an implied accessorial, a promise about a second stop, a lumper assumption, a pickup time nobody on your side can hold. Where it finds one, the award does not stand. It reverts to pending approval and reaches a person with the quoted line and its timestamp attached.
| Moment | Without the safeguard | With it |
|---|---|---|
| Before dialling | Agent negotiates well with a carrier that cannot legally haul the load | Vetting fails, the call is never placed |
| Opening the call | Rate agreed, trailer discovered at the dock | Equipment and dock time settle before the first number |
| Closing the price | Rate inside policy, detention and accessorials outside it | Acceptance blocked until every term clears, award taken atomically |
| After hang-up | A spoken half-promise shows up as an invoice line | Transcript audit pulls the award back to pending approval |
The market is asking for exactly this
The Transportation Pulse Report 2026 from Transporeon, a Trimble company, surveyed more than 230 shipper and carrier executives across the EU and US in August and September 2025. Two-thirds of shippers and more than half of carriers said AI's primary role is augmenting human decisions rather than replacing them, with most preferring a human in the loop. The same respondents named inconsistent data as their biggest obstacle to getting value from it.
Read that as a specification rather than as hesitancy. Buyers will accept an agent that negotiates, provided the irreversible step stays near a person. Four safeguards are how that person's involvement becomes rare instead of constant, and the data point about inconsistent data explains why: every gate above reads from your carrier record, your load record and your rate policy. A guardrail is only as firm as the field it checks.
Where to start
Take last month's spot awards and read them as a liability exercise. For each one, write down what a fast, persuasive agent could have agreed to that you would have refused. Then name which of the four layers would have stopped it, and be honest about the awards where the answer is none of them.
That list is your actual autonomy boundary, and it is usually narrower than the one in the slide deck. Book a walkthrough to hear Miles negotiate a live lane and watch each block fire in sequence.