Issue 10 · Automation · Jul 21, 2026

Automate the typing. Never the deciding.

The one back-office workflow I'd automate first if I ran a ten-truck fleet, and the three I wouldn't touch. The popular answers rest on numbers nobody ever measured.

Last week I promised you the one back-office workflow I’d automate first if I ran a ten-truck fleet, and the three I wouldn’t touch.

I had an answer in mind when I made that promise. It was the obvious one — rate confirmations. Every vendor in freight says the same thing: your people are retyping load details out of PDFs, it takes four to eight minutes a load, hand it to a machine.

Then I went looking for where that number came from.

Nobody has ever measured this

There is no independent time-motion study on rate-confirmation entry. None. Not from ATRI, not from FMCSA, not from any university transportation program, not from OOIDA.

Every “minutes per load” figure circulating in freight automation traces back to a vendor blog. When you open those posts, most of them are honest enough about it if you read closely. One document-automation company builds its entire case on a scenario it introduces with the words “consider a 10-person brokerage processing 800 loads per month” — the four-minutes-per-document figure inside that example is attributed to nothing. It’s an assumption dressed as a finding, and it has been quoted onward for years until it hardened into common knowledge.

The most-repeated statistic in the category is worse. “Three to five check calls per load” appears in dozens of vendor posts, conference decks and LinkedIn threads. I could not trace it to a single study, survey, or dataset. The clearest version I found presents it inside a paragraph that begins “a typical day looks like this” — an illustration, not a measurement.

I’m not saying these numbers are wrong. Four minutes to retype a rate con sounds about right to anyone who has done it. I’m saying nobody has checked, and the people telling you the number are the people selling the fix. That’s worth knowing before you spend money against it.

So I went looking for a number in this space that somebody actually measured.

The one number that was measured

The OOIDA Foundation ran a detention-time survey of its members. Twenty-seven questions, emailed to 18,788 members, 253 responses, 95% confidence, ±6% margin of error. The methodology is published. The instrument is described. It is, as far as I can tell, the only rigorous measurement of where a small carrier’s non-driving hours actually go.

The finding, verbatim:

“the average weekly wait time for loading is 7.2 hours, and the average weekly wait time for unloading is 7.1 hours, for a total of 14.3 hours.”

Fourteen point three hours a week, at the dock, waiting.

The rest of the survey is worse. Members on the 60-hour rule lose 24% of their possible compensated drive time to detention. Half lose one or two loads a week to it; another 13% lose three or more. And the part that should decide your automation priorities:

17% collect no detention pay at all. Of the operators who say they always attempt to collect, 9% never receive it.

One respondent, in his own words:

“Detention time is crucial to my operation. The time to load or offload affects me considerably. I am a one truck company. My next load depends on detention. If I’m held up, I can lose thousands of dollars for loads missed.”

Now hold those two things next to each other. The industry’s automation conversation is organized around a four-minute task that nobody has measured. The largest measured drain on a small carrier’s week is fourteen hours of waiting — and almost nobody sells software against it.

That’s because detention isn’t an intelligence problem. It’s a documentation problem. You don’t win a detention claim by being smart. You win it by having a timestamped record of when the truck arrived, when it was loaded, and when it left — and by having that record attached to the right load number, in the right format, before the broker’s clock runs out.

Which is the same capability you need to stop retyping rate cons.

So: automate document capture

Here is the answer, and it isn’t the flashy one.

Point a cheap model at your documents once, and let that single extraction feed three things:

The load entry — carrier, shipper, pickup, delivery, rate, load number, accessorials.
The invoice packet — the same fields, assembled with the POD, in the format your factor or your customer wants.
The detention record — arrival and departure timestamps, captured off the BOL and the POD, attached to the load, sitting in a sheet the day it happens rather than reconstructed from memory six weeks later when you’re arguing about it.

One extraction. Three uses. And critically:

The machine’s job here is transcription, not judgment. It is reading a number off a page and putting it in a box. When it gets that wrong, you get a correction. Nothing else in the back office has that property.

That’s the whole reason this is the right thing to automate first — not because it saves the most time (though it probably does), but because it’s the only one of these tasks where being wrong is cheap.

What it costs

A two-page rate confirmation runs roughly 3,000 input tokens and about 500 output tokens of structured JSON. At Haiku 4.5 pricing — $1 per million input tokens, $5 per million output — that’s about half a cent per document. Two hundred rate confirmations a month comes to roughly a dollar ten.

I want to be precise about the status of that figure: the token count is my estimate, not a measurement. I’m publishing the arithmetic so you can check it, and next week I’m running ten real rate confirmations through and publishing what it actually costs, field by field, including what the model gets wrong. If the estimate is off I’ll say so.

What to build

A Gmail filter that catches rate confirmations by sender or subject.
Google Apps Script (free, and it will call an API for you) that sends the attachment to Haiku and gets structured fields back.
A Google Sheet with a “pending” tab. The script writes there. It writes nowhere else.
A human clicks before anything moves to the TMS or the invoice.

For a seven-truck carrier at around 200 rate cons a month, that’s under $25 a month all-in.

If you don’t want to touch code: Parseur’s free tier handles 20 pages a month, and Make.com’s free tier will move the output into a sheet. That’s genuinely $0, and 20 pages is genuinely the ceiling — name it honestly and decide if it’s enough to prove the idea.

The approval step is not a policy you adopt. It’s a piece of architecture. Nothing posts until a person clicks, which means you cannot skip it on a bad Tuesday when you’re behind. That distinction is the entire thesis of this issue, expressed as a design choice.

One thing worth saying out loud about the vendors

While pricing this out, I checked what the packaged tools cost. Not one freight-specific document tool publishes a price. Rossum, Klippa, HubTran, Loadsure — all contact-sales. Meanwhile the generic parsers publish theirs plainly: Docparser starts at $39 a month, Nanonets at $100, Parseur has a free tier.

Draw the obvious conclusion. If a company won’t quote you on its website, it isn’t built to sell to a ten-truck fleet. You are not its customer; you’re a lead to be qualified out.

And the three I wouldn’t touch

The rate. The carrier pick. Any compliance answer.

There’s one test that separates these from the typing, and it’s worth memorizing:

Does a wrong output cost you a correction, or a claim?

Transcription errors get corrected. These three get invoiced, stolen, or cited.

1. The rate

I’ll be straight with you about the evidence here, because it’s the weakest of the three and I’d rather you hear the limits from me.

I could not find a single documented case of a broker or carrier losing quantified money to an AI quoting engine mispricing freight. Not in FreightWaves, not in FreightCaviar, not in Trucking Dive, not in litigation. No vendor pullback, no brokerage publicly reverting to human quoting. That absence might mean the tools are fine. It might mean nobody publishes their losses. I don’t know, and I’m not going to invent a cautionary tale to fill the gap. (If one burned you, hit reply — that’s this week’s question, and I mean it.)

What I can show you is that the companies building rate-negotiation agents won’t let the model set the price either.

Chain launched an autopilot booking agent in June. From their own announcement: start, target and max rates “pull directly from the broker’s TMS, so the agent knows where to open every negotiation and where the ceiling is.” When an offer lands outside those bounds, “the agent escalates to the assigned rep.” Their own framing of the boundary, to FreightWaves: “The goal isn’t to automate the really hard freight. It’s to pre-book the freight that’s already moving with carriers in your network.”

Read that as an operator. The agent is not pricing. It is executing inside a price a human set. That’s a materially different product from the one being marketed as “AI quoting,” and the distinction is the whole ballgame.

Then there’s the silence. C.H. Robinson receives about 11,000 pricing emails a day, serves roughly 2,268 truckload customers with automated email quotes, generates around 2,000 quote replies daily, and responds in just over two minutes. Those numbers are all published. The accuracy rate is not. No error rate, no human-review percentage, nothing. The largest automated quoter in North America publishes its volume and its speed and says nothing whatsoever about whether the quotes are right.

I’m not alleging the quotes are bad. I’m pointing at a verifiable absence: the metric that would matter most to you is the one nobody discloses.

And even a vendor will tell you the failure mode if you read their marketing carefully. Vooma, which sells auto-quoting, on its own category: without proper implementation, AI produces “inconsistent quotes, missed service requirements, and subpar customer service” — because “in freight, data sometimes lacks the density to shape consistent trends.” That’s an admission against interest, and it’s correct.

Now the arithmetic that makes this concrete. A mid-market brokerage runs roughly $189 of gross margin on a $1,912 load. The accuracy bar vendors themselves cite for granting an agent autonomy is 95%. Take that at face value: 95% accurate means one quote in twenty is wrong, against $189 of headroom. You do not need many bad twentieths to erase a good month.

One more thing, if you’re tempted to let a bot read inbound shipper email and quote off it. In 2023 a car dealership’s chatbot was talked into “selling” a Chevrolet Tahoe for one dollar, with the bot volunteering “that’s a legally binding offer – no takesies backsies.” The dealer refused and roughly 300 dealer sites were patched inside two days. Inbound email is untrusted input. A quoting bot that reads it is reading text written by someone with an interest in the number.

As a baseline: unilateral pricing mistakes are generally enforceable unless the error was obvious enough that the other side should have caught it. A rate 10% under market is a contract. A rate 90% under probably isn’t. The dangerous zone is exactly the one a machine lands in most often — wrong enough to hurt, not wrong enough to void.

And the ten-truck version of all this: you don’t have a TMS with start, target and max rate fields for an agent to read. So an LLM quoting on your behalf isn’t executing inside your strategy — it’s quoting from nothing. Let it draft the counteroffer email. You type the number.

2. The carrier pick

This one isn’t a judgment call anymore. It’s a legal exposure, and it changed on May 14, 2026.

In Montgomery v. Caribe Transport II, LLC, No. 24-1238, a unanimous Supreme Court, Justice Barrett writing, held that a negligent-hiring claim against a freight broker is not preempted by the FAAAA. The facts: Shawn Montgomery was severely and permanently injured when his tractor-trailer was struck by a truck hauling plastic pots for Caribe Transport II. C.H. Robinson had arranged the shipment. Montgomery alleged the broker knew or should have known from Caribe’s safety rating that hiring it was likely to result in a crash.

The sentence to remember, verbatim from the opinion:

“Here, requiring C.H. Robinson to exercise ordinary care in selecting a carrier ‘concerns’ motor vehicles—most obviously, the trucks that will transport the goods.”

And the holding, from the syllabus:

“A claim that one company negligently hired another to transport goods is not preempted by the FAAAA because States retain authority to regulate safety ‘with respect to motor vehicles’ under the Act.”

Here’s why this belongs in an automation article rather than a legal one.

Your carrier-vetting file is now a discovery target. If you are ever sued over a crash involving a carrier you selected, the question will be whether you exercised ordinary care in selecting it. A fully automated approval produces a record showing that no judgment was exercised at all — which is the precise thing the duty requires you to have exercised. The automation doesn’t just fail to help. It manufactures the exhibit.

And the fraud data says the machine can’t do this job anyway. Cargo theft losses reached roughly $725 million in 2025, up 60% year over year — while the number of incidents stayed essentially flat (3,594 versus 3,607). Flat incidents, soaring losses, average theft value up 36% to $273,990. That combination means one thing: theft got selective. These are not crowbars. This is deception.

The clearest illustration is a Manhattan District Attorney indictment: eight defendants, $4.49 million in cargo, six loads including $2.6 million in cigarettes and $432,000 in cheese. The method? The defendants used real, clean MC and DOT numbers belonging to legitimate carriers.

Run that through your automated vetting stack. Every database check returns PASS. The MC number is real. The authority is active. The safety rating is clean. The fraud lives at the identity layer — who is actually standing at the dock — which no API call reaches.

It cuts the other way too, and honesty requires saying so. Automated vetting platforms also flag legitimate owner-operators as suspicious; roughly 80% of owner-operator businesses have now run into one. So you get both failure modes: false negatives let thieves in, false positives lock honest carriers out. Both are resolved the same way — a human looking at the exception.

Which gives you a clean rule you can hand to your team today:

The machine may auto-DECLINE. It must never auto-ACCEPT.

An auto-decline is a filter; the cost of a mistake is a phone call. An auto-accept is an unsecured loan to a stranger — and since May, it’s also an exhibit.

3. Any compliance answer

Start with the text. 49 CFR 390.11:

“Whenever in part 325 of subchapter A or in this subchapter a duty is prescribed for a driver or a prohibition is imposed upon the driver, it shall be the duty of the motor carrier to require observance of such duty or prohibition.”

There is no provision anywhere in the FMCSRs that transfers that duty to a vendor, a consultant, or a model. If the answer is wrong, you’re the one cited. You can outsource the drafting. You cannot outsource the duty.

Now, how wrong do these things get? I have to be straight with you again: no study exists testing LLMs on FMCSA, HOS or DOT questions. I looked hard. Nobody has measured it. That’s a genuinely empty space.

What we do have is the legal domain, which is the closest analogue — a body of dense, citation-heavy regulatory text where being confidently wrong has consequences. The Stanford RegLab study published in the Journal of Legal Analysis found that on specific, verifiable questions about federal court cases, GPT-4 hallucinated 58% of the time. Ask about a court’s core holding and the rate exceeded 75%.

The finding that should worry an operator most, verbatim:

models “struggle to correct incorrect legal premises and often cannot recognize when they produce inaccurate information.”

It will not tell you it’s wrong. It has no mechanism for doubt. It will answer a question about your hours-of-service exemption in exactly the same confident register whether it’s right or inventing.

And if you’re about to say but my compliance tool is connected to the actual regulations — that was tested too. A follow-up study examined purpose-built, retrieval-grounded, paid legal research tools. They still hallucinated 17 to 33 percent of the time. Lexis+ AI was accurate on 65% of queries; Westlaw’s AI-Assisted Research, 42%. Retrieval narrows the problem. It does not close it. Hold that number up against any compliance-AI demo you’re shown.

We also know how the accountability question resolves, because it’s been litigated in a smaller setting. When Air Canada’s chatbot invented a bereavement fare, the airline argued the chatbot was “a separate legal entity responsible for its own actions.” The tribunal member called that “a remarkable submission” and rejected it. (That was a British Columbia tribunal — persuasive as principle, not binding as US precedent.) The damages were trivial. The principle isn’t: you own what your machine tells people.

And there’s a live example sitting in front of us this week, courtesy of the agency itself.

Last week’s issue covered the July 22 rule removing the federal CDL self-reporting requirement. Buried in that rule, FMCSA acknowledged that some states still require the report, declined to say which ones, and wrote, verbatim:

“However, FMCSA will not be compiling this list… Nothing in this rule absolves a CDL holder from having to comply with a State requirement if that requirement exists.”

Ask a general-purpose model after tomorrow whether a driver still has to self-report an out-of-state conviction, and it will tell you no. Confidently. And it will be wrong in an unknown number of states — a number the agency itself explicitly refused to publish.

That’s not a hallucination the model invented. It’s one the regulatory record set up. No amount of retrieval fixes it, because the missing information doesn’t exist in retrievable form. Somebody has to call a state licensing agency and ask.

We’re building that list, by the way. It’s slow precisely because FMCSA wouldn’t do it.

The line

Here’s where I landed, and it’s simpler than I expected when I started.

Anywhere the machine is copying something down, let it. Rate cons, BOLs, PODs, timestamps, addresses, load numbers. It’s fast, it costs half a cent, and when it’s wrong you fix a cell.

Anywhere it’s deciding something — what to charge, who to trust, what the rule says — your name is on the authority, not its. A wrong rate is invoiced. A wrong carrier is stolen freight and, since May, a deposition. A wrong compliance answer is a citation, and 390.11 says the duty was always yours.

The vendors selling you the second thing will tell you the first thing is too small to bother with. It isn’t. It’s the part that works.

Automate the typing. Never the deciding.

Issue #10 of Freight/Signal · back to all issues · subscribe to the Tuesday newsletter

Sources

Read this in full each Tuesday