By Sagar Shankaran, Founder of CallSphere
Three bills, three clocks, three payers. Why general chatbots blend freight vocabulary, what tuning on your own documents fixes, and the error-rate math.
Key takeaways
You tried this in 2024. Somebody on the drayage desk pasted an ocean carrier's invoice into a chatbot and asked it to explain the charges, and back came a fluent, confident paragraph about the container incurring "detention fees for excessive dwell time." Except the charge was per diem, the payer was different, the clock was counted differently, and the dispute window had a hard edge the answer sailed past. Forty minutes on the phone with the terminal later, the company ate $1,700.
That was not a bad chatbot. That was a general-purpose model doing what they all do with trade vocabulary: picking the meaning most common on the open internet rather than the one that is correct on a pier in Long Beach.
In freight, "detention," "demurrage" and "per diem" are three separate bills, with three clocks, three issuers and often three payers. Demurrage is charged by the marine terminal or ocean carrier while your container still sits inside the terminal past its free days. Per diem — which half the industry also calls detention, which is where the confusion is born — is charged by the ocean carrier once the box is out of the terminal and not returned within its free days. Driver detention is a completely different animal: the truck waiting at the consignee's dock past two free hours, billed by the carrier to you, and passed to your customer only if you have an accessorial approval.
Mix them up and you bill the wrong party, miss a dispute window, or eat a charge you could have passed through. A general model mixes them up constantly, because in ordinary English they all mean "waiting."
Which is the whole argument for the category that became distinct in 2026. A vertical model is one trained on a single trade's own documents and vocabulary, so that it reads "class 92.5," "TONU," "per diem" and "34 degrees continuous" the way a fifteen-year dispatcher reads them, rather than the way a dictionary does.
Take one import container discharged at a Southern California terminal. Four free days at the terminal before demurrage starts. Five free days on the box after gate-out before per diem starts. A chassis on it that may or may not be yours, depending on whether you did a chassis split. The 7501 entry summary cleared on day two. The receiver's appointment is three days out.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for logistics in your browser — 60 seconds, no signup.
Miss the appointment and you now have a decision: is the cheaper mistake another day inside the terminal, or gating the box out to a yard and starting the per diem clock? That is a real arithmetic question with a real answer, and it changes with the tariff. A system that thinks the two charges are the same word cannot even frame the question, let alone answer it.
Demurrage is the loud example. The quiet ones cost more in aggregate, because they happen forty times a day:
flowchart TD
A["Ocean carrier invoice hits the drayage inbox"] --> B{"Which clock started?"}
B -->|"Box still inside the terminal"| C["Demurrage: terminal free days"]
B -->|"Box out, not yet returned"| D["Per diem: ocean carrier free days"]
B -->|"Driver waiting at the consignee"| E["Detention: 2 free hours, then hourly"]
C --> F["Match to vessel discharge and gate-out records"]
D --> F
E --> F
F --> G["Bill the right party, or dispute inside the tariff window"]
Vertical AI became its own category this year for an unglamorous reason: general models miss the vocabulary, units, abbreviations and edge cases a trade takes for granted, and no amount of clever instructions fully fixes it. Tuning does two things instructions do not.
First, it teaches the model your documents' actual shapes: what a rate confirmation from your own TMS looks like, where the accessorial block sits, what your carriers write in the exceptions box, how your top ten customers phrase a tender email. Second, it teaches the resolutions your people have already made — that when this particular consignee writes "detention," they mean driver detention and it is capped at four hours by the broker-shipper agreement; that a temperature written as "34F" without a mode means continuous on this lane.
You are not building a model. You hand a supplier a pile of your own real documents with the right answers attached, and get back something that reads your paperwork rather than paperwork in general. The test is simple: give it fifty invoices and fifty tender emails from last month and count how many it reads correctly.
The ocean carrier's statement arrives with 41 line items across 26 containers. The tuned reader splits them into demurrage, per diem and chassis rental, matches each to the gate-out and discharge dates, and produces three lists: charges that are correct and pass through with your markup, charges that are correct but were caused by the receiver's appointment window and belong in a conversation with the customer, and charges that look wrong — free days counted through a terminal closure, or per diem billed on a box returned the day before.
Your operations coordinator works the third list, which last month was eleven line items worth $4,300, inside the dispute window instead of three days after it closed. That is the whole change. Nobody's job disappeared; the disputable items stopped being invisible.
Illustrative numbers, stated so you can swap in your own. Suppose your desk processes 400 documents a week — ocean carrier invoices, LTL bills, rate confirmations, tender emails. Suppose a general reader misreads 5 percent of them in a way that costs money, and a tuned reader with human review on flagged items misreads 1 percent.
Still reading? Stop comparing — try CallSphere live.
See the logistics AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
| Measure | General model | Tuned on your freight documents |
|---|---|---|
| Documents per week | 400 | 400 |
| Costly misreads | 5% = 20 | 1% = 4 |
| Average cost per misread | $180 | $180 |
| Weekly cost | $3,600 | $720 |
| Annual cost, 50 weeks | $180,000 | $36,000 |
The gap is $144,000 a year on assumptions you should test rather than trust. Note the shape of the average: it mixes $40 LTL reclass adjustments with the occasional $1,700 per-diem week and one $6,000 rejected reefer load. Before you spend anything, pull last quarter's credits, reclass bills and eaten accessorials out of your accounting system and work out your real average. That number is the entire business case and it takes an afternoon to produce.
It will not know that a terminal changed its free-time rule last Tuesday. Tariffs move, and the current rule lives on a terminal's website, not inside any model. Whatever you build has to read the current tariff, not remember an old one.
It will not read the handwriting in the margin of a BOL reliably — "24 of 26 short, driver notified" in ballpoint at a loading dock defeats most readers, and that note is precisely the one that matters for a claim. It will not make the commercial call about eating a legitimate charge to keep a customer who does $400,000 a year with you. And it will not replace the person who calls the terminal and talks a charge down.
The first step is small: for two weeks, have your coordinator keep a tally sheet of every document the current system reads wrong, with the word or field it got wrong and what it cost. Twenty entries in and you will know exactly which part of your vocabulary is worth teaching.
You hand over documents, and you should ask the supplier in writing what happens to them, whether they train anything shared with other customers, and how you get them deleted. Plenty of workable setups use only your own documents for your own use. Get that answer before the first upload, not after.
Sometimes yes. If your problem is ten confusing terms, a well-written page of house rules attached to every request will fix most of it, and you should try that first because it is free. Tuning earns its keep when the confusion is in the shape of your documents — hundreds of layouts from dozens of carriers and terminals — rather than in a short list of words.
The ones that repeat and cost money when misread: ocean carrier statements, LTL invoices with NMFC lines, rate confirmations, and tender emails from your top twenty customers. Skip the one-off documents. A few hundred correctly labelled examples of a repeating form beats ten thousand random pages.
Yes, with different words. Your version of the demurrage problem is freight class and accessorial naming: detention versus layover versus TONU, and whether "class 70" on a customer's tender matches what the carrier's inspector will actually assign on the dock.
The same vocabulary problem shows up on the phone, and it is worse there because nobody spells anything. A carrier calling about "the Laredo load" and quoting a pro number needs to be understood the first time. CallSphere builds AI voice and chat agents for business phone lines and web chat that can be set up with your own load numbers, terminology and rules — answering, capturing details and booking follow-ups around the clock. It reads and speaks your trade's words; it does not audit your ocean carrier's tariff.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Four-week is 28 days and a GS-1932 will not clear a 36-inch door. What industry-tuned AI actually fixes at the rental counter and storage office in 2026.
Why general AI misreads pack sizes, buydowns and dyed diesel on c-store paperwork, what a trade-trained model fixes in the price book, and what stays human.
ROH, BAR, DQQB, roll-in shower: the lodging shorthand general AI mangles on the phone, what 2026 trade-trained models fixed, and what bad bookings really cost.
Why an AI model that “improves” a foreclosure notice costs a community weekly a republication and an affidavit — and what a newsroom-tuned model does instead.
Days open, DAs, recips, pencil shrink and price slides: the trade shorthand general AI gets wrong on dairy and cattle operations, and what training on it fixes.
AOS, (c)(9), Epstein credits, MSC receipt prefixes and two-surname names: the shorthand a general model gets wrong, and what industry-tuned AI actually fixes.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI