AI-native is not a chatbot. It is who does the work.
At Tide, large language models draft the quote, plan the load, and prepare the customs entry. People approve what matters. This page shows exactly where that line sits.
Four layers, one accountable loop.
Data feeds ground the models, the models draft the decisions, systems execute them, and a person signs the ones that matter. Nothing ships from the reasoning layer directly.
Grounded in the record.
The models never invent a rate, a rule, or a date. Every number they work with comes from a data feed or a document on the record, and every consequential output crosses a human gate before it becomes real. The division of labor is explicit:
What agents do alone
Read rate sheets, schedules, and tariff tables. Draft quotes and load plans. Assemble entry paperwork from the shipment's own documents. Watch every milestone and compute what a delay does downstream. None of this ships anywhere until a person or a standing rule says it does.
What always needs a signature
Prices customers see. Customs entries, which a licensed broker reviews and files under their own license. Exception decisions that cost money, like re-routing a delayed container or authorizing an expedite. The model proposes with its reasoning attached; the person disposes.
How errors surface
Errors surface as exceptions. When the model's confidence is low, when documents disagree, or when an output fails a deterministic check, the item routes to a person with the discrepancy highlighted. A wrong draft costs a review before it ever becomes a filing.
Three loops. They run on every shipment, all day.
A quote
- request parsed
- 3 routings priced
- margin checked
- draft assembled
- approved and senthuman
An exception
- rolled sailing detected
- downstream impact computed
- replan proposed
- option pickedhuman
- record updated
A customs entry
- documents extracted
- HS proposed with confidence
- broker reviewhuman
- entry filed
- release tracked
Models are only as good as what they read.
The model questions, answered straight.
The questions technical buyers actually ask about an AI-native forwarder.
What does AI-native mean, concretely?
It means the default worker on a shipment is a model. Agents draft the quote, plan the load, assemble the entry, and watch the milestones; people review, approve, and handle judgment calls. Software does the work here, and people supervise it. That is the whole product.
What do the agents do without a human in the loop?
They read and draft. They ingest rate sheets, schedules, tariff tables, and shipment documents, compute routings and landed costs, and prepare paperwork. They also track milestones and flag exceptions continuously. Anything that binds money or a legal filing, a customer-facing price, a booking, a customs entry, crosses a human gate or an explicit standing rule you have approved first.
Whose models does Tide use, and is my data used to train them?
Tide runs on commercial frontier models behind a provider-agnostic layer, paired with deterministic pricing and validation logic that does not hallucinate because it does not generate. Customer shipment data is used to serve that customer's shipments. It is not used to train foundation models, and account data is segregated per customer.
What happens when the model is wrong?
The failure mode is a flagged draft that stops for review. Low confidence, disagreeing documents, or a failed deterministic check all route the item to a person with the discrepancy highlighted. Because consequential outputs are gated, a wrong draft costs a human a review. The error surfaces in the exception queue with the model's reasoning attached, which is also how the system gets audited.
Can a licensed broker really sign an AI-drafted entry?
Yes, the same way brokers have always signed entries prepared by staff. The agents assemble the entry from the shipment's own documents; a licensed customs broker reviews the classification and the numbers, and files under their own license. Accountability sits with a named person whose license is on the line, which is precisely why the review step is not optional.
Can we turn the automation down?
Yes. The human gate is configurable per customer and per action type: you can require sign-off on everything, set value thresholds under which standing rules apply, or let routine re-bookings run automatically once you trust them. Most teams start with everything gated and loosen it decision type by decision type.
See the agents work a real lane.
Bring one shipment. Watch the quote, the plan, and the paperwork happen while you are still on the call.