Where Logistics Companies Should Use AI (and Where They Should Not)

Logistics is a chain of promises: a pickup window, a dock slot, a temperature band, a customs code, a last-mile SMS. AI helps when it reads the documents, ranks the exceptions, and retrieves the SOP for the person who still owns the promise. It hurts when it invents an ETA, books a lane you cannot run, or hides a delay in fluent English. This article is a map of where logistics companies should use AI — and where they should not — for operators and CTOs in the US, UK, Europe and Australia.

DayDecoder is in Ahmedabad. We integrate AI into existing software, including warehousing and industrial-adjacent products, without publishing fake savings. For fleet-heavy operators, read AI for transport businesses first; this piece zooms out to 3PL, warehouse, and network problems. Industry context: manufacturing, IoT and warehousing.

A simple map of the chain

Documents are the unglamorous gold

Bills of lading, packing lists, commercial invoices, PODs, temperature PDFs, warehouse receipts. These arrive as email attachments and photos. A schema, a queue, and a human confirm is the same pattern as other SMB automations, with more stamps and coffee stains. OCR plus extraction beats a chatbot you paste the PDF into.

Customs and dangerous goods are high-regret. Use models to pre-fill; a licensed human or your existing broker process still signs. Do not let a completion pick a HS code that “looks right”.

Exceptions beat dashboards

You already have status codes. The work is the weird ones: “arrived but seal broken”, “consignee unknown”, “reefer in defrost”. Classify, retrieve the SOP (RAG on your playbooks — RAG explained), draft the customer update from events that exist. If the GPS never arrived, the prose must not pretend it did.

Control towers fail when they are another dashboard. They work when they are a ranked queue with an owner. AI’s job is ranking and drafting, not owning.

Warehouse

Cycle count photos, damaged-goods notes, ASN discrepancies. Slotting and labour planning are often MILP or heuristics in the WMS. Ask the WMS vendor what they already sell before you commission a science project. Vision on a live floor needs safety and works-council/union conversations, especially in the EU. “Just put cameras and GPT” is not a design.

IoT (temperature, dock occupancy) is a time-series problem first. Anomaly detection can be statistical. An LLM can summarise a night of alerts for the morning shift. It should not be the sensor.

Network design and “optimisation”

Network optimisation is operations research. If you do not have clean costs and constraints, a model will give you a confident wrong map. Use analysts and solvers. Use AI to clean the inputs (geocoding messy addresses, extracting capacities from PDFs) so the solver is less starved.

We will not quote a “15% mile reduction” we did not measure on your network. Anyone who does on a first call is selling a slide.

Build vs buy

TMS, WMS, and visibility vendors ship AI features. Use them when they sit on your data. Build when your customer portal, your exception policy, or your documents are the differentiator. Hybrid is normal. See custom vs buy.

A 90-day sequence

1. Count one document type and one exception type.
2. Spike extraction or classification on staging, with logs and a kill switch.
3. Review desk for two weeks. Measure error types, not vibes.
4. Production flag for a subset of accounts or warehouses.
5. Only then RAG on SOPs or a customer-facing draft.

Architecture matches AI in Laravel/Node: jobs, tenancy, no model in the hot sale/scan path.

DayDecoder: 10+ years, 103 customers. contact@daydecoder.com.

EDI, APIs and the unglamorous majority of the work

Many logistics “AI” delays are EDI mapping delays. If 214s and 210s are messy, a model will summarise mess. Budget an integration slice. We would rather map statuses correctly than generate a paragraph about a missing milestone. Once events are trustworthy, generation is cheap. Before that, it is fiction.

Carrier APIs rate-limit. Cache. Queue. Do not ask a completion to “try the API again” in a loop. Your worker does that with backoff, like any other backend job.

Multi-party data: shipper, 3PL, consignee

Who may see a rate, a POD photo, a customs file? The assistant must use the same ACLs as the portal. A 3PL chatbot that can retrieve another shipper’s invoice is a tenancy bug. Design roles first. Then retrieve. Then generate. The order is not negotiable because the demo looks nicer the other way.

Subcontracted carriers add another hop. If you do not have their event, do not let the model interpolate one. “No event yet” is an honourable status.

Temperature, pharma and high-regret freight

If you move controlled temperature or regulated goods, the first AI is still document and exception work, with tighter audit. Do not auto-close a temperature alert because a paragraph sounds calm. A human signs. Models can assemble the pack (logger file, door events, SOP citation). They do not certify the product is safe. If your quality team would not allow a junior to auto-close it, they should not allow a model to either.

How we would staff it

A dispatcher or warehouse supervisor as product owner, an engineer who has shipped integrations, and an eval set of real anonymised documents. A two-week spike, then a production slice. Offshore overlap with US/UK/EU/AU hours as in our offshore guide. We will not staff a twelve-person “innovation lab” to classify POD stamps.

If your actual bottleneck is master data (locations with three spellings), we will say so and offer to clean that first. It is not glamorous. It is why the second project works.

Visibility products vs your own assistant

The market is full of “end-to-end visibility” suites. Some of them are the right buy: they already speak carrier EDI, they already have a map, they already page when a container dwells. Adding a chatbot on top of a suite you do not have is backwards. Adding extraction on top of a suite you do have can be right, if their AI button is generic and your exception policy is not.

Ask vendors where the event store lives, how you export, and whether their assistant can see another customer’s freight in a shared environment. If they cannot answer tenancy, keep walking. The same question we ask of ourselves on Laravel apps.

If you build, keep the event store yours even if the model is a vendor. The store is the product. The sentences are a skin. When the vendor changes their model name next quarter, your milestones should still be queryable.

Ports, rails and partners you do not control

Dwell at a port, a rail ramp, or a customer DC is often a phone culture. Models cannot call the clerk in a way that replaces a relationship, and generated email that sounds like a lawyer can make the clerk slower. Use AI to prepare a pack for a human who already has the number. Measure dwell with the partner’s timestamps, not with a paragraph.

Where partners offer APIs, use them. Where they offer PDFs at 6 p.m., extraction is the feature. Where they offer nothing, your assistant should say so, not simulate a milestone. Trust is the scarce asset in logistics. Fluent fiction spends it.

A 3PL that serves retailers in the US and grocers in the UK will have different appointment cultures and different food-safety paperwork. Do not train one global tone that papered over that. Per-customer playbooks in RAG, with ACLs, beat one “logistics personality”. If a playbook is missing, refuse and route to a human. Missing is a data problem you can count: how many accounts have a playbook, how many do not. That number is a better OKR than “AI adoption”.

Insurance, claims, and OS&D files are another document pile. Extraction plus a pack for a claims handler is in-scope. A model that admits liability in a customer email is not. Drafts should be unable to send without a role that already could send. The same rule as refunds in a SaaS app, applied to freight money.

If you operate bonded warehouses or file to customs systems, your software change control may already be audited. Treat prompts and models as part of that change control: versions, approvals, rollbacks. “We tweaked the prompt on Friday” is not an audit trail. Logs are.

Handover checklist

You should leave a first logistics AI slice with: event dictionary (your codes vs carrier codes), ACL matrix for shipper/3PL/consignee, a folder of anonymised documents used in eval, the prompt and model versions, the kill switch, and the name of the operations owner who will look at dead letters. Without the event dictionary, the next engineer will reinvent statuses and the assistant will drift. Without the ACL matrix, someone will demo a HQ view on a laptop at a conference. Put the matrix in git, not in a slide. Add a one-page “when we refuse to auto-send” list so a night operator does not have to guess. That page is product, not decoration, and it should be linked from the tool itself.

If EDI is still messy, the checklist includes a mapping owner on your side, not only ours. Integrations fail when that name is blank. Fill it in before the second warehouse goes live. Copy the same pack for the second warehouse; do not assume the first warehouse’s playbook applies to a site with different dock rules. Sites lie in similar sentences and different timestamps. When the second site goes live, rerun the eval on that site’s documents; a model that read one warehouse’s stamps may stumble on another’s handwriting, and that is a data fact, not a reason to buy a larger model on the spot. Keep the eval size honest: ten documents is a smoke test, not a programme, and we will say so if that is all you have.

FAQ

Should we build a logistics GPT with all our tracking data in the prompt?

No. Tracking data is an API and a database. Retrieve the shipment, then generate a sentence. Dumping the company into a context window is expensive and leaky.

Can AI replace customer service for “where is my freight?”

It can draft from events. It cannot invent events. Start with authenticated tracking and a draft; auto-send only when the event types are boring and evaluated.

Is computer vision required?

Not for a first project. Documents and exceptions pay sooner. Vision is a specific warehouse or yard problem with hardware and safety attached.

What about carbon reporting?

If you already have a methodology, a model might help classify missing data. If you do not, AI will write a paragraph that looks like a report. Finance and sustainability owners should set the methodology first.

Typical cost band?

Indicative five figures for a focused extraction or exception slice when systems have APIs; more if EDI mappings are the real project. Not a quote. Bring volumes to a call.