AI for logistics companies: use cases and first steps.
Where AI pays off in freight, 3PL and distribution today, where a language model is the wrong tool, how to measure it, and a 90-day plan for the first project.
For mid-size freight forwarders, 3PLs and distributors, the first AI system worth building usually isn’t route optimisation or a forecasting model. It’s a system that reads shipping documents and customer emails, writes the result into your TMS or WMS, and answers “where is my shipment?” Language models are good at that work. Optimisation and forecasting need different tools, covered below.
This guide covers the use cases that pay off, a table of which kind of AI fits which logistics problem, how to measure the result, what your TMS integration and data need, and a 90-day first-project plan.
What are the main AI use cases in logistics?
Oliver Wyman’s review of 25 real-world logistics use cases groups them into six categories: back-office automation (inbound email, pricing, scheduling, shipment tracking), customer interaction, expert support such as routing and network flow, predictive analytics for ETAs and visibility, real-time data processing with computer vision, and robotics. It puts AI penetration in logistics at around 50%, behind technology and finance.
Those categories use different technology. Only the first two are mainly language-model work: reading text and acting on it. The rest are optimisation, forecasting, vision and robotics. That split decides what to build first. The Logistics Trend Radar from DHL (September 2026) says AI is starting to plan, coordinate and act across supply chains alongside people, so limits on what an agent may do matter more.
Where AI pays off for forwarders and 3PLs today
1. Reading shipping documents
Bills of lading, commercial invoices, packing lists, customs declarations and proofs of delivery arrive as PDFs, scans and photos, in every layout imaginable. A document AI system extracts the fields, checks them against the booking and each other, and writes them to your system. It flags anything it isn’t sure of for a person. The checking is what makes it safe: totals that don’t add up, weights that don’t match the booking, container numbers that fail a format check, and missing signatures are caught before they become a delay.
2. Quotes from inbound emails
Requests for quotes arrive as free-text emails: origin, destination, dimensions, dates, special handling, often half-specified. An AI agent can read the request and ask the customer for what’s missing. It then pulls rates from your tariffs and drafts the quote for approval, or sends it within limits you set. The language model reads and writes. The rate comes from your rate tables through code, never from the model’s memory. Response time is often what wins a load, so this is usually the use case with the clearest revenue effect.
3. “Where is my shipment?”
A large share of customer messages are status questions. An assistant on email, web chat or WhatsApp can look up the shipment and answer in the customer’s language. It hands over to a person when the answer is bad news or the customer is upset, with the conversation attached. The same design runs in a production WhatsApp sales agent we describe in detail. Our guide to AI customer service covers what else to automate in support.
4. Exceptions
Delays, damaged goods, failed deliveries and missing paperwork trigger chains of emails and calls. An agent can spot an exception from tracking events (a carrier status message, an EDI 214, a missed milestone) or from a customer email, gather the facts, draft the notifications to the customer and carrier, and open the right case. A person decides what to do; the agent does the gathering and the typing.
5. Carrier communication and freight audit
Chasing confirmations, appointment slots and documents is repetitive, rule-bound work, a good fit for an agent with clear limits on what it can agree to. Matching carrier invoices against agreed rates, accessorials and proof of delivery is mostly reading and comparing: AI handles the reading, and your rules, enforced in code, do the comparing.
Where a language model is the wrong tool
Some of the most-discussed logistics problems don’t need a large language model at all, and are solved better without one. Use this table to match the problem to the tool before you buy or build anything:
| Problem | Right tool | Data needed | First-project risk |
|---|---|---|---|
| Reading bills of lading, invoices, packing lists, PODs | Document AI with a language model, plus validation rules | A few hundred real documents per type, including the worst scans; the booking record to check against | Wrong fields written to the TMS. Needs confidence checks and a review queue |
| Quote requests from email | LLM agent plus your rating logic | Tariffs and rate tables in structured form; past quotes and outcomes | A wrong or off-policy price sent. Needs approval limits enforced in code |
| Status questions and customer updates | LLM agent reading shipment and tracking data | Clean shipment status and customer-to-shipment mapping | Wrong answer on a stale status. The agent must show what it knows and when |
| Exceptions and claims | Rules to detect, LLM agent to gather facts and draft | Tracking events, carrier messages, past cases | Over-promising to a customer. A person decides the remedy |
| Route and load optimisation | Optimisation solver or routing engine | Orders, vehicle and driver constraints, travel times, time windows | Missing real-world constraints, so plans look good and fail on the road |
| Demand and volume forecasting | Time-series and machine-learning models | Years of clean history by lane, customer or SKU | Too little history, or history distorted by one-off events |
| ETA prediction and shipment visibility | Machine-learning models on tracking history, plus data feeds from carriers and devices | Tracking events, carrier and port data, telematics | Gaps in carrier data. A model can’t predict what no one reports |
| Warehouse inventory checks and damage detection | Computer vision on camera feeds | Labelled images from your own sites and lighting | Camera placement and lighting, not the model, decide accuracy |
| Warehouse robotics and slotting | Robotics and WMS optimisation | Layout, SKU velocity, WMS and equipment data | Heavy site changes and integration. Not a quick first project |
A language model still has a job next to the other tools. It can turn a planner’s request into solver inputs and explain a plan in plain words. It can summarise what changed in a forecast for the planning meeting, write the customer update when an ETA moves, and draft the claim from a damage detection and the paperwork. It is the reading and writing layer beside the specialist system, not a replacement for it. BDC’s guide for logistics firms lists the same split: forecasting, paperwork, tracking and warehousing are separate problems with separate tools.
Where should a person stay in charge?
- Anything that commits money or capacity. Booking a load, accepting a rate, approving a claim or releasing a payment starts as a draft that a person approves. Automate it later, within limits written in code, once the numbers support it.
- Carrier and broker identity. Freight fraud often arrives as a convincing email. The FMCSA says fraud and identity theft occur when someone uses another carrier’s USDOT number without authority or acts as an unregistered broker, and advises confirming phone numbers against the official SAFER record. An agent can run those checks and flag a mismatch. A person decides whether to trust the carrier.
- Customs classification and declarations. An agent can pull the data and propose a classification. A licensed customs broker or other qualified professional signs off, because the legal responsibility sits with them.
When the model is unavailable, a field fails validation or a source system is down, the right behaviour is a clean handover, not a guess. Our note on designing AI that fails gracefully covers how.
How do you measure whether it’s working?
Agree the numbers before building, and record two to four weeks of baseline from your current process so you have something to compare against. Four measures cover most logistics flows:
| Measure | How to count it | What it tells you |
|---|---|---|
| Touches per shipment | Manual actions (typing, re-keying, forwarding, chasing) between booking and delivery, sampled across a set of shipments | Whether work is removed or just moved |
| Quote turnaround | Minutes from the request arriving to a quote sent, median and slowest 10% | Whether response time improved where it wins loads |
| Exception handling time | From the first event or message to the customer being told and a case owner assigned | Whether problems are caught and communicated sooner |
| Document error rate | Wrong or missing fields per 100 documents, counted per field type, and the share a person had to correct | Whether the extraction and checks are accurate enough to trust |
Add quote win rate once the quote flow has run long enough to mean something. Count errors per field type rather than per document. A document that is right on 19 of 20 fields can still be the wrong container number. Re-run a fixed set of real documents and messages on every change. Our guide to measuring AI quality before shipping explains how.
BDC’s guide also warns that adoption can temporarily reduce productivity while systems are integrated and staff are trained.
How to choose the first project
Pick the flow that scores best on four questions:
- Is it high-volume? Hundreds of documents or messages a week give the system enough to pay back, and enough real examples to test on.
- Is it measurable? Touches per shipment, quote turnaround, exception handling time or document error rate. Agree the number before building.
- Is a mistake recoverable? Start where a person reviews the output, or where errors are easy to fix, before automating anything that commits money or capacity.
- Can it reach the data? The system is only as good as its access to your TMS, WMS, ERP and inboxes. If integration is blocked, pick another flow first.
For many forwarders and 3PLs, quote requests or document extraction come out on top: high-volume, measurable and easy to keep under human review.
The first system’s real job is to earn trust. Pick one where people can check its work, and let the numbers decide what to automate next.
How ready is your data, and how does it connect to your TMS?
Integration with an older TMS or WMS is usually the biggest piece of work, bigger than the AI. BDC lists older systems that don’t share data easily and operational data that isn’t collected consistently among the main obstacles, and advises assessing your data and gaps before choosing tools. Check these before you commit to a flow:
- Which route into the system exists? In order of preference: a documented API, EDI messages (load tenders, status messages and invoices) or a database view, and last of all scheduled file exports or a monitored mailbox. Each lower route works, but adds delay and failure modes.
- Read first, write later. Start with read-only access so the system can look up bookings and shipments. Add writes through a staging step that a person approves, then open direct writes field by field, with every change logged and reversible.
- Reference data. Customer and location codes, charge codes, units of measure, carrier IDs and Incoterms must match between documents and the TMS. If one customer exists under four names, fix that first.
- Rates in a machine-readable form. Quote automation fails when tariffs live in PDFs and in one person’s head.
- A sample archive. A few hundred past documents and emails per flow, with the correct outcome for each, including the worst ones.
You don’t need to replace your TMS. The AI layer reads from and writes to it, or reaches it through email and files where nothing better exists. See our services for how we scope the integration alongside the AI work.
What makes logistics AI hard
- Messy inputs. Scans, handwriting, stamps, and forms in several languages. Test on your worst documents, not your best, and test each language separately.
- Accuracy that compounds. One wrong container number causes a chain of problems downstream. Every extracted field needs a confidence check and a cross-check against data you already hold.
- Accountability. An agent that emails customers or carriers on your behalf needs limits enforced in code and a log of everything it sent. Our note on prompt injection explains why inbound emails must be treated as untrusted input. In freight, that includes emails that impersonate a carrier or customer.
Which rules apply to logistics AI?
Most of the rules that touch this work concern disclosure and trade documents. As of October 2026:
- Telling customers they’re talking to AI. In the EU, Article 50 of the AI Act requires AI systems that interact directly with people to be designed so those people are informed they’re dealing with an AI, unless that’s obvious from the context. It applies from 2 August 2026. Our EU AI Act guide for builders explains the rest.
- Electronic trade documents. In the UK, the Electronic Trade Documents Act 2023 defines when information in electronic form is an “electronic trade document”: where a reliable system is used. In practice, electronic and paper documents will both flow through your system for years. The ICC Digital Standards Initiative shows an electronic bill of lading adoption rate of 12.8%, sourced to the Digital Container Shipping Association (DCSA), up from 5% in 2024 (the page gives no date for the newer figure). A 2024 industry survey found 49.2% of respondents used electronic bills of lading, exclusively or alongside paper, and 41.7% used both formats. The figures measure different things, but both mean your extraction has to handle paper and electronic documents. Which rules apply to your shipments is a question for a qualified professional.
What does a 90-day first project look like?
- Days 1 to 30: choose, baseline and prepare. Pick one flow using the four questions above. Measure today’s touches, turnaround, handling time and error rate. Collect the sample archive, map the fields to your TMS, and agree what the system may never do. Build the first version against the archive, not live traffic.
- Days 31 to 60: run it in shadow. The system processes real incoming documents or emails, but a person does the work as usual and compares. Fix the failure patterns the comparison shows.
- Days 61 to 90: go live with review. The system handles real work, with a person approving what it writes or sends. Track the four measures weekly against the baseline.
- After day 90: widen by evidence. Move the fields and cases that stayed accurate to automatic handling, keep the rest under review, and choose the second flow. Keep running the fixed test set on every change.
This is a typical shape, not a promise: a locked-down legacy TMS will push more integration work into the first 30 days.
Questions logistics companies ask
What are the main AI use cases in logistics?
For document-heavy operations, the main ones are reading shipping documents, drafting quotes from emails, answering shipment status questions, handling exceptions and matching freight invoices. Beyond those, route optimisation, demand forecasting, ETA prediction and warehouse vision use other kinds of AI, such as solvers and machine-learning models.
Is AI replacing logistics jobs?
The evidence points to changed work rather than none. DHL’s Logistics Trend Radar says logistics will remain a people business while AI reshapes jobs and skills. In the systems we design, people keep decisions on money, capacity, customs and upset customers, and the AI takes the reading, typing and chasing.
How can a small logistics firm start?
Pick one repetitive flow, such as quote emails or document entry, and measure it for a few weeks. BDC advises tying AI to a concrete operational problem and notes that focused, off-the-shelf tools can often deliver value without a custom platform. Bring in outside help when the integration is complex or the data is sensitive.
Do we need to replace our TMS to use AI?
No. AI systems sit alongside your existing platforms, reading from and writing to them through their APIs, EDI or, as a last resort, their email and file exports.
How long until it’s live?
Our plan above puts a first flow live with review inside 90 days, including a shadow period where it runs alongside your team before it handles any work. The TMS integration is usually what stretches or shortens that.
Modulus Labs AI builds document AI, AI agents and the integrations around them. Describe the workflow that eats your team’s time, and we’ll tell you what we’d build first and what it would take.