RAG for Logistics Exceptions

What Problem Does It Solve?

Exceptions are the shipments that break the plan. A pickup is missed, a consignee refuses a load or a delivery arrives short or damaged. A truck waits at a dock long enough to trigger detention, a fee for driver time beyond the agreed free time.

Each exception raises the same questions. What does this customer’s contract say? Which accessorial charges apply? Accessorials are fees beyond the base freight rate, such as detention, lumper or redelivery fees. What does the carrier agreement allow? What should we tell the customer, and by when?

The answers are spread across contracts, customer routing guides, internal SOPs, rate sheets and email threads. Third-party logistics providers (3PLs) and brokers may manage dozens of customers, each with its own rules. Experienced coordinators know many of those rules by heart. Newer staff spend time searching, asking colleagues or guessing.

An exception assistant puts the relevant clause and its source in front of the coordinator. It can draft a customer update or carrier message. The coordinator checks the draft, edits it and decides whether to send it.

What Questions Can It Answer?

Which Content Should It Search?

• Customer contractsService terms, liability limits, notice periods and accessorial rules, each tagged with customer code, account team and effective dates.
• Customer routing guidesCarrier selection rules, appointment requirements, labeling rules and penalties for noncompliance, kept at the current approved version.
• Internal SOPsStep-by-step playbooks for missed pickups, refusals, shortages, damage and delays, including escalation paths.
• Rate and accessorial schedulesCodes, rates and free-time rules by customer and lane, with start and end dates.
• Carrier agreementsService expectations, claims terms and chargeback rules, visible only to teams allowed to see them.
• Message templatesApproved wording for customer updates, carrier notices and claim acknowledgments.
• Shipment recordsRead-only lookups in the transportation management system (TMS) for status, stops and reference numbers, limited to fields the exception needs.

How Does Hybrid RAG Work Here?

Hybrid RAG runs two searches at once and merges the results. Keyword search catches exact strings such as PRO numbers, customer codes and accessorial codes. Vector search matches by meaning, so “consignee would not accept the freight” can find a clause about refused shipments.

1. Contracts, routing guides, SOPs and rate schedules are split into sections, tagged with customer code, lane, document type, access group and effective dates.
2. Each section goes into a keyword index and a vector index. The vector index stores embeddings, lists of numbers that capture meaning.
3. A coordinator asks a question, often pasting a PRO number, bill of lading number or customer code alongside plain wording.
4. The system identifies the customer, lane and ship date, then filters both searches to documents the coordinator may see and that were in force on that date.
5. Keyword and vector results are merged, and a re-ranker, a model that compares the question with each passage, puts the strongest passages first.
6. The LLM answers only from those passages, citing the document, section and effective date for each point.
7. When asked, the assistant drafts a customer or carrier message from approved templates and the cited terms.
8. The coordinator reviews the answer and draft, edits as needed and decides what to send or escalate.

We start with hybrid search because exception questions are full of identifiers. Meaning-based search alone can blur two similar lane IDs or accessorial codes. Keyword search alone misses a routing guide that describes the same situation in different words. Combining both, with date and customer filters applied to each search, gives a dependable starting point. For how this compares with the other seven patterns, see RAG architectures explained.

When Should You Add Other RAG Patterns?

Launch hybrid RAG first, with a set of real exception questions and answers agreed by experienced coordinators. Add another pattern only when those results show a gap it fills.

Pattern: Multimodal RAG for Delivery Documents

The symptom is simple. Coordinators keep asking about evidence that lives in scans and photos, and the assistant cannot see it. Shortage and damage questions often hinge on a handwritten note on a bill of lading or a proof-of-delivery photo.

Multimodal RAG makes those files searchable. Optical character recognition (OCR) reads printed and handwritten text, and a vision-capable model describes photos. The results are indexed with the shipment’s reference numbers and linked to the original image. Handwriting and phone photos vary in quality, so unreadable fields should be flagged for a person rather than guessed. Processing images adds cost. For extracting fields at scale, intelligent document processing is often the better fit.

Pattern: Agentic Lookups Across the TMS and Email

The symptom here is that coordinators still open three systems for every answer. They check status in the TMS, search email for the carrier’s explanation and then look up the contract.

Agentic RAG hands those steps to an AI agent, software that plans steps and chooses its own tools. The agent checks shipment status in the TMS through read-only access. It reads the related carrier emails, finds the contract terms and drafts a reply or claim summary for approval. Agents are slower and less predictable than a single search. Emails can also contain text that tries to steer the agent. That calls for step limits, least-privilege tools and logging, covered in our agentic AI architecture guide.

Where Do People Stay in Control?

• Coordinators send every messageThe assistant drafts customer and carrier messages, and a person reviews, edits and sends them.
• No automatic claims or creditsFiling claims, issuing chargebacks and granting customer credits always require human approval.
• Disputes stay with peopleAccount managers and claims staff decide how to handle disagreements over charges, liability or service failures.
• Contract questions get escalatedUnclear or conflicting terms go to the client’s contract or legal team for the final call.
• Uncertain answers are labeledWhen sources do not cover a situation, the assistant says so and points to the right escalation path.
• Experts own the contentOperations leaders approve SOPs, templates and routing guide updates before they enter the index.

What Security and Compliance Controls Matter?

• Customer contract confidentialityEach customer’s terms are visible only to the teams assigned to that account, enforced before the model sees any text.
• Carrier and partner dataCarrier agreements and partner rates are handled according to the contract terms that govern them.
• Driver location privacyDriver location data is used only for the exception at hand, with clear purpose limits and access rules.
• Read-only system accessConnections to TMS, warehouse management system (WMS) and EDI data can read records but never change them.
• Current terms by effective dateRouting guides and rate agreements carry start and end dates, and superseded versions are excluded from answers.
• Data quality checksMissing tracking events, inconsistent reference numbers and unscanned documents are measured early and shown as gaps.
• Private deploymentThe assistant runs in your cloud tenant on Azure, AWS or Google Cloud.
• No public model trainingYour contracts, shipment data and messages are never used to train public models.
• Audit trailsQuestions, sources, drafts and approvals are logged for later review.

We do not provide legal advice, and the client’s legal and contract teams make the final calls on contract interpretation. Our AI governance, security and compliance service helps set controls, monitoring and oversight.

How Do You Measure Whether It Works?

• Retrieval accuracyHow often search finds the correct contract clause, SOP step or rate for the customer and date.
• Citation accuracyWhether each cited source truly supports the statement it is attached to.
• Effective-date correctnessHow often answers use the version of a routing guide or rate agreement in force on the ship date.
• Draft edit rateHow much coordinators change drafted messages before sending them.
• Handling time per exceptionTime from exception report to customer or carrier update, compared with the pre-launch baseline.
• Escalation qualityWhether cases the assistant flags as unclear were the right ones to escalate.
• Coordinator feedbackRatings and comments on wrong, missing or outdated answers.
• Cost per answerModel, search and tool costs combined.

How Do You Roll It Out?

1. 1

   Pick One Exception Type

   Choose a common, well-documented exception, such as detention or refused loads, for a small group of customers. Gather real questions from past cases and have experienced coordinators agree on correct answers.

2. 2

   Prepare the Content

   Collect the contracts, routing guides, SOPs and rate schedules for those customers. Tag each with customer code, access group and effective dates. Fix outdated or conflicting versions before indexing.

3. 3

   Build Hybrid Search

   Set up keyword and vector search with customer, access and date filters. Connect read-only TMS lookups. Test against the agreed questions and tune until the right clauses appear consistently.

4. 4

   Pilot With Coordinators

   Give the assistant to a few coordinators for everyday exceptions. They review every answer and draft before acting. Log feedback and failed questions to guide content fixes and search tuning.

5. 5

   Expand Carefully

   Add customers and exception types one group at a time. Consider multimodal document search or agentic lookups only when pilot results show the need. Keep measuring after each change.

Frequently Asked Questions

Start Handling Exceptions With Cited Answers

Tell us which exceptions take your coordinators the most time and where the answers live today. We will help you choose a starting point and build a pilot on your own documents. To get started, book a free AI consultation.