Contract Entitlement RAG Check

A contract entitlement RAG check sits in the confirm stage of freight AP, after the invoice is parsed and before a human or downstream system posts approval. For each accessorial line (detention, layover, fuel surcharge, liftgate, residential delivery, reweigh, and similar charge codes), a retrieval model searches the indexed carrier contract corpus, pulls the governing clause span, and states whether the billed amount is entitled under the contracted terms.

The output is a per-line entitlement result: charge code, cited clause span, and a clear pass, fail, or empty finding. Empty means the relevant contract is not indexed or the clause cannot be located with enough confidence to cite. In that case AP still owns the approval decision. The check never silently invents entitlement or blocks payment on missing retrieval.

This pattern is retrieval-augmented generation applied to commercial terms, not a general chat assistant over PDFs. The model is constrained to retrieve, quote, and compare. It does not renegotiate rates or rewrite contract language.

Accessorials are where contracted price and billed price diverge most often. Base linehaul rates are usually easier to match against a rate table. Detention windows, free time, fuel formulas, and exception fees live in denser prose, schedules, and amendments. Auditors who skim invoices without opening the contract tend to approve charges that look familiar rather than charges that are entitled.

A retrieval check forces every accessorial to answer the same question: does this charge code appear in the contract, under what conditions, and at what rate or formula? When the answer is no, or the billed amount exceeds the contracted formula, the discrepancy is visible before cash leaves. When the answer is empty, the gap is also visible: the team knows the contract was not available to the model, so the human review is deliberate rather than assumed.

Cost impact compounds across carriers and lanes. A single missed detention rule on a high-volume carrier can exceed the annual cost of indexing and running the check. The confirm-stage placement matters because the invoice is already structured into lines and codes; the model does not have to invent the charge taxonomy from free text alone.

The pipeline starts with a normalized invoice object: carrier identity, shipment references, charge codes, amounts, units, and dates that matter for free-time and surcharge formulas. For each accessorial line, the system builds a retrieval query that includes the carrier, charge code synonyms, and any temporal context (effective date of the invoice vs. contract version).

The retriever returns ranked contract chunks from the indexed corpus for that carrier relationship. The generation step is narrow: given the charge line and the candidate spans, decide entitlement and emit the supporting clause text with enough surrounding context that an AP analyst can open the source document to the same place. Citation of the clause span is mandatory for pass and fail. Empty is reserved for no usable span.

Charge-code mapping is a first-class concern. Carrier invoices and contracts rarely use identical labels. A robust check maintains a controlled vocabulary that maps invoice codes and free-text descriptions to contract concepts (for example, “DET” and “driver wait time” both resolving to detention). Without that mapping, retrieval either misses the right clause or over-matches unrelated fees.

Versioning belongs in the same loop. Amendments, rate riders, and dated schedules change entitlement. The index should retain contract identity, effective dates, and supersession so the model cites the version that applied on the ship or invoice date, not the latest PDF uploaded to a shared drive.

Entitlement RAG is complementary to other confirm-stage agents, not a replacement for them.

An automated 3-way match agent confirms that invoice, PO or load tender, and proof of delivery align on quantities, services, and references. Match answers “was this service authorized and delivered?” Entitlement answers “under the contract, may this accessorial be billed at this amount?”

An invoice anomaly detector flags statistical or rule-based outliers (unusual amounts, duplicate patterns, unexpected code combinations). Anomaly detection is a prior or parallel signal; entitlement RAG is the contract-grounded explanation for why a flagged accessorial should be paid or disputed.

An AI rate benchmarking engine compares billed rates to market or peer benchmarks. Benchmarks inform negotiation and exception policy. Entitlement RAG stays inside the four corners of the signed agreement.

A billing discrepancy classifier routes and labels disputes once a mismatch is known. Entitlement findings (wrong formula, outside free time, charge not in schedule) become structured inputs to that classifier so dispute packets cite the same clause the model used.

Together, these agents reduce cost by catching unauthorized or miscalculated charges early, while keeping AP as the approval authority when retrieval returns empty or confidence is insufficient.

Empty-if-not-indexed is a product requirement, not a failure mode to paper over. Teams that force a yes/no on every line when the corpus is incomplete train users to distrust citations. Prefer explicit empty, with a queue for contracts still to be ingested, over a low-confidence guess.

Human approval remains the gate. The check attaches evidence to the invoice workspace; it does not auto-deny payment unless the organization separately defines that policy for high-confidence fails. Most logistics AP groups start in advisory mode: surface the clause, recommend dispute language, let the analyst decide.

Corpus quality dominates model quality. Scanned amendments without OCR, missing rate appendices, and carrier packs stored only in email produce empty findings even when entitlement exists on paper. Treat contract ingestion (identity resolution, OCR, clause segmentation, effective dating) as part of the same program as the RAG model.

Auditability matters for finance and for carrier relationships. Store the retrieved span, model version, charge code mapping used, and timestamp with the approval record so a later dispute can replay what the system saw.

Enterprise contract platforms such as Ironclad and Icertis already hold executed agreements, clause metadata, and often obligation or commercial-term structures. They are natural sources of truth for the indexed corpus and for clause-level citations, especially when logistics MSAs and rate schedules live alongside other supplier contracts. Coupa sits closer to the procure-to-pay and invoice workflow: entitlement checks that run where invoices are coded and approved reduce swivel-chair work between CLM and AP. Cass AI and related freight-audit offerings focus on carrier invoices, accessorial taxonomies, and transportation spend patterns; they are strong when the primary corpus is freight contracts and historical audit outcomes rather than a general CLM library.

In practice, many teams combine a CLM or contract repository (Ironclad or Icertis) for authoritative text with a freight AP or audit layer (Cass AI or Coupa invoice flows) for charge codes and approval UX. The RAG check is the glue: retrieve from the contract system of record, decide entitlement per line, write the citation back into the invoice confirm step.

Selection criteria should emphasize citation fidelity, charge-code coverage for your carrier mix, contract versioning, and clear empty behavior, not model marketing claims alone. Pilot on one high-volume carrier with a complete, dated contract pack. Measure dispute yield, false fail rate, and empty rate before expanding the index. Cost outcome is real when entitled fails turn into recovered or avoided accessorial spend, and when empty rates fall as the corpus matures, without removing AP from the final approval.