Celadon — Systems & architecture

Making contracts & policies searchable with AI

CeladonUpdated July 20264 min read

Obligations and answers are buried in contracts and policy documents nobody has time to read end to end. AI makes them searchable, citable, and auditable.

Contracts, policies, and regulatory documents are where obligations live, and answers are hardest to find there. Teams either read hundreds of pages themselves, ask around and hope someone remembers, or guess and hope they guessed right. None of those scale, and the cost of guessing wrong in a legal or compliance context is rarely small. Contract and policy RAG turns those documents into a system that answers questions with citations, so people get the right answer quickly and can prove exactly where it came from.

What it delivers

In legal and compliance contexts, a citation is not a nicety — it is the whole point. An answer you cannot trace to a clause cannot be relied on, no matter how fluent it sounds.

Where teams start

The highest-value starting points are usually the questions that already get asked constantly and answered inconsistently: what a vendor contract obligates you to do, whether a proposed action is allowed under company policy, which contracts contain a specific clause across a portfolio of hundreds, and how a regulatory change affects existing agreements. These are bounded, repeatable questions with a right answer sitting in a document somewhere. That is exactly the shape of problem this kind of system solves well, and exactly the kind of task that otherwise falls on whoever in legal or compliance happens to have the most institutional memory.

Portfolio search is often the unlock. Finding every agreement with a most-favored-customer clause, a specific liability cap, or a non-standard termination right is work that used to mean weeks of associate time. A well-built retrieval system collapses that to minutes, with citations a lawyer can still verify before anyone relies on the result.

How permissioning actually works

The technical challenge that separates a real system from a demo is access control. A well-built implementation mirrors the permissions already set in your contract management or document system, so a person only ever sees, in an AI-generated answer, exactly what they would be allowed to see by opening the source file themselves. That has to be enforced at retrieval time, not bolted on afterward, or the system will eventually surface something it should not.

Versioning matters just as much. Contracts amend; policies revise; an answer grounded in last year’s MSA is a liability. The system should retrieve from the active version by default and make superseded versions available only for historical questions, with those versions clearly labeled as such. Logging should record which version supported each answer, so an audit months later can reconstruct what the system said and why.

Evaluation is not optional

Fluency is a poor proxy for correctness in legal text. Build an evaluation set from real questions your team already asks, with gold answers and clause references reviewed by counsel. Track citation accuracy as carefully as answer quality. A confident summary that points to the wrong section is worse than a slow human search, because it creates false confidence under deadline pressure.

Define what the system must refuse: legal advice to clients outside counsel review, interpretation of ambiguous clauses as settled law, or answers when retrieval returns weak support. Escalation to a human is a feature, not a failure. The organizations that succeed treat refusal as part of product design, not as an afterthought when something goes wrong.

Governance is the foundation

These systems handle sensitive material, so data handling and access control come first: before accuracy, before speed, and before anything else. Zero Data Retention and training exclusions belong in the vendor conversation before a single contract is uploaded. See AI without breaking confidentiality and What is RAG for how the underlying grounding and citation mechanics work.

Enterprise AI programs that skip governance rarely make it from pilot to production with real data, a pattern MIT NANDA’s 2025 research documents across industries. Contract RAG is especially exposed because a single permission miss ends the program. Build the controls first, then expand the corpus.

Start with an AI Decision Sprint.

Sources

Accessed July 2026. Vendor terms and benchmark methodologies change; verify current primary documentation before making a decision.

Turn the pattern into a production system

Build covers architecture, integration, evaluation, delivery, and handoff against written acceptance criteria.

Explore Build →