What this automation does
The tool does the reading a lawyer would otherwise do line by line, then hands over an issues list the lawyer can work through quickly.
- Submit. A lawyer or contract manager uploads the agreement, or opens it in Word with an add-in, and picks the agreement type and which side you are on.
- Normalize. The document is converted to clean text with clause numbers, defined terms and cross-references preserved. Existing tracked changes from the counterparty are extracted separately.
- Resolve definitions. Defined terms are collected first, so a clause using “Losses” or “Confidential Information” is read with its actual definition.
- Check against the playbook. For each playbook topic, such as limitation of liability, indemnity, IP ownership, confidentiality, termination, auto-renewal, assignment, governing law and data protection, the model finds the relevant clause, quotes it and compares it with your standard and fallback positions.
- Flag missing terms. The model reports playbook topics with no matching clause, such as a vendor agreement with no data security obligations.
- Suggest language. For deviations, it proposes replacement text drawn from your approved clause library, not freshly invented wording.
- Lawyer decides. The issues appear as Word comments or a review memo with risk levels. The lawyer accepts, edits or rejects each one, and their choices are logged to improve the playbook.
When it is worth building
The volume test: if your legal team reviews dozens of similar commercial agreements a month and most of the time goes to spotting the same handful of issues, a playbook-driven first pass saves real hours. It is especially useful for vendor paper, where every counterparty’s template is different and the issues are in different places.
It is not worth building for low-volume, high-stakes work such as M&A documents or bespoke financing agreements, where every clause is negotiated from scratch. It is also not the first step if you lack a playbook; writing one with your lawyers is often the most valuable part of the project, with or without AI. And if an off-the-shelf review product fits your agreement types and data rules, compare it honestly with a build. Our custom agent vs off-the-shelf comparison lays out that decision.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Document intake | Word add-in, or upload from SharePoint, Box or Google Drive | Lawyers stay in Word, and documents stay in the repository your firm or company already secures |
| Parsing | A DOCX parser that keeps numbering, definitions and tracked changes | Clause structure matters; flattening a contract to plain text loses meaning |
| Playbook | Structured records per topic: standard, fallbacks, escalation triggers | Legal edits rules directly, and every flag cites the rule it applied |
| Clause library | Approved language stored with tags and version history | Suggested edits come from text your lawyers already approved |
| Model | Claude through the API, with long context for full agreements | Reads a complete MSA with its exhibits and follows a strict output schema |
| Quote verification | Code that checks every quoted passage against the source | Removes issues based on text that is not in the contract |
| Matter link | Clio or your matter system through its API | Review results attach to the right matter with the right permissions |
| Notifications | Microsoft Teams or email | Tells the requester when the first pass is ready for legal |
How to build it
- Write or formalize the playbook. Start with the two or three agreement types that make up most of your volume. For each topic, write the standard position, acceptable fallbacks and walk-away points, in your lawyers’ words.
- Build the clause library. Collect approved fallback language from past negotiations, with the context in which each was accepted.
- Assemble a test set. Gather 30 to 50 past agreements with the issues your lawyers actually raised. This is how you will measure the tool.
- Build parsing and definitions. Get numbering, defined terms and cross-references right before any prompt work. Most quality problems trace back to bad parsing.
- Run the playbook check topic by topic. One focused pass per topic is more reliable than one giant prompt for the whole agreement.
- Verify quotes and score against the test set. Measure what the tool flagged that lawyers also flagged, what it missed, and what it flagged wrongly.
- Deliver in the lawyer’s workflow. Output as Word comments or a memo attached to the matter. Pilot with one or two lawyers for a month before wider use.
What breaks in production
- Order of precedence. An MSA says one thing and the order form overrides it. The tool must read the full contract stack and the precedence clause, or it will approve terms that a later document changes.
- Hidden counterparty edits. Incoming drafts sometimes contain tracked changes or unflagged edits to your own paper. Compare against your template version, not just the playbook.
- Absence errors. A model reports a clause as missing when it sits in an exhibit, or misses that it is missing. Lawyers should confirm every “missing” flag.
- Long scanned agreements. Image-only PDFs need OCR first, and quote verification gets weaker. Mark these as lower confidence.
- Playbook drift. Lawyers start accepting a fallback that is not in the playbook. Review logged decisions monthly and update the rules.
- Over-trust by business users. A sales lead reads “no issues found” as approval. The output must always say who still needs to approve.
Timeline and cost
Expect four to eight weeks, with playbook work and test-set evaluation taking a large share. Lumeor builds this in the range of $18,000 to $45,000, depending on the number of agreement types, whether output goes into Word or a matter system, and how much clause library work is needed. Model usage per agreement is usually small next to the cost of an hour of legal time, but long agreements with many passes add up, so we estimate it on your test set.
How to measure success
- Lawyer minutes per first-pass review, by agreement type
- Precision: share of flagged issues the lawyer agrees are real
- Recall: share of issues the lawyer found that the tool also found, measured on a sampled set each month
- Share of suggested edits accepted with little or no change
- Turnaround time from request to first response for the business
- Playbook updates made from logged decisions each quarter
Contract review sits in the middle of a larger flow. Contract intake and routing gets agreements to legal with the right context, and NDA generation takes the simplest agreements off legal’s desk entirely.