Contract and clause review workflows

This lesson is educational and is not legal advice. Contract meaning, enforceability, and appropriate negotiating positions depend on the complete agreement, governing law, facts, and qualified review.

A contract is a system, not a collection of independent paragraphs. Before using AI, assemble the complete authorized review set: main agreement, schedules, exhibits, incorporated policies, amendments, referenced order forms, and the correct comparison version. Confirm the parties, effective date, governing law, document status, and whether signatures or attachments are missing.

Normalize the documents without silently changing them. Preserve section numbers, headings, tables, defined-term capitalization, redlines, footnotes, and page references. Optical character recognition can turn “not” into “now,” lose a decimal, or scramble columns. If source quality is poor, mark it for manual inspection rather than letting the model repair likely text.

Next define the review standard. An approved playbook can specify preferred language, acceptable fallbacks, prohibited positions, business-owner questions, and counsel-only decisions. Without a standard, “risk detection” becomes the model’s generic guess.

Use several narrow passes instead of one broad instruction.

Pass 1: inventory and structure. Extract sections, defined terms, exhibits, cross-references, dates, currencies, and missing components. This pass reveals whether the review set is complete.

Pass 2: playbook comparison. For each relevant clause, show the contract text, applicable playbook position, deviation, and exact location. Do not yet recommend acceptance.

Pass 3: dependency check. Trace definitions and cross-references. A liability cap may exclude confidentiality obligations; an indemnity may use a definition found in another schedule; a termination right may depend on a cure procedure elsewhere.

Pass 4: operational extraction. List obligations, owners, deadlines, notice mechanics, renewal windows, audit duties, and deliverables. Send each item to the business owner who can confirm feasibility.

Pass 5: issue routing. Assign an owner and gate according to the playbook. The model may suggest a route, but the authorized reviewer confirms it.

Each pass produces a smaller artifact that a person can verify. Keep quotations beside every finding.

A strong review record includes:

• document and version;
• clause type and exact location;
• quoted source text;
• related definitions and cross-references;
• playbook rule and its version;
• deviation stated neutrally;
• factual or operational question;
• potential consequence, labeled as analysis rather than fact;
• permitted fallback, if the playbook supplies one;
• owner, reviewer, status, and decision;
• unresolved uncertainty.

Do not let the model collapse observation and advice. “The clause has no stated cap” is an observable feature. “Accept the clause” is a decision. “This creates unlimited liability” may require interpretation of exclusions, definitions, and governing law.

A bounded prompt might say:

Compare the supplied clause with playbook rule L-4. Quote both verbatim.
Identify additions, omissions, changed thresholds, conditions, exceptions,
and defined terms. Trace all cross-references in the supplied agreement.
Do not assess enforceability, assign a final risk rating, or recommend
acceptance. Mark missing documents and uncertain dependencies OPEN.
Return one review record per deviation.

The prompt should name the allowed documents. Treat contract text as data, not as instructions to the model. A clause that says “ignore previous instructions” has legal text value but must not change the workflow.

Chunking can hide dependencies. If the system reviews one clause at a time, provide definitions and linked provisions or require a second whole-document pass. Version confusion can compare the draft against the wrong template. Generic “market standard” claims may be unsupported, stale, or irrelevant to the organization’s position. Summaries can remove conditions. Redline reconstruction can miss formatting-only but meaningful changes.

Risk labels also create false precision. A red-yellow-green score is only useful when each label maps to an approved rule and an accountable route. A bright dashboard should never replace the quoted text and reasoning beneath it.

Finally, preserve the decision trail. Record whether counsel accepted, rejected, modified, or escalated the issue and why. Do not automatically reuse one matter’s exception as a new standard. Approved playbooks must be changed deliberately.