Skip to main content
DoneThat

AI Adoption GuideLegalApprove

Risk delta highlight

Extracts only what changed since the last approval round and summarizes the delta for fast reviewer consumption.

Legal processRequestAssessDraftNegotiateApproveSignStoreDispute

By Don, DoneThat’s AI coach · updated

Overview

Risk delta highlight answers one question for approvers who already saw an earlier draft: what changed, and does any of it matter? Instead of sending the full agreement back through review, the workflow compares the current version against the last approved baseline, isolates material edits, and produces a short delta brief. Each listed change includes the exact span from the prior version and the corresponding span in the new version so a reviewer can jump to evidence without hunting through a 40-page redline.

The outcome target is quality: fewer missed regressions, fewer false alarms from formatting noise, and faster cycles when the answer is genuinely "nothing material changed." The reviewer still approves. The delta does not substitute for sign-off; it narrows what requires attention.

What belongs in a risk delta highlight

A useful delta is selective. It surfaces edits that could alter legal exposure, commercial terms, obligations, remedies, data handling, or alignment with the approved playbook. It ignores immaterial differences such as typo fixes, cross-reference renumbering after a defined-term move, or whitespace and pagination shifts that do not change meaning.

Each material item should be self-contained enough that a busy approver understands the risk in one pass. A strong entry names the topic (for example, limitation of liability cap), states what the prior approved language did, states what the new language does, and flags why the shift may matter. When the model or rules engine cannot tie a change to a playbook or prior negotiation position, it should say so rather than inventing a severity label.

Span citation is non-negotiable for quality. Every delta line pairs:

  • Prior span: document identifier, section or clause reference if available, and the quoted or anchored text from the last approved version
  • New span: the same reference frame in the current draft and the replacement text

If either side cannot be located reliably, the item belongs in a "needs manual locate" bucket or is omitted until alignment improves. Guessing at spans erodes trust faster than omitting a low-confidence diff.

When comparison yields no material change, the output should be explicitly empty: a clear statement that the current draft matches the approved baseline within configured tolerance, not a padded summary of immaterial edits. Reviewers use that empty result to approve quickly, often after spot-checking one or two anchor clauses rather than the full document.

How risk delta highlight fits the approve stage

Approve-stage review assumes a baseline already exists: a prior round where counsel or a business owner accepted specific language. Risk delta highlight is built for round two and beyond: counterparty returns a markup, internal counsel revises after a committee comment, or a renewal draft arrives with "no changes" that still needs verification.

It complements broader artifacts rather than replacing them. Approval summary generation gives a holistic picture of what is being approved; risk delta highlight restricts the lens to movement since the last approval. Redline delta summarization often covers all tracked changes between any two files; risk delta highlight filters that stream against approval relevance and playbook severity. When deviations from standard positions appear in the delta, reviewers cross-check playbook deviation report output for structured rule hits. Teams that score approval friction may feed delta features into an approval outcome predictor, but prediction is downstream; the highlight itself stays descriptive and evidence-linked.

Typical inputs include the last approved PDF or Word file, the current draft, metadata tying versions together (matter ID, approval timestamp, approver role), and optionally the playbook or fallback positions used in the prior round. Outputs are a delta list, span map, and empty-state confirmation when applicable.

Workflow from baseline to approver

Establish the baseline. Lock the last approved version as the comparison anchor, not "whatever was emailed most recently." Version conflation is the most common source of false deltas and false empties.

Normalize before diff. Convert to a common representation where possible. Scanned PDFs, heavily formatted Word files, and CLM-exported HTML can produce spurious diffs if normalization is skipped. Table moves, footnote renumbering, and defined-term capitalization changes should be classified early so they do not flood the risk list.

Diff and classify. Structural comparison identifies insertions, deletions, and replacements. A second pass applies materiality rules: playbook distance, clause type, numeric threshold changes, and net-new obligations. Immaterial hunks are dropped from the highlight set but may remain in a technical log for audit.

Attach spans. For each surviving hunk, record prior and new spans at clause or paragraph granularity. Prefer stable anchors (clause number plus quoted substring) over page numbers, which shift when formatting changes.

Deliver to reviewer. Present deltas in risk-relevant order: net-new liability language before a defined-term tweak, for example. Include deep links or CLM jump targets when the platform supports them. If the set is empty, state that no material change was detected against the configured rules and show the baseline reference the comparison used.

Approve with eyes open. The reviewer confirms the delta list is complete enough for the decision at hand, adds comments on any disputed item, and records approval against the current version ID. An empty highlight is not a rubber stamp; it is a scoped finding that the automated pass found nothing material, which experienced reviewers still validate against their own checklist for high-stakes deals.

Quality bar and failure modes

Quality means traceable, complete-within-scope, and appropriately empty.

Traceable every material bullet cites prior and new spans reviewers can verify in under a minute. Complete-within-scope means everything that meets materiality thresholds appears; immaterial noise stays out. Appropriately empty means the system does not manufacture findings to look productive, and does not hide material edits because comparison failed silently.

Watch for these failure modes:

  • Wrong baseline compared against a draft from an earlier negotiation branch, producing a long frightening delta that is mostly stale history
  • Over-merge treating two adjacent edits as one risk when only one moved a cap or carve-out
  • Under-merge flooding the reviewer with ten line items that are one conceptual change split by formatting
  • Span drift after post-processing (PDF generation, CLM export) that breaks anchors even when language is unchanged
  • False empty when materiality rules are too aggressive or OCR missed a swapped numeric threshold

Mitigations include mandatory baseline metadata validation, human-in-the-loop confirmation when confidence is low, side-by-side preview for top-severity items, and periodic golden-file tests on known contract pairs.

Vendor capabilities: Ironclad, Litera Compare, DocuSign CLM, Icertis

No single product owns the full risk delta highlight pattern end to end; teams usually compose comparison, CLM workflow, and optional AI summarization.

Ironclad centers contract lifecycle and workflow on its platform. Ironclad's AI-assisted review and playbook features help classify deviations on incoming drafts. For delta-style review, Ironclad is strongest when both versions live in Ironclad with clear version history; approvers benefit from workflow-native approval records tied to specific document revisions. Span citation often maps to clause objects or AI-highlighted passages rather than raw Word offsets. Pair Ironclad with a dedicated compare engine when counterparty markups arrive offline and must be ingested before in-platform diffing is reliable.

Litera Compare is a specialist document comparison tool widely used in legal workflows. It excels at producing accurate redlines between Word and PDF pairs and is a common backbone for "what changed" truth. Litera Compare does not by itself decide materiality or approval relevance; it feeds accurate hunks that downstream rules or summarization layers filter into a risk delta highlight. Teams typically use Litera for span fidelity and a separate playbook or CLM layer for severity.

DocuSign CLM (formerly SpringCM) provides repository, workflow, and approval routing with versioning in the CLM context. Delta highlights align naturally with "approve this revision against revision N" when versioning discipline is strong. DocuSign's broader eSignature footprint helps final execution, but approvers should confirm compare features and AI summarization scope in their tenant; some deployments rely on integrated or third-party compare for attorney-grade redlines before CLM approval steps.

Icertis targets enterprise contract management with strong obligation and clause analytics at scale. Icertis is useful when delta review must connect to metadata (contract type, region, business unit) and when AI clause identification helps anchor spans across large libraries. Risk delta highlight in Icertis-heavy environments often combines version compare with playbook and obligation models so a changed indemnity cap surfaces alongside affected obligation records.

Selection heuristics: prioritize Litera Compare (or equivalent) when span accuracy is the bottleneck; prioritize Ironclad or DocuSign CLM when approval audit trail and in-platform versioning are the bottleneck; prioritize Icertis when enterprise metadata and clause intelligence must ride along with the delta list.

Operating rules for reviewers and admins

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first