Workflow Automation · Read + Write

Contract clause review vs playbook

Augment Rep · Med Reason · High

The problem

Legal reads every incoming contract end-to-end. Spotting deviations from the playbook takes ~4 hr per MSA. Legal is the bottleneck.

The AI approach

LLM parses each clause; vector search matches it to the closest playbook position; risk-scored deviations come back with suggested fallback language.

The outcome

4 hr → 25 min

First-pass redline in minutes. Legal focuses on the few clauses that need judgement.

Try itInput → Process → Output

Input — Incoming MSA
MSA — Vendor Xv2 · 14 pages
§ 4 · TermThis Agreement shall be effective for 12 months from the Effective Date.
§ 7 · Limitation of liability"Liability under this Agreement shall not exceed USD 10,000 in the aggregate, regardless of cause."
§ 9 · Governing lawSingapore law shall govern, with disputes settled by SIAC arbitration.
§ 11 · Data protectionVendor will process personal data per PDPA and Customer DPA Annex 1.
Process — AI pipeline
1Parse + segment clausesReadPredictive
2Match to playbook (vector)ReadSymbolic
3Flag deviations + risk scoreWriteGenerative
4Draft fallback languageWriteGenerative
Output — Redline pack
Click Run demo to flag deviations and propose fallback clauses.
CLAUSE REDLINE§ 7 · LoL
Risk · liability capHIGH ▲
As proposed "… shall not exceed USD 10,000 in the aggregate" Playbook standard "… cap = greater of 12 months fees or USD 250,000, excluding IP indemnity"
Suggested fallback "Liability cap shall be the greater of (a) fees paid in the 12 months prior; or (b) USD 250,000. The cap shall not apply to IP indemnity, confidentiality breach, or wilful misconduct."

Three AI types in this use case

SymbolicPlaybook database (clause → standard position → fallback ladder); risk-tier rules; redline output format (track-changes).
PredictiveClause segmentation, type classifier (liability, IP, term, gov-law), similarity ranking against playbook embeddings.
GenerativeLLM compares each clause to the playbook position, explains the deviation, drafts fallback language in the playbook voice.

The stack

  • Parser · DocAI / unstructured
  • Embeddings · text-embedding-3-large
  • LLM · Claude Sonnet
  • Sink · Word track-changes / CLM

When this works

  • Standard contract types (MSA · NDA · DPA)
  • Playbook is current + signed-off
  • English contracts at first
  • Reviewer signs off before sending

When it fails

  • Bespoke contracts (M&A · IP licensing)
  • Playbook is stale or contradictory
  • Multi-language without per-language playbook
  • Politically loaded clauses needing judgement