Explainer · Agentic legal AI
How AI reviews a contract against your own playbook.
Five steps, four hard problems, and the calls I made leading Paralegent AI from idea to production.
The Supplier’s total liability under this Agreement shall not exceed the fees paid in the prior six (6) months.
Cap: the greater of USD 1,000,000 or 12 months of fees. Below that: negotiate.
Cap below 12 months of fees and below the USD 1M floor.
Suggest: “the greater of (i) USD 1,000,000 or (ii) fees paid in the prior twelve (12) months.”
The short answer
AI contract review against a playbook compares every clause with the positions a legal team has already agreed: its preferred wording, the fallback it accepts and its limits. The system reads the playbook once, matches each clause to the right term, has specialist agents assess the gap, and returns a finding with the rule, the reason, replacement language and a confidence score. The lawyer decides.
Five steps, from playbook to a decision in Word.
This is the pipeline in Paralegent AI. The order is the point: the standard comes first, the lawyer’s decision comes last.
- 80–150 terms, once
Turn the playbook into rules.
The legal team’s playbook, in PDF or DOCX, is read once and broken into 80–150 terms. Each term carries the preferred position, the fallback language, the GREEN / ORANGE / RED limits, the rationale and who approves an exception. No written playbook? It is built from the team’s approved contracts. This is the real onboarding work, and it is what makes every later finding defensible.
- 15–20 seconds
Match each clause to its rule.
Each clause of the incoming contract is embedded as a 1,536-dimensional vector and matched to the playbook term it relates to.
- 18+ specialists
Send it to the right specialist.
There is one specialist per playbook category — 18+ today: warranty, liability, indemnity, IP, termination, payment, data and more. An orchestrator sends each clause only to the two or three specialists whose terms it matches most closely. When two disagree, the higher-confidence finding wins and both rationales are kept for the reviewer.
- 2–8 min · 40–50 findings per MSA
Write a finding a lawyer can check.
Each finding carries a risk level (GREEN approve, ORANGE escalate, RED negotiate), the playbook rule, a plain reason, replacement language and a confidence score. The replacement language is the playbook’s own preferred or fallback wording, not text the model invents. Only high-confidence findings are surfaced.
- Accept · Modify · Reject
The lawyer decides, in Word.
Findings arrive as native Word comments on the exact clause. The reviewer accepts, modifies or rejects each one, and the reviewed .docx stays the deliverable.
The four problems every contract-review AI has to solve.
The pipeline is the easy part to describe. Whether lawyers trust it and use it comes down to four product problems. These are the calls I made on each.
| The problem | The easy route | My call | What it cost |
|---|---|---|---|
| GroundingCan a lawyer defend every finding? | Ask a general model whether a clause looks risky. | Review against the customer’s own playbook, so each finding cites the rule it departs from. | Onboarding takes real work: 80–150 terms extracted once. |
| AdoptionWill lawyers actually use it? | Another web app to log into, or copy-paste into a chatbot. | A Word add-in. Findings land on the exact clause as native comments. | Office.js limits the interface. |
| Data controlWill legal and IT approve it? | Run it in the vendor’s cloud and ask security to sign off. | Deploy in the customer’s own Azure, AWS or Google Cloud, on the LLM they choose. Contract data never leaves. | An 8–10 week implementation per customer instead of a sign-up page. |
| TrustCan the reviewer check it quickly? | One model, one answer, no way to see why. | 18+ specialists and an orchestrator; every finding carries its rule, reason and confidence. | More moving parts to coordinate. |
Two shapes of legal AI. They fit different buyers.
The category now has two shapes. Neither is better in general.
A broad legal AI workspace
Research, document review, drafting and multi-step agent work across a legal team’s day.
Fits when a team wants one AI workspace for many kinds of legal task.
A playbook system in your own cloud
One job — contract review against the team’s own playbook — inside Word, deployed in the customer’s cloud and owned after a one-time implementation.
Fits when the standard is your own playbook and contract data must stay in your environment.
Harvey and Legora are described from their own public descriptions, October 2026. This is not a ranking.
Five questions to ask before you buy.
Whichever shape you choose, these answers decide whether lawyers will trust it and IT will approve it.
- Whose standard does it apply — ours, or a generic one?
- Where does a finding appear — on the clause, or in another tool?
- Can we see the rule, the reason and the confidence behind each finding?
- Where do our contracts go, and which model reads them?
- Who owns the system after implementation?
Questions
What people ask about playbook review.
Credit and limits
Paralegent AI is a Cognilium AI product; Mudassir Marwat is Founder & CEO. The multi-agent engineering was a team effort; I led the product. Figures are Paralegent’s published product statistics, confirmed with its founder (2026); how the product works is described from Paralegent’s own product pages (June 2026). The clause above is fictional. This page explains how the product works; it is not legal advice.
How does AI contract review against a playbook work?
AI contract review against a playbook compares every clause with the positions a legal team has already agreed: its preferred wording, the fallback it accepts and its limits. The system reads the playbook once, matches each clause to the right term, has specialist agents assess the gap, and returns a finding with the rule, the reason, replacement language and a confidence score. The lawyer decides.
What is a contract playbook?
A legal team’s agreed positions for each type of clause: the preferred wording, the fallback it will accept and the point at which a deal needs sign-off or negotiation. In Paralegent AI, a playbook becomes 80–150 terms that every contract is checked against.
How is Paralegent AI different from Harvey or Legora?
Harvey and Legora offer broad AI workspaces for legal teams: research, document review, drafting and multi-step agent work. Paralegent AI does one job. It reviews contracts against the customer’s own playbook, inside Microsoft Word, and it is deployed in the customer’s own Azure, AWS or Google Cloud on the LLM they choose, as a one-time 8–10 week implementation the customer owns. It is not a ranking: they fit different buyers.
What happens in the 8–10 week implementation?
A dedicated pod of 3–4 engineers builds it with the legal team: playbook design in weeks 1–2 (from existing guidelines or approved contracts), deployment in the customer’s cloud in weeks 3–4, Word and contract-system integration in weeks 5–6, testing on real contracts in weeks 7–8, and training and go-live in weeks 9–10.
Can AI contract review run in our own cloud?
Yes. Paralegent AI deploys inside the customer’s own Azure, AWS or Google Cloud environment, on Azure OpenAI, AWS Bedrock or Google Vertex AI, so contract data never leaves.
Does AI contract review replace the lawyer?
No. It scores routine clauses and explains each deviation; the reviewer accepts, modifies or rejects every finding. In Paralegent’s published figures, an 80-page master agreement takes about 30 minutes of review instead of 20–30 hours by hand.
What did Ali Ahmed do on Paralegent AI?
As AI Solutions Engineer at Cognilium AI, I have led Paralegent AI as its product manager, from idea to production, since January 2026. The four calls on this page are mine; the multi-agent engineering was a team effort.
Building AI for experts?
Talk to me about AI product leadership, agentic systems for document-heavy work, or a role that joins product judgment with hands-on delivery.