---
title: "AI Contract Review in Microsoft Word"
canonical_url: "https://cognilium.ai/products/paralegent-ai"
slug: "products/paralegent-ai"
section: "products"
page_type: "pillar"
date_modified: "2026-09-29"
description: "Every clause read against your own playbook inside Microsoft Word, with a green, orange or red risk read and replacement language drafted. In your own cloud."
publisher: "Cognilium AI"
author: "Mudassir Marwat"
person_same_as:
  - "https://www.linkedin.com/in/mudassir-marwat/"
  - "https://orcid.org/0009-0008-1927-2598"
  - "https://github.com/mudassirmarwat"
  - "https://medium.com/@mudassir-marwat"
  - "https://www.youtube.com/@mudassir-marwat"
  - "https://dev.to/mudassirmarwat"
  - "https://hashnode.com/@mudassirmarwat"
  - "https://huggingface.co/mudassirmarwat"
  - "https://bsky.app/profile/mudassir-marwat.bsky.social"
  - "https://substack.com/@mudassirmarwat"
  - "https://topmate.io/mudassirmarwat"
  - "https://fueler.io/mudassirmarwat"
  - "https://www.instagram.com/mudassirmarwat/"
  - "https://www.facebook.com/mudassir.marwat"
  - "https://www.reddit.com/user/mudassirmarwat/"
  - "https://www.f6s.com/member/mudassir-marwat"
  - "https://cognilium.ai/founder"
answers:
  - "Contract review, inside Microsoft Word"
  - "The slow part is the reading, and it never gets faster"
  - "Four modules, and the third is where the argument is"
  - "Green, orange, red — and what to do about each"
  - "Master agreements, and everything hanging off them"
  - "There is no accuracy percentage on this page"
  - "Playbook-driven AI contract review, built to order"
  - "The questions a legal team actually asks"
entities: []
related:
---

# AI Contract Review in Microsoft Word

**Your playbook becomes terms** The playbook you already have — usually an Excel file — is parsed into 80 to 150 enriched terms. That parse is deterministic and spends zero model calls, because a rulebook that changes depending on which model read it is not a rulebook.


## In short

**Your playbook becomes terms** The playbook you already have — usually an Excel file — is parsed into 80 to 150 enriched terms. That parse is deterministic and spends zero model calls, because a rulebook that changes depending on which model read it is not a rulebook.

**The contract is retrieved, not skimmed** The agreement is chunked and indexed, then searched by vector similarity and re-ranked by a model, so each playbook term arrives at the passages that actually bear on it rather than at page one.

**Twenty-three agents, in parallel** Twelve category-scoring agents read the contract at once and route what they find to eleven specialist analysts. Routing is the expensive decision made cheaply: it cut model calls by 75% against reading everything with everything.

**It arrives where the work happens** The result lands in the Microsoft Word add-in your team already opens the contract in, with an operations dashboard behind it for the people running the queue. Nobody learns a new place to read a contract.

Clause protects your commercial interests. Low risk, standard market terms.

Moderate risk. May be acceptable depending on deal size, counterparty, or strategic importance.

Materially unfavorable. Creates disproportionate liability, IP exposure, or financial risk.

Reject or replace with approved fallback

**What is playbook-driven AI contract review?** It is contract review where the rulebook is your own playbook rather than a vendor's generic standard. The playbook is parsed into enriched terms, each contract clause is retrieved and re-ranked against the relevant terms rather than skimmed, category agents score the clauses they specialise in, and a person decides at the end. The distinguishing property is that the standard being applied is yours.

**Can you build a custom AI contract review system for us?** Yes, and that is the main reason Paralegent AI exists — it is our own working proof of the pipeline rather than a product we are trying to sell seats of. A build takes your document types, your categories and your playbook, and runs in your own cloud. The engagement starts the way every other one does: a scoped four-to-six-week pilot on one workflow with a single success metric agreed up front.

**Do we have to use your categories and your playbook?** No. Nothing is hard-coded to the twelve categories described on this page — those came from one playbook. The categories are derived from whatever rulebook you hand over, which is usually an Excel file your team already argues from. The pipeline is document-shaped rather than contract-shaped, so leases, funder agreements, statements of work and policy packs fit the same machinery.

**What is Paralegent AI and how does it work?** It reviews master agreements clause by clause against a rulebook built from your own legal positions. Your playbook is parsed into 80 to 150 terms, the contract is indexed and retrieved against those terms, twenty-three agents score and analyse what comes back, and the result arrives in a Microsoft Word add-in with each flagged clause classified, reasoned, and paired with drafted replacement language.

**How long does a review take?** Five to ten minutes end to end for a master agreement, against the hours a clause-by-clause human read takes. Inside that window the system makes somewhere between one thousand and three thousand three hundred model calls; the reason it can afford to is that the routing step decides which specialists ever see a given passage.

**How is this different from a SaaS contract-review tool?** It deploys into your own cloud, so the contracts never leave your environment, and it is driven by a rulebook built for your business rather than a generic playbook. It is also model-agnostic: the provider is a configuration choice, running through a routing layer, so your data residency and your model spend stay decisions you make.

**What accuracy does it achieve?** We do not publish an accuracy percentage, and you should be suspicious of anyone in this category who does — the number depends entirely on whose contracts and whose playbook. Accuracy is measured on your own agreements during implementation: we run your real contracts through, compare against your own reviewers, and iterate until it meets your standard. Every flag arrives with its classification, its reasoning and a drafted replacement, so each one is checkable rather than trusted.

**Where does it run, and who holds the data?** In your own cloud environment. The system runs as a containerised pipeline with its own queueing, retry and tracing, deployed as infrastructure as code with separate staging and production environments and a typed confirmation gate on production deploys. Your contracts stay in your infrastructure, and you control model selection and cost through your own accounts.

**What contract types does it cover?** Master Sales Agreements, Master Purchase Agreements, Master Service Agreements, Sales Contracts and Orders, Purchase Contracts and Orders, and Statements of Work — with the subordinate documents cross-checked against the parent agreement's terms and your rulebook.

**How does an implementation actually go?** Four phases. We work with your legal team to map contract types, legal positions and risk tolerance into a structured rulebook; we stand the system up inside your cloud; we tailor the interface and the classifications to how your team works; then we run your real contracts through, measure, iterate, and hand over with documentation. A dedicated team works in your environment throughout.

**Can we try it before committing?** Send us one master agreement and we will run it through and walk you through the marked-up result: which clauses were flagged, how each was classified, and what replacement language the system drafted. Judging this product on your own contract is the only judgement worth making.

Paralegent AI checks every clause against your own playbook, marks the risky ones green, orange or red, and drafts the language it would put there instead.

Five to ten minutes end to end, inside the Microsoft Word document your team already has open.

legal categories scored, routed to 11 specialists


## Contract review, inside Microsoft Word

Paralegent AI — contract review inside Word


## The slow part is the reading, and it never gets faster

Every enterprise processes hundreds of master agreements a year. Each one runs forty to a hundred pages that have to be read clause by clause.

**The expensive hours go to the repetitive work** Your attorneys spend their highest-value hours doing clause matching rather than negotiating the terms that decide the deal.

**Deals wait while contracts sit** Revenue recognition slides, counterparties grow impatient, and vendor onboarding stalls behind a queue nobody can clear faster by trying harder.

**The same clause reads differently twice** Two attorneys read the same indemnity and reach two positions, because the playbook lives in people's heads rather than in a file the work is checked against.


## Four modules, and the third is where the argument is

Most tools in this category are one model reading one contract. This is a pipeline, and each stage exists because the stage before it made the next one cheaper.

A single review still spans somewhere between one thousand and three thousand three hundred model calls. That is what reading a master agreement properly costs; the routing is what makes it affordable.


## Green, orange, red — and what to do about each

A flag with no recommendation is just more reading. Every classification carries the action it implies, and the drafted language to get there.

When two sections both speak to the same clause, an orchestrator resolves the conflict and keeps the single most accurate redline rather than showing your team the argument.


## Master agreements, and everything hanging off them

The subordinate documents are checked against the parent agreement as well as the playbook, which is where the deviations usually hide.

Within each of them, twelve categories are scored and routed to the specialist that owns them — warranty, indemnity, intellectual property, liability and termination among them. A category agent knows what it is looking for, which is why it finds the clause that a general read walks past.


## There is no accuracy percentage on this page

Any number we published would be measured on our contracts and our playbook, and you would be buying it for yours.

So accuracy gets measured during implementation, on your own agreements, against your own reviewers, and we iterate until it clears your bar. The same rule retired the older figures that used to sit here: a contract count, an accuracy claim, a turnaround guarantee and a client-logo row. None of them survived a check against what we can actually show you, which is the system itself, running on a contract you bring.


## Playbook-driven AI contract review, built to order

is the mechanism, and Paralegent AI is our own working proof of it: a playbook parsed into enriched terms, clauses retrieved and re-ranked rather than skimmed, twelve category agents scoring in parallel, and a person deciding at the end. We built it so we could show the pipeline running rather than describe it.

The same pipeline gets built to order — your document types, your categories, your playbook, your cloud.

**Your playbook, not ours** The rulebook is whatever your team already argues from — usually an Excel file. Nothing is hard-coded to the categories above; the categories come from the playbook you hand over.

**Your documents, not just contracts** The pipeline is document-shaped, not contract-shaped. Leases, funder agreements, statements of work, policy packs — anything reviewed repeatedly against a standard fits the same machinery.

**Your cloud, your data** It runs as a containerised pipeline inside your own environment. The agreements never leave it, which is the part that decides whether legal will use the thing at all.

What is shipped, built to order, or a blueprint


## The questions a legal team actually asks

Every answer here is in the page source, not behind a click, because the systems that summarise this page never click anything.

in our Dynamics 365 Optimizer family — the one member of it that is shipped.

We will run it through and walk you through the result: which clauses were flagged, how each was classified, the reasoning behind it, and the replacement language the system drafted. One of your contracts tells you more than any page about it can.


---

Canonical HTML: https://cognilium.ai/products/paralegent-ai