Read summarized version with
About project
Working time:
2025
Industry:
Legal technology
The service:
Generative AI
Overview
The customer is an international virtual data room provider whose platform hosts confidential documents for mergers and acquisitions, fundraising and audits. Its clients include investment banks, law firms and private equity deal teams who review thousands of documents under deadline.
The Challenge
Deal teams kept asking the platform for the same thing: help answering diligence questions across thousands of documents, without sending confidential data to public AI tools.
1. Repetitive Diligence Questions
Advisers answered the same questions about change of control, termination and assignment on every deal by rereading documents.
2. Terms Buried in Schedules
Key terms hid in schedules, amendments and side letters that keyword search did not surface.
3. Strict Confidentiality
Deal data could not leave the provider’s controlled environment, which ruled out public AI tools.
4. Bidder Isolation
Any AI feature had to respect folder and bidder permissions, so no bidder could see another bidder’s documents through an answer.
Technology & Approach
We designed AI search as a feature of the data room itself, running inside the provider’s AWS environment and enforcing the permissions the platform already uses.
Implementation
Phase 1: Design on a Sample Deal
Indexed an anonymised sample deal, agreed question types and answer format with the product team, and designed retrieval that filters by deal, folder and bidder before any passage reaches the model.
Phase 2: Ingestion and Retrieval
Built document ingestion with Amazon Textract for scanned files, a retrieval index per deal in Amazon OpenSearch Serverless, and cited answers with Amazon Bedrock through Amazon Bedrock Knowledge Bases.
Phase 3: Clause and Term Extraction
Added extraction of change of control, termination, assignment, exclusivity and governing law clauses into a review table that links each term to its source page.
Phase 4: Review Queue and Rollout
Answers and extracted terms go to a review queue where deal team members approve, edit or reject them before sharing, released first to a group of pilot clients.
Customization & Client Adaptation
The design started from the provider’s permission model, not from the model.
Permission-Aware Retrieval
Access checks run before retrieval, so the model only sees passages the requesting user is allowed to read.
Citations on Every Answer
Each answer links to the exact pages it relied on, and questions without support in the documents get an honest no-answer.
Data Stays in the Provider’s Account
Documents, indexes and model calls stay in the provider’s AWS account, encrypted with AWS KMS and logged in AWS CloudTrail.
Guardrails for Deal Language
Amazon Bedrock Guardrails keep answers within the deal’s documents and filter sensitive personal data from outputs.
If you find this case interesting, we recommend taking a closer look at
The Results
Pilot deal teams worked through standard diligence question lists in hours instead of days, with every answer cited to its source page.
Long-Term Value
- An AI capability the provider offers inside its data room, not as a separate tool.
- Extracted terms that export directly into diligence reports.
- A retrieval design that scales deal by deal without mixing data across deals.
Business Impact
- Advisers spend their time on judgment instead of rereading documents.
- Confidential deal data never leaves the provider’s controlled environment.
- Bidder isolation is enforced by design, not by convention.
