The Challenge: Four Thousand Contracts, One Deadline
A mid-cap UK manufacturer is acquiring a competitor. The target company’s contract portfolio contains 4,000 active supplier, customer, and partnership agreements. The acquiring company’s legal team needs to identify every contract containing a change-of-control clause, most-favoured-nation provision, or termination-for-convenience right before the deal closes in six weeks. Manually reviewing 4,000 contracts — many running to 80+ pages — would require a team of eight paralegals working full-time for the entire period, at an estimated cost of £180,000 in temporary staffing and £40,000 in external counsel review of flagged clauses.
The contracts include confidential pricing terms, customer identities, and competitive intelligence that the acquiring company’s board considers highly sensitive. Uploading these to a SaaS contract review platform means placing the target’s most commercially sensitive information on infrastructure the legal team does not control — a risk the board will not accept during a deal where information security is paramount.
AI Solution: LLM-Based Clause Identification and Extraction
An open-source LLM can read each contract and identify specific clause types with high accuracy. The model receives the full contract text (or relevant sections for very long agreements) and a structured prompt specifying the clause types to search for. It returns identified clauses with their exact location, a plain-English summary of their effect, and a risk rating based on how the clause might impact the transaction.
The pipeline handles the diverse formats these contracts arrive in. Document AI processes native PDFs and Word documents, while PaddleOCR tackles scanned legacy agreements and faxed amendments. The extracted text feeds into the LLM, which applies a consistent analytical framework across the entire portfolio regardless of drafting style, governing law, or document age.
GPU Requirements: Batch Processing a Full Contract Portfolio
Contracts are text-heavy: a typical supplier agreement occupies 5,000-15,000 tokens. Analysing for multiple clause types requires either multiple passes or a single long-context pass. With modern LLMs supporting 32K-128K context windows, most contracts fit in a single inference call. The 4,000-contract portfolio needs sustained batch throughput over 24-72 hours.
| GPU Model | VRAM | Contracts per Hour (Mistral 7B) | 4,000 Contracts |
|---|---|---|---|
| NVIDIA RTX 5090 | 24 GB | ~45 | ~89 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~65 | ~62 hours |
| NVIDIA RTX 6000 Pro | 48 GB | ~75 | ~54 hours |
| NVIDIA RTX 6000 Pro 96 GB | 80 GB | ~110 | ~37 hours |
An RTX 6000 Pro through GigaGPU processes the entire portfolio in under two days. For ongoing contract management beyond the deal — monitoring new agreements, tracking renewal dates, flagging risky clauses at execution — an RTX 6000 Pro handles daily volumes easily at a lower price point.
Recommended Stack
- Mistral 7B-Instruct or LLaMA 3 8B-Instruct for clause identification — these models handle legal language well and fit comfortably in memory alongside OCR and preprocessing workloads.
- vLLM for batch inference with continuous batching optimisation.
- Apache Tika for extracting text from native PDFs, DOCX, and legacy formats.
- PaddleOCR for scanned contracts and image-based PDFs.
- Pydantic structured output parsing to enforce consistent JSON output from the LLM — clause type, location, risk level, summary.
- Airtable or PostgreSQL for the clause register, feeding into a review dashboard for the legal team.
Post-acquisition, the same infrastructure powers an AI chatbot that lets the commercial team query the merged contract portfolio: “Which supplier agreements allow price increases exceeding RPI?” or “List all contracts with European counterparties that reference GDPR specifically.”
Cost vs. Alternatives
Commercial contract analysis platforms (Kira Systems, Luminance, eBrevia) charge per-contract or per-page fees that for a 4,000-contract review typically total £60,000-£200,000. These platforms are capable, but they require data to flow through the vendor’s infrastructure. A self-hosted approach on dedicated GPU provides comparable extraction quality at significantly lower cost, with the decisive advantage that every contract stays on UK infrastructure under the firm’s direct control.
The speed advantage compounds the value. Receiving a complete clause register within 48 hours rather than six weeks gives the deal team actionable intelligence early enough to negotiate indemnities, request consents, or restructure the transaction around problematic contracts.
Getting Started
Assemble 200 contracts with known clause classifications from a completed transaction. Run the LLM extraction pipeline and measure precision and recall against the known ground truth. Fine-tune if needed — typically one iteration brings recall above 92% for standard clause types. Deploy on the live portfolio with lawyer review focused on the borderline-confidence extractions.
GigaGPU offers private AI hosting with the throughput M&A contract reviews demand. Process thousands of agreements on GDPR-compliant infrastructure you control, with no per-contract licensing fees.
GigaGPU’s UK-based dedicated GPU servers process contract portfolios at deal speed with full data sovereignty.
View GPU Server Options
