FOI Request: AI Redaction on GPU

The Challenge: 3,200 FOI Requests and a 20-Day Statutory Deadline

A metropolitan council in Northern England receives approximately 3,200 Freedom of Information requests per year — an average of 13 per working day. Each request requires identifying relevant documents (often scattered across multiple departments and legacy filing systems), reviewing every page for personal data requiring redaction under GDPR, identifying material exempt under sections 36, 40, 41, or 43 of the FOI Act, applying redactions, and preparing the disclosure package with exemption justifications. The FOI team of five spends an average of 14 hours per request on this process. With the 20-working-day statutory deadline, the team operates under constant time pressure, and compliance rates have dropped to 78% — below the ICO’s expected 90% standard. Two formal complaints to the ICO last year resulted in practice recommendations.

Documents contain highly sensitive material: social care records, safeguarding reports, commercial contracts, and personnel files. Processing through cloud AI services is strictly prohibited under the council’s information governance policy. All document handling must occur on UK-hosted, council-controlled infrastructure.

AI Solution: Automated Document Review and Redaction

Document AI and PaddleOCR running on a dedicated GPU server process the document set for each FOI request. The pipeline first digitises any scanned documents (the council holds 40% of records as paper scans). A named entity recognition model identifies personal data: names, addresses, phone numbers, email addresses, national insurance numbers, and dates of birth. An LLM via vLLM then reviews each document section against FOI exemption criteria, flagging material that may be exempt under specific sections and generating draft exemption justifications.

The FOI officer reviews the AI’s proposed redactions and exemptions, approving, adjusting, or overriding as needed — a 2-hour review process versus 14 hours of manual document-by-document analysis.

GPU Requirements

FOI document sets vary from 10 to 500 pages per request. The pipeline runs OCR on scanned pages, NER on all text, and LLM analysis on each section. Processing the daily intake of 13 requests (averaging 80 pages each) requires moderate GPU throughput.

An RTX 5090 processes the daily FOI workload in a single morning batch. Private AI hosting ensures all council documents remain within GDPR-compliant UK infrastructure — a non-negotiable requirement for local government.

Recommended Stack

• PaddleOCR for digitising scanned council documents with high accuracy on typed and handwritten text.
• SpaCy with custom NER models for UK-specific PII detection (NHS numbers, NI numbers, council reference numbers).
• vLLM serving a 7B model fine-tuned on FOI exemption criteria for section-by-section legal analysis.
• PDF redaction engine (pikepdf, PyMuPDF) for applying irreversible redactions to disclosure documents.
• Audit trail database recording every AI decision, officer override, and final redaction for ICO compliance.

For handling telephone call recordings included in FOI requests, add a vision model for CCTV frame analysis when footage is requested. Deploy an AI chatbot for public-facing FOI submission guidance and tracking.

Cost Analysis

The five-person FOI team costs approximately £200,000 annually. AI-assisted processing reduces per-request time from 14 hours to 2 hours — a 86% reduction. The same team can handle 7x the current volume, critical as FOI request volumes continue to grow at 8% annually. Alternatively, the team capacity freed by AI can address the compliance backlog, pushing the 78% on-time response rate back above the ICO’s 90% target.

Avoiding further ICO enforcement action — which can result in practice orders, fines, and reputational damage — protects the council from costs estimated at £50,000-£150,000 per enforcement case. The AI system’s consistent, auditable approach to redaction also reduces the risk of accidental personal data disclosure, which carries penalties of up to £500,000 under the Data Protection Act.

Getting Started

Compile 200 completed FOI responses with the original document sets and applied redactions. Use these as training examples for the NER model and exemption classifier. Process 50 new FOI requests through the AI pipeline in parallel with the manual process, comparing redaction accuracy and exemption identification. Target 98% recall on personal data detection (missing PII is far worse than over-redacting) before deploying as the primary workflow.

GigaGPU provides UK-based dedicated GPU servers for local government document AI workloads with full information governance compliance. Guaranteed UK data residency and no shared tenancy.