Allen & Gledhill

Allen & Gledhill, one of Singapore's largest law firms, faced a structural productivity problem common across the legal profession: lawyers at every seniority level were spending hours manually reviewing documents exceeding 100 pages, drafting case summaries, and verifying citations — work that added no analytical value but could not be delegated away. A 2025 LexisNexis survey found over half of legal professionals in Singapore and Malaysia now face client pressure to work faster using Gen AI. Beyond inefficiency, A&G faced a harder constraint: client data confidentiality rules ruled out public cloud LLMs entirely, while uncontrolled models introduced hallucination risk — plausible but legally incorrect outputs — that could expose the firm to serious liability.

A&G partnered with Pand.AI through IMDA's GenAI x Digital Leaders initiative to build A&GEL (pronounced 'Angel'), a custom on-premise Gen AI knowledge management system running entirely on A&G's own servers with stringent access controls and data encryption — eliminating cloud data exposure risk. The system was benchmarked across multiple open-weight Large Language Models, including Meta Llama, Microsoft Phi-4, DeepSeek R1, and Alibaba Qwen 3, before selecting the best-performing configuration. Critically, raw model capability was not the only design lever: A&GEL uses structured templates with fixed clauses, jurisdiction-specific headings, and citation frameworks to constrain and validate outputs. The pilot was opened to all levels — paralegals, junior lawyers, and directors — across banking, corporate real estate, maritime, and aviation practice groups, with A&G's technology lawyers coaching colleagues on Gen AI techniques to accelerate adoption.

The productivity impact was immediate and measurable. Lawyers who previously spent hours manually reviewing 100+ page contracts now use A&GEL to produce accurate document summaries with verified citations in minutes. Key outcomes include:

• Document summarisation: 100-page contracts summarised with citations in minutes, versus hours manually
• Compliance checking: Non-compliant clauses flagged automatically against statutes and prior cases
• First-draft generation: Case summaries and document analyses produced as structured first drafts ready for lawyer review
• Firm-wide adoption: Pilot spanned all seniority levels and multiple practice groups, surfacing domain-specific nuances that improved output quality

The time freed from document grunt work allows lawyers to focus on higher-value client counselling and hearing preparation.

• On-premise deployment is non-negotiable for law firms: client confidentiality obligations make public cloud LLMs an unacceptable risk regardless of vendor assurances.
• Model selection is necessary but not sufficient: structured output templates, fixed clause libraries, and jurisdiction references matter as much as which LLM is under the hood.
• Pilot breadth accelerates quality: including paralegals through directors surfaces workflow nuances that a narrow pilot would miss — and builds firm-wide buy-in simultaneously.
• Position AI as augmentation, not replacement: framing the tool around 'letting subject matter experts be subject matter experts' reduced resistance and kept lawyers engaged as co-designers.
• Compliance use cases are strong entry points: flagging non-compliant clauses is verifiable, low-risk, and immediately demonstrates value to skeptical practitioners.