Allen & Gledhill deploys custom on-premise GenAI assistant A&GEL to streamline contract review and legal workflows

Allen & Gledhill, one of Singapore's leading law firms, faced persistent inefficiencies driven by high-volume manual workflows: lengthy contract reviews, document translation, and administrative paperwork that consumed billable lawyer hours without generating proportionate client value. The legal sector's sensitivity around client confidentiality compounded the problem — off-the-shelf GenAI tools running on public cloud infrastructure were fundamentally incompatible with professional obligations to protect privileged client information. Simultaneously, the factual accuracy limitations of public LLMs posed unacceptable risk in a profession where errors carry legal and reputational consequences the firm could not audit or control.

Allen & Gledhill partnered with Pand.AI through IMDA's GenAI x Digital Leaders Programme to develop A&GEL (pronounced 'Angel') — a fully on-premise, custom-built GenAI assistant deployed entirely within the firm's own infrastructure. Starting from a consolidated list of over 100 candidate use cases, the team prioritised those with the greatest impact and feasibility, ultimately building specialised modes for contract review, contract translation, document generation, and advanced legal document summarisation. The solution evaluated multiple open-weight Large Language Models — including Llama2, Phi-4, DeepSeek, and Qwen 3 — before settling on configurations tuned to produce verifiable, lawyer-reviewed outputs. Cross-functional collaboration between technology lawyers, business users, and Pand.AI ensured the system aligned with real-world legal workflows rather than theoretical use cases.

A&GEL delivered measurable productivity gains across both legal and business service staff, reducing time spent on laborious tasks such as document summarisation, contract review, and translation. Key outcomes include:

• Improved efficiency on high-volume document workflows that previously consumed significant lawyer hours
• Enhanced document search and management capabilities across the firm's internal knowledge base
• Minimal onboarding friction — staff required little training to adopt basic functionality
• Knowledge transfer from Pand.AI built lasting in-house AI expertise within the firm
• Thought leadership positioning at the GenAI–legal intersection, opening new client service and advisory opportunities

• Build on-premise when client confidentiality is non-negotiable — cloud-based public LLMs may be structurally incompatible with legal professional obligations, regardless of contractual safeguards.
• Prioritise use cases by impact and feasibility before building; consolidating 100+ candidates into a focused product prevents scope creep and improves delivery quality.
• Resist model-chasing: evaluating multiple LLMs is necessary, but committing to specialised, verifiable outputs matters more than adopting the latest release.
• Cross-functional collaboration between lawyers, business users, and technologists is essential — subject matter experts must remain the authoritative voice on what 'good output' looks like.
• AI should augment professional expertise, not replace it; outputs must be designed for lawyer verification, not blind acceptance.