Ahmed Al Mubarak on the importance of context engineering in article intelligence
Howden Re’s Director of AI and Data Science, Ahmed Al Mubarak, recently wrote the cover story for Issue 12 of Data & AI Magazine which explores a growing shift in how organisations should think about deploying large language models (LLMs). His piece argues that context engineering, rather than prompt optimisation, is what enables AI systems to produce outputs that are reliable, auditable, and fit for enterprise decision-making.
Al Mubarak noted: “As businesses look to embed LLMs into critical workflows, particularly in data-dense sectors such as insurance and reinsurance, reliance on well-constructed prompts can be limiting. While prompts can shape responses, they cannot, on their own, guarantee accuracy, auditability, or consistency when models are exposed to messy, multi-source, and often unstructured enterprise data.”
At a system level rather than the level of individual interactions, the solution, Al Mubarak argues, is context engineering: the deliberate combination of content, constraints and capabilities that make LLMs useful inside complex workflows.
To make this practical, he outlines a five-layer framework designed to turn LLMs into reliable enterprise tools: knowledge, operational, structural, behavioural, and integration.
Underpinning this approach is the use of AGENTS.md files, a lightweight convention for documenting and versioning an AI system’s tools, rules, and operating assumptions as a first-class software artefact. By making a model’s assumptions explicit and versioned, this approach reduces reliance on individual prompt craft, strengthens governance, and ensures consistent AI behaviour across teams, projects, and time.
For Howden Re, designing context around policies, broker submissions, wordings, and external market data enables a move beyond surface-level summarisation toward grounded analysis that preserves source structure, provenance, and intent. This philosophy underpins NOVA, Howden’s (re)insurance market insight platform, which brings together placement data and broader market financials from hundreds of curated sources to turn fragmented information into strategic clarity. By structuring and contextualising these diverse datasets before they reach the model layer, NOVA allows AI systems to generate insights that are rooted in evidence rather than inference.
The result is AI that supports underwriting, claims, and advisory work with outputs that are consistent, explainable, and suitable for review, rather than opaque responses that require manual correction. Instead of relying on isolated prompts applied to disconnected data, the system is engineered to understand how information fits together across the market, enabling more confident and defensible decision-making.
As Al Mubarak commented, “When you engineer the context system properly, models become competent collaborators rather than gifted improvisers. At Howden Re, this approach transformed how we process complex insurance documents from scanned policies to multi-source market data; delivering grounded, auditable insights at scale.”