Context Engineering: The Enterprise Brain Behind AI in Banking

What does the AI know about the bank it works for?

Most financial institutions possess vast amounts of valuable information and while AI is becoming mainstream in banking, this fundamental question is often overlooked.

We have spent three decades building banking technology in specialized layers. A tier-one financial institution today typically operates between 800 and 1,500 production applications and core banking, cards, lending, treasury, risk, compliance and CX each carry their own data models, reference vocabularies and quiet assumptions about what a customer or an account means. Those assumptions become expensive when we ask AI agents to reason across them.

A large language model deployed against a bank’s data estate without a proper context layer is, in effect, a fluent stranger. It speaks banking fluently in the abstract, but it does not know your bank’s people, products, history or specific dialect — the definitions your risk team uses, the way your product taxonomy differs from a competitor’s, the operational history that explains why a particular exception exists. Every answer it produces has to be translated back into the bank’s own terms by someone on the staff, thus eroding the promised productivity gains.

Closing the gap between a fluent stranger and a trusted colleague is what context engineering does. It is the layer of the AI stack that banks are most consistently underinvested in, even as they continue to automate processes and client interfaces.

What is Context Engineering?

Enterprise AI architectures increasingly organize around three concerns:

• A Core that provides the governed technical foundation for enterprise AI — model management, identity, security and observability.
• A Coordination layer where people, agents, applications and systems collaborate in defined workflows.
• And between them, Context, which we describe as the bank’s trusted enterprise brain.

Context is the layer that gives every AI application reliable and traceable access to the institution’s data, processes, knowledge and history, enriched with domain awareness, lineage, business rules and compliance markers.

The raw material of that brain is scattered widely across the bank. It lives in transactional systems and reference data stores, in process documentation and standard operating procedures, in audit findings and complaint logs, in call-center transcripts and relationship-manager notes, in RAG pipelines and curated document repositories. Often, it also includes tacit knowledge that has never been written down, such as training videos, whiteboard sessions or tribal expertise held in the heads of long-tenured staff. A context layer is what makes this material findable, comparable and interoperable so an AI system can reason across it, rather than each application rediscovering the same knowledge from scratch.

Making that brain intelligent rests on three components working together – taxonomy, ontology and knowledge graph.

A controlled vocabulary or taxonomy, gives the enterprise a single canonical name for every entity class it cares about. Product classifications, entity types, event categories and regulatory concepts get named once, versioned and owned by a domain steward. It sounds pedestrian until you consider that most banks cannot answer the question “how many active mortgage customers do we have” with a single number, because five downstream systems each define “active” differently. A taxonomy will not settle that disagreement on its own, but it makes it visible and traceable where governance can begin to do useful work.

On top of the vocabulary sit ontologies that encode how those terms relate to each other. Taxonomy names entities and ontology models the connections between them: a Customer holds an Account, a Risk Event affects a Position and a Data Asset is governed by a Regulatory Rule. Expressed formally, these relationships let a reasoning engine surface implicit connections that no human has manually tagged.

The knowledge graph is the persistent structure where entities from source systems are resolved, de-duplicated, linked and made available for traversal. A well-populated graph in a large bank runs into millions of nodes, drawing from the data catalogue, the API registry, the schema registry, business glossaries and the regulatory corpus. What matters is less the size of the graph and more the discipline of the pipeline feeding it, from source discovery through normalization against the taxonomy to publication into search and traversal endpoints with confidence scores and freshness commitments attached.

In our engagements across large-bank data estates, we consistently see between 40%-50% of AI-application build effort consumed by context assembly work encompassing entity resolution, glossary reconciliation or lineage reconstruction, before the application logic is complete.

Why Does This Layer Matter Now?

Three forces are compressing the timeline for getting the context layer right.

The volume of AI-driven decisions is rising sharply. NVIDIA’s 2026 State of AI in Financial Services survey found that 65% of respondents are actively using AI and 42% are using or assessing agentic AI. The share of decisions that touch a model is growing across underwriting, servicing, fraud, KYC and treasury. Each of those decisions inherits the quality of the substrate underneath.

The economics have shifted. 70%-80% of our BFSI clients identify data quality, governance and availability as leading barriers to scaling AI across financial services. An AI pilot deployed against a semantically fragmented data estate produces uneven answers and the promised productivity gain then dissolves into the cost of human review.

Regulatory expectations have hardened. The European Union AI Act, UK Financial Conduct Authority (FCA’s) evolving guidance on AI model governance, Monetary Authority of Singapore (MAS)‘s Fairness, Ethics, Accountability and Transparency (FEAT) and the US Federal Reserve’s the Office of the Comptroller of the Currency (OCC) and the Federal Deposit Insurance Corporation (FDIC’s) model risk management framework have converged on a common expectation: the institution must be able to explain, in defensible terms, what a model considered when it produced a particular answer. Technical lineage alone will not carry that explanation. A regulator asking why a customer was flagged as high-risk expects an answer grounded in the bank’s own definitions, business rules and decision history. This is what the context layer holds.

What Distinguishes the Programs That Work?

Banks build the taxonomy for domains where vocabulary is partially governed, populate the graph from the two or three most authoritative source systems in it and deliver an AI-driven search or Q&A capability aimed at internal specialists. The programs that generate compounding returns resist the temptation to declare victory and move to the next domain. They extend the taxonomy horizontally, build cross-domain equivalence mappings between finance, risk and compliance definitions, treating the context layer as shared enterprise infrastructure rather than a domain-specific asset. Every new domain plugged into a common semantic backbone increases the value of every domain already connected to it and this is where the investment starts to compound.

The strongest programs also treat the context layer as a product with a roadmap and a service-level commitment against which downstream teams can plan. Confidence scoring is attached to every node, so that downstream applications can filter on quality. Freshness commitments of 24 hours for batch sources and 60 seconds for streaming ones are becoming table stakes.

What This Looks Like in Practice

One mega US wealth management firm was fielding client queries through a service organization of representatives searching a SharePoint repository of 1000+ internal documents. It supported 36 million+ active accounts, average query resolution ran four to seven minutes and accuracy varied by representative. With target to grow the customer account base to 50 million+, scaling the service organization on the existing model would have required a steep and expensive hiring curve.

We built a service knowledge assistant powered by GenAI, backed by a context layer that reorganized the SharePoint corpus including client history, product documentation, operational procedures and compliance guidance accumulated across years of the firm’s growth into a form the model could reason over. A feedback rating loop from service representatives improved the response quality over time.

Query resolution dropped from four to seven minutes to four to seven seconds. Accuracy rose above 93% against enterprise-verified documentation. The AI assistant handled more than 16,000 interactions daily across all service tiers and the variance in answer quality from the old model largely disappeared.

A search engine bolted on top of the same SharePoint would have inherited its inconsistencies and stalled inside the first cohort. The AI Assistant worked because the knowledge underneath had been engineered for it. The fluent stranger became a colleague who spoke the firm’s dialect.

Where This Leaves the AI/Technology Leadership

The AI conversation has been dominated for two years by model selection, agent frameworks and coordination protocols. Those choices continue to matter. But the Context layer deserves the same attention, because it decides whether the models are reasoning over the bank’s real knowledge or over an approximation of it. Every capability built above inherits the quality of that decision.

The banks that will pull ahead over the next twenty-four months are the ones treating context as a central program rather than as a workstream inside individual AI projects. That means a named owner accountable to the executive committee, a shared roadmap that consolidates taxonomy, ontology and knowledge-graph work across domains and a funding model that lets downstream teams draw on the context layer as a service rather than rebuild it every time.

The agent count on any given roadmap is a lagging indicator of AI maturity. The state of the enterprise knowledge underneath is the leading one and that is where the technology leaders’ attention will pay the highest return over the next twenty-four months.