The Question Enterprise AI Is Not Asking
Every major banking AI programme is investing in context.
Context engineering, retrieval systems, knowledge bases, agent frameworks and increasingly sophisticated memory architectures have become central pillars of enterprise AI strategy. The logic appears obvious. If AI produces inaccurate answers, provide it with more information. Connect more systems. Ingest more documents. Increase the context window. Improve retrieval.
The assumption is so widely accepted that it rarely attracts scrutiny.
Yet beneath this consensus lies a question that receives remarkably little attention:
How do we know the context is true?
It is a deceptively simple question, but one with profound implications for the future of banking. Enterprise AI is becoming increasingly effective at accessing information, yet many organisations have invested far less effort in establishing whether that information is complete, connected, current and trusted.
The banking industry may be approaching the challenge from the wrong end.
The Race for Context
The last two years have transformed enterprise AI. The conversation has evolved from prompt engineering to retrieval-augmented generation, from retrieval to agents, and from agents to context engineering. The progression makes sense.
Every organisation that has experimented with AI has discovered the same reality. Generic intelligence is useful. Enterprise intelligence is valuable. A large language model may possess broad knowledge of economics, risk management, financial products and regulation. Yet the moment it enters a bank, those capabilities become insufficient.
The model needs access to:
- Internal policies
- Product definitions
- Risk methodologies
- Regulatory interpretations
- Operational procedures
- Customer processes
- Governance controls
In other words, it needs context. The industry's response has been to focus on delivering more of it. But context may not be the asset we think it is.
Context Is What We Provide
At its core, context is information assembled for a specific task. A policy document provided to answer a compliance question is context. A knowledge-base article retrieved by an agent is context. A collection of emails, reports or code supplied to a model is context. Context exists to make reasoning more relevant.
It answers the question:
What information should the AI consider right now?
Without context, enterprise AI is largely disconnected from the reality of the organisation. It is little surprise that so much effort is being invested in improving it. Context engineering is essential because it helps AI work with the right material at the right moment. But it has an important limitation. Context tells us what information is available. It does not tell us whether that information should be trusted.
The Hidden Assumption
Most enterprise AI initiatives contain an assumption so deeply embedded it often goes unnoticed. The assumption is that enterprises already possess knowledge. In reality, many possess something else entirely. They possess information.
Banks in particular are extraordinary generators of information.
- Policies.
- Models.
- Reports.
- Logs.
- Code.
- Data.
- Risk assessments.
- Meeting minutes.
- Architecture diagrams.
- Control inventories.
The challenge is rarely whether these assets exist. The challenge is establishing how they relate to one another.
- Which policy created this control?
- Which model implements that policy?
- Which process supports that model?
- Which reports depend upon that process?
- Which change requests altered the outcome?
Banks are rich in policies, controls, models, code, processes, data definitions and regulatory obligations. What is often missing is the durable, governed connection between those assets.
The information exists. The relationships do not. Without those relationships, information remains information. It may be useful, relevant and accurate in isolation. But it is not yet enterprise knowledge.
Before AI, This Was Good Enough
For decades, banks were able to function despite this fragmentation. Experienced people became the connective layer between policies, systems, processes, models and controls. Architects understood system dependencies. Risk specialists understood regulatory intent. Business analysts understood operational processes. Developers understood implementation details.
The enterprise was effectively held together by institutional memory. Ask a subject matter expert which reports depended upon a specific calculation and the answer could often be provided immediately. Ask the same question across disconnected repositories, undocumented processes and fragmented ownership, and the answer frequently became much harder.
The relationships existed. They simply existed inside people. That was manageable whilst humans remained the primary consumers and interpreters of enterprise information. AI changes the economics of that model because organisations can now process, compare and reason across information at a scale that no individual expert can match. But AI can only work reliably with relationships that have been identified, validated and made explicit.
AI did not create the need for enterprise knowledge. AI revealed its absence.
AI Did Not Create The Problem
There is a temptation to view the challenges of enterprise AI as AI problems.
- Hallucinations.
- Missing context.
- Incomplete answers.
- Lack of explainability.
Whilst these symptoms are real, they often originate elsewhere. The uncomfortable reality is that AI has exposed weaknesses that already existed inside the enterprise. If business logic is fragmented, AI inherits fragmentation. If lineage is incomplete, AI inherits incomplete lineage. If provenance is unclear, AI inherits unclear provenance. The model is not creating these uncertainties. It is revealing them.
For the first time, organisations have technology capable of reasoning across vast quantities of enterprise information. Yet when that reasoning begins, an uncomfortable truth emerges. The enterprise may know far less about itself than it assumed. This also creates an opportunity. Historically, enterprise knowledge scaled mainly through people. AI creates the possibility of amplifying trusted knowledge across employees, applications, processes and agents. The opportunity is not to replace expertise. It is to capture, connect and amplify it.
For most of history, knowledge scaled through people.
AI may become the first technology capable of scaling knowledge itself.
Knowledge Is Different
Knowledge begins where information ends. Knowledge is not merely the existence of facts. Knowledge is the establishment of governed relationships.
It introduces:
- Provenance
- Lineage
- Traceability
- Contextual meaning
- Governance
- Trust
Knowledge answers questions that information alone cannot.
Not simply:
But:
- Why does it exist?
- How is it connected?
- What depends upon it?
- What happens if it changes?
Most importantly:
This distinction becomes increasingly important as AI moves into decisions involving risk, compliance, auditability and transformation. In these environments, relevance alone is insufficient. Context determines what AI can see. Knowledge determines what AI can trust.
Context is task-specific. It is assembled at the point of need to support a particular question or workflow. Enterprise knowledge should be different. It should be durable, governed, connected and reusable. The two disciplines are not competitors. Context engineering helps AI retrieve and use the right material. Knowledge engineering helps ensure that material carries meaning, provenance and trust before it is retrieved.
A Different Starting Point
The enterprise AI industry has spent substantial effort helping models access information. The next challenge may be helping enterprises establish understanding. That does not reduce the importance of context engineering. Far from it. Context engineering solves a critical problem. It helps AI determine what is relevant at a particular moment. But relevance and trust are not the same thing. A perfectly retrieved document may still be wrong. A complete set of reports may still be disconnected. A larger context window does not automatically produce understanding.
Perhaps the challenge facing banking is not simply how to provide AI with more context. Perhaps it is how to establish trusted enterprise knowledge before that context is assembled. The AI industry has spent enormous effort teaching machines how to reason. The next challenge may be helping organisations understand themselves. Before AI can answer better questions, enterprises may first need better knowledge.
That challenge has a name.
The Enterprise Knowledge Gap
Next in the series: The Enterprise Knowledge Gap, exploring why banks possess more information than at any point in history, yet still struggle to explain how their organisations actually work.
