Artificial Intelligence

Beyond AI Copilots: Building Decision-Ready Intelligence in Banking

Written by: Diego De Aragao | CFA, SVP at CITI

Updated 8:00 AM EST, November 18, 2026

post detail image
Diego De Aragao | CFA, SVP at CITI Diego De Aragão, CFA, is a former Citi senior vice president specializing in finance, data, and AI. He advises financial services organizations and teaches generative AI at CUNY.

The AI conversation in banking has gotten ahead of the architecture. Over the past two years, financial institutions have launched copilots for almost everything, including research, underwriting, relationship management, finance, operations, risk, software development, and internal knowledge. Many of them are useful — some are genuinely impressive — but deploying another assistant is becoming the easy part. The greater challenge comes when all of these tools begin consuming the same institutional data, producing recommendations that influence real decisions, and acting on behalf of employees.

That is why I think the copilot is a useful transition point but a poor target architecture. The real opportunity for chief data officers (CDOs) is to create a governed intelligence layer that sits across the institution’s data, knowledge, analytics, and workflows. Instead of building intelligence independently inside every application, the institution develops reusable capabilities for assembling context, retrieving proprietary knowledge, applying analytical reasoning, enforcing entitlements, and delivering intelligence into the places where decisions already happen. In financial services, this distinction is critical because a useful answer and a defensible answer are not necessarily the same thing.

I have seen this problem from a slightly different angle through years of work in balance-sheet management, treasury, finance, and advanced analytics. When we built increasingly sophisticated decision support and optimization capabilities around capital, liquidity, funding, and balance-sheet allocation, the mathematics was rarely the hardest part. The real work was getting the data right, preserving the economics behind it, understanding the constraints, and making the output useful to an executive who ultimately had to make a decision. AI changes the technology, but it does not change that basic reality. If anything, it raises the bar because the machine can now produce a very convincing answer before the institution has established whether that answer deserves to be trusted.

The Copilot Was a Useful Prototype, Not the End State

The current pattern inside many large institutions reminds me of what happened during earlier waves of analytics transformation. Every team finds a problem worth solving and builds a capability around it. Commercial banking develops client intelligence while credit teams build document analysis tools and finance introduces natural-language querying. Meanwhile, operations creates an internal knowledge assistant, and risk develops its own environment to meet tighter control requirements. Each initiative makes sense in isolation, yet the institution gradually ends up with duplicated retrieval logic, different definitions of the same information, inconsistent controls, and another generation of technology fragmentation.

That is exactly what we should avoid with AI. Retrieval, identity, entitlement, provenance, model routing, orchestration, and policy enforcement should increasingly behave like reusable institutional services rather than features recreated inside individual copilots. The employee may still experience AI through an application, but the intelligence underneath should be portable. A relationship manager should be able to ask an AI assistant whether the economics of a relationship have changed without leaving the CRM, just as an underwriter should be able to identify missing information in a credit package without opening another tool. Intelligence should be available within the workflows where its context and consequences are understood.

This may ultimately make some of the most valuable AI almost invisible. I am not convinced that employees want another destination, another prompt box, or another AI interface to manage. They want better information at the point where they have to make a decision. That pushes the architecture away from a collection of copilots and toward something closer to an enterprise intelligence utility — one that can serve different workflows without losing the controls that distinguish banking from a generic knowledge-management problem.

A Client 360 That Only Aggregates Data Is No Longer Enough

Banks have been pursuing some version of Client 360 for decades, so simply attaching AI to a consolidated client record is not particularly transformational. The more interesting opportunity is to turn that record into a living decision asset: one that captures what the institution knows about a client, along with what has changed, why it may matter, and what should be investigated next.

Take something as mundane as a reduction in operating deposits. In most systems, that appears as another movement in an account balance. However, in the context of the broader relationship, it may mean something very different. Have balances migrated to another institution? Have external transfers increased? Has payment activity declined? Is the client reacting to pricing? Has there been a service issue, a corporate event, or a deterioration in the underlying business? The point is not for AI to declare that the client is at risk but to assemble enough context to tell the banker that something deserves attention and to explain why.

This is where I think the industry sometimes confuses having more information with having more intelligence. The institution may already possess virtually every data point required to understand the situation. The problem is that the signals sit in different systems, under different ownership, and at different levels of granularity. The value comes from connecting them around a business question without flattening the controls around the underlying information. A genuinely unified client view is therefore not a bigger database. It is a continuously evolving hypothesis about the relationship, grounded in evidence and open to challenge.

Moving Beyond Retrieval to Decision-Ready Intelligence

The same distinction applies to retrieval-augmented generation (RAG). RAG has been enormously useful because it gives models access to the institution’s own information rather than forcing them to reason entirely from public training data. But retrieving 10 relevant documents and producing a good summary is still not the end product in most financial workflows.

An underwriter reviewing a credit file, a CFO examining financial results, and a relationship manager reviewing client interactions all need more than summaries of information they already have. They need to know what changed, what is unusual, what information is missing, where the evidence conflicts, which assumptions are driving the conclusion, and which questions should be answered before the decision moves forward. That is a much higher bar than enterprise search with a language model attached to it.

I learned this very directly while working with balance-sheet optimization. An optimization engine can produce a mathematically attractive allocation, but no serious executive simply accepts the number because the algorithm says it is optimal. The conversation immediately turns to the assumptions, including how changes in funding costs might affect the recommendation, which liquidity or capital constraints are binding, and what the decision means for risk-weighted assets (RWA), interest expense, and other balance-sheet objectives. Executives also need to understand the trade-offs involved and the scenarios that could change the recommendation. The analytical output becomes valuable when it helps management navigate the trade-offs, not when it simply produces an answer faster.

I believe that is the right mental model for enterprise AI as well. The copilot taught the industry a new way of interacting with information. The harder and more valuable step is to make the intelligence underneath it shared rather than rebuilt inside every application, anchored to the decisions it is meant to improve, and delivered inside the workflow with the institution’s controls intact. A client view should explain what changed and why it matters. A retrieval layer should surface what is missing and where the evidence conflicts, not only what is relevant. None of that requires another destination for employees to visit. The real objective is decision readiness: reducing the distance between a messy information environment and a decision that can survive informed scrutiny.

Related Stories

November 19, 2026  |  In Person

Cincinnati Leadership Summit

Westin Cincinnati - Downtown

Similar Topics
Artificial Intelligence
Data Management
Diversity
Testimonials
background imagebackground image
Community Network

Join Our Community

starElevate Your Personal Brand

starShape the Data Leadership Agenda

starBuild a Lasting Network

starExchange Knowledge & Experience

starStay Updated & Future-Ready

logo
Social media icon
Social media icon
Social media icon
Social media icon
About