Connected intelligence: Decision

Nearly every organization now has access to frontier AI models. OpenAI, Anthropic, Google, and others continue to improve rapidly, and for many common enterprise tasks, their capabilities are converging. So why are outcomes still so uneven?

Even the most powerful model can only work with the information it receives. And in high-stakes, regulated environments, getting that information right, getting it there in the right structure at the right moment, verified and auditable, is an engineering challenge that rivals the models themselves in complexity.

The gap between powerful and reliable

When a large language model (LLM) generates a response, it works with whatever information is in its context window: the finite set of tokens it can process at once. Current frontier models offer context windows of a million tokens or more. That sounds enormous until you consider what a complex financial workflow actually requires: ratings, research, financials, ownership data, and regulatory filings assembled across dozens of entities simultaneously. At that scale, context fills fast, and what gets included, excluded, or compressed directly determines the quality of the output.

The industry has learned, often through expensive failure, that the real engineering challenge extends well beyond the model. It encompasses the configuration of everything surrounding it: the retrieval pipelines that find the right evidence, the evaluation frameworks that verify the output, the governance infrastructure that makes it auditable, and the structured domain knowledge that determines whether a given piece of information is relevant, current, and authoritative. Andrej Karpathy, an OpenAI co-founder, offered a useful analogy: If the LLM is the CPU, the context window is RAM, and context engineering is the operating system that decides what to load into working memory and when.

This is the infrastructure that bridges the gap between AI that sounds right and AI you can act on. In regulated industries, where yesterday’s data can be actively dangerous and every output must be defensible, it is where the difference between a powerful model and a decision-grade outcome is made.

Most organizations will spend years building this infrastructure. Moody’s already has.

What connected intelligence delivers

At Moody’s, we call this infrastructure connected intelligence: the systems, domain knowledge, and governed data architecture that determine which information reaches the model, in what structure, at what moment, and subject to what constraints. It is what transforms a capable model into a decision-grade system.

Connected intelligence is built on four subsystems, each independently engineered and continuously refined:

Underpinning all of it is a unified knowledge graph spanning more than 600 million entities and 2 billion ownership links; interconnecting credit ratings; research; financials; ownership structures; news sentiment; environmental, social, and governance data; and risk signals. This is not a data warehouse. It is a living architecture of interconnected risk knowledge that is continuously updated, rigorously validated, and encoded with the interpretive frameworks that only a century of credit and compliance experience can produce.

That knowledge base is the piece that cannot be replicated. A retrieval pipeline can be rebuilt. The connected intelligence it draws on, accumulated over decades of real-world deployment and refined through continuous customer feedback, cannot.

Why financial services demand this

Every context engineering challenge is amplified in financial services across four dimensions.

1. Temporal validity. A refinancing reported this morning can invalidate last week’s debt ratios entirely. Yesterday’s data is not merely stale; if it contradicts current reality, it is actively dangerous. Connected intelligence must detect when new filings, ratings actions, or market events have been published and update the retrieval index before the next query arrives.

2. Regulatory auditability. The EU AI Act imposes direct obligations from August 2026. The US Securities and Exchange Commission’s 2026 examination priorities explicitly include reviewing whether firms can demonstrate how AI-driven decisions were reached. The UK Financial Conduct Authority’s Mills Review is expected to deliver practical guidance on audit trails and explainability by the end of 2026. If an AI system cannot explain why it produced an output, it cannot be deployed at scale in regulated markets.

3. Cross-entity complexity. An issuer’s credit profile connects to sector risk factors, which connect to macroeconomic conditions, which may be subject to regulatory actions that vary by jurisdiction. A general-purpose retrieval system has no way of knowing that a change in sovereign risk methodology should trigger reevaluation of every corporate issuer in that jurisdiction. Connected intelligence built on a century of accumulated domain experience does.

4. Stakes of the output. In general-purpose enterprise AI, a flawed response is an inconvenience. In financial services, the cost of error is financial, legal, and regulatory exposure. Connected intelligence in this domain is not a technical enhancement. It is risk management.

Moody’s Agentic Solutions are the product layer that activates connected intelligence, and we designed them to meet customers where they already work.

Through Anthropic’s Claude, Moody’s became the first financial services firm in the world to launch an MCP app, with fully built credit and compliance workflows running natively inside the Claude environment. Through Microsoft, Moody’s MCP servers and a dedicated Moody’s agent are embedded directly inside Microsoft 365 Copilot, Researcher, and Excel. Via OpenAI, Moody’s MCP servers are available inside ChatGPT Enterprise. Through AWS Marketplace, Moody’s Agentic Credit Memo workflow is deployable directly into existing cloud infrastructure. And via the Databricks Data Intelligence Platform, Moody’s GenAI-ready data is available to teams building and scaling AI solutions on Databricks.

No new platform. No new login. Decision-grade intelligence, wherever you choose to work.

The model layer is commoditizing. IBM’s chief AI architect has described 2026 as a “buyer’s market” where the model itself is not the main differentiator. The question for enterprise leaders is no longer which model you have access to. It is whether the intelligence surrounding that model is engineered to make its output high quality, auditable, and grounded in the most authoritative data available at the moment of decision.

That is exactly what Moody’s connected intelligence delivers.

Read the full white paper: Connected intelligence: Where enterprise AI actually gets built

Learn more about Moody’s Agentic Solutions.

For more information on how Moody’s can support your risk and compliance processes, including automated screening that leverages AI, please get in touch – we would love to hear from you.