BI needed a warehouse. AI needs a gateway.

Business Intelligence turned data into knowledge so that humans could close the loop. Data became dashboards, dashboards informed decisions, and people translated those decisions into action.

AI agents flip that model. Instead of producing insights for someone else to execute, they execute. Every action carries just enough reasoning in the moment - small analytics over a handful of records - to relate these accounts, reconcile an order against the ticket blocking it, or read the contract before answering the customer. Intelligence no longer lives in a separate analytical layer. It lives where the work happens.

For two decades we built data warehouses because people needed a consistent view of the business. We centralized data, standardized metrics, and built semantic layers so everyone looking at revenue, or churn, or pipeline, saw the same answer. The warehouse became the company’s shared source of truth for analytics. It didn’t just make analytics faster - it made analytics consistent.

AI inherits exactly the same requirement. The difference is that the warehouse is no longer enough.

An agent rarely answers questions using analytical data alone. It needs the CRM, the ticketing system, the ERP, the application database, recent events, documents, conversations, and whatever system currently owns the truth for the task it’s performing.

The industry’s embrace of MCP makes this concrete. Anthropic introduced the Model Context Protocol in November 2024 as an open standard for connecting AI to external systems. One year later it reported more than 10,000 active public MCP servers and over 97 million monthly SDK downloads. AI is no longer expected to operate on a curated snapshot of enterprise data. The workforce is already wiring it into live operational systems at runtime, with or without a strategy.

Connecting is the easy part. The hard part is everything data teams already know.

Point an LLM directly at a CRM and it burns tokens exploring schemas, chooses the wrong fields, misses business rules, or retrieves records it shouldn’t even see. These are the same problems data teams solved for the warehouse: performance, governance, and trust. The work still has to be fast and affordable, access still limited to the right identity, the answers still correct and consistent. None of that disappeared when AI arrived - it moved closer to the operational systems where AI now works.

This is where the BI model breaks down. Building BI meant centralizing data. Building AI means preserving the authority of distributed systems.

Operational systems contain things a warehouse intentionally abstracts away.

• Business logic. The workflow that determines whether an opportunity is actually qualified lives inside Salesforce automation, not in a flattened status column.
• Permissions. The source system knows who can see or modify a record. Downstream copies approximate those rules, but they are rarely identical.
• Operational reality. The latest transaction, the new field someone added last week, the event that arrived thirty seconds ago - these often exist only in the source.

Warehouses were designed to transform data for analysis.

That is why enterprise AI needs a data gateway. The warehouse centralized three responsibilities. The gateway has to perform those same responsibilities across many systems instead of one.

• Access becomes federated. Instead of moving every dataset into one platform, the gateway reaches into the system where the authoritative data already lives. The warehouse remains an important source for standardized analytics, but it becomes one connection among many.
• Governance can no longer assume there is a person in the loop. Agents act autonomously or on behalf of users. The gateway has to enforce policy consistently across every system it fronts while preventing credentials and sensitive information from leaking into prompts or responses.
• Context becomes the hardest problem. BI standardized answers to questions people knew they wanted to ask by building semantic layers. AI has to answer questions nobody modeled ahead of time. An agent has to discover, at runtime, which systems matter, which records belong together, which business rules apply, and which procedures should be followed. Two people asking different questions about the same customer should still receive answers that agree, even though neither path was designed in advance. That is a much harder standardization problem than BI ever faced.

The gateway becomes the checkpoint where that happens. Every agent finds its route through enterprise systems at runtime, in response to a user’s goal - and still has to be performant, governed, and trustworthy on every path it takes.

This architectural shift is what we’re focused on at MarcoPolo. We’re working with data and AI teams trying to connect agentic AI to enterprise systems without giving up the performance, governance, and trust that took years to build in the warehouse era. I’d love to compare notes with others wrestling with the same transition.

Sources

• Anthropic, “Introducing the Model Context Protocol”, November 25, 2024.
• Anthropic, “Donating the Model Context Protocol and establishing the Agentic AI Foundation”, December 9, 2025.