Three months ago, the conversation I was having with enterprise technology leaders was about which model to fine-tune. Today, it’s about why the pipeline feeding that model is the reason their AI project is six months behind schedule. Or, more important, why adding more pipelines and more cloud capacity is not translating into measurable value from AI in production.
This is not a coincidence. The 2025 enterprise AI architecture—vector databases, RAG layers, orchestration frameworks and ingestion pipelines pulling from operational systems—was built on an assumption that does not survive contact with production: that enterprises can keep moving data fast enough to make AI agents useful in real time and then reconstruct governance downstream after every move.
That assumption came from pre-AI blueprints. It was like adding more horses versus building horsepower. AI in production needs brake-horsepower infrastructure that puts data and AI together in real time in a sovereign infrastructure, not in different places.
The next generation of successful enterprise architecture starts and finishes at the engine inside that vehicle: the data layer.
This is the new world of an engine, where all the parts fit and work together in real time. It’s not a set of fragmented pieces glued together with great intent but optimized only for reducing drag and friction—it means building a whole new sovereign systemic design for AI success.
That old assumption cannot hold. And the data layer is where it is breaking first.
The pipeline tax no one has on their balance sheet
Look at the architecture most large organizations actually run. Transactional systems feed pipelines, which feed warehouses, lake houses, feature stores and models. Each hop is a translation. And each translation is a place where governance policies have to be reapplied, lineage gets murky and a masking rule defined in one system can silently fail to propagate to the next.
By the time data reaches an AI agent, it may have been copied four times and governed by three regimes, none of which fully agrees with one another. Then a regulator asks a simple question—“Can you show me where this customer’s data went and who touched it?”— and the answer takes six weeks and a consulting engagement.
This is the pipeline tax. It does not appear as a line item in any budget, but it shows up as audit findings; AI hallucinations; stalled migrations; and the reason why 95% of enterprises say they want to operate as their own sovereign AI and data platforms, while only 13% report they are actually thriving at it. Those figures come from EDB’s recent customer research, but the broader pattern is visible across the market: Gartner has tied GenAI project abandonment to poor data quality, inadequate risk controls, escalating costs and unclear business value. And McKinsey’s 2025 State of AI survey found that AI adoption is broadening, but most organizations have not yet scaled the technology into enterprise-wide impact.
The retreat from 2025’s architecture is happening at scale—and fast
The market has started to figure this out. The retreat from the RAG infrastructure enterprises spent 2025 building is real—VB Pulse found that organizations that “went wide on RAG in 2025” are now hitting a common failure point: architectures built for document retrieval do not hold at agentic scale. Single-method vector similarity is no longer enough for production agentic workloads that require accuracy, access control and context across systems.
Vector database categories are shifting as a result. The issue is not that retrieval is going away; it is that the simple RAG-to-vector-database pipeline is being rebuilt for a different era of AI. Hyperscalers are beginning to rebuild their data stacks around agents rather than pipelines. Even lake house incumbents are publishing research arguing that when queries span databases and documents, stronger models alone do not fix the problem—architecture does.
What is missing from most of those stories is the next move. If pipelines are the problem, what replaces them?
Always-on-governance is the new model, at the data layer
The architectural answer now forming is straightforward: Stop moving the data and bring agents and AI to the data. Governance should live inside the data layer by design, not be bolted onto every downstream system after the fact.
Treat governance as a property of the architecture itself. Think of it like the human body: Organs perform different functions, but they are interdependent and governed by the same system 24x7x365. Enterprise AI needs the same principle. Different systems and agents may serve different purposes, but they have to operate from the same rules of governance, policy and sovereign control.
The pieces required to do this are no longer speculative. Postgres®, where much of the enterprise’s operational data already lives, can serve as a governance control plane, with row-level security, column masking and lineage native to the engine. Apache Iceberg has won the open table format argument. The Model Context Protocol gives AI agents a standardized, governed way to reach operational data without requiring a custom integration for every application.
None of this is a 2027 roadmap conversation. It is a procurement conversation happening now.
Migration is a capability, not a project
The same logic applies to the modernization backlog blocking everything else. Migration has historically been treated as a project: scope it, staff it, suffer through it and deliver it 18 months late.
The reason it remains painful is that the work itself—discovering schema dependencies, translating embedded business logic, validating functional equivalence—is exactly the kind of high-context, repetitive reasoning that coordinated AI agents are now genuinely good at.
The COBOL-translation demos getting attention this year are the leading edge of something larger: migration becoming an autonomous, continuously running capability rather than a one-off program. That changes the unit economics. It also changes the strategic question. The interesting question is no longer, “How long will this Oracle migration take?” It is, “How quickly can we evolve our entire platform strategy?”
The data layer is where the next decade gets decided
The vendors that win the next decade of enterprise infrastructure will not be the ones with the fastest query engine or the slickest notebook experience. They will be the ones that recognize data movement is breaking enterprise AI.
The pipeline tax has been paid long enough. The interesting work now starts at the data layer—and it starts when enterprises stop paying that tax.
The pipeline model breaks at agent scale. It was born of good intent, but in a world moving toward 1 billion agents delivering 217 billion instructions a day, it is architecturally medieval. The agentic era will be won at the data layer.
