Agentic AI
,
Artificial Intelligence & Machine Learning
,
Data Security
CIOs Must Match Architecture to Workloads, Governance and Business Context
Artificial intelligence is forcing enterprises to reconsider data platforms they spent years modernizing.
See Also: How to Build Faster, More Scalable AI Applications with Elastic and NVIDIA
Cloud data warehouses built for dashboards, financial reporting and business intelligence are now being asked to support AI models and supply autonomous agents with the information needed to make operational decisions. Those workloads require faster processing, stronger governance and a deeper understanding of what enterprise data means.
CIOs must decide whether to extend the warehouses they already have, move toward a data lakehouse or assemble a more specialized architecture. For many organizations, the answer will depend on the workloads they expect to automate and the complexity they can realistically manage.
But architecture alone won’t make enterprise AI work. Companies also need semantic and governance foundations that help AI systems select the right data, apply consistent definitions and understand how information relates to business processes.
“The LLM, everybody can get,” said Bharat Bansal, a partner in Bain & Company’s technology and data practice. “It’s your data” that will distinguish one company’s AI strategy from another.
Your Modern Platform Isn’t So Modern Anymore
The architecture problem begins with the variety of work companies are now trying to place on a common data foundation, and agentic AI is showing the cracks. Agents retrieve information, make decisions and act on them. A data-quality problem that would produce an incorrect dashboard could cause an autonomous agent to make purchases or supply-chain decisions before a human intervenes. Data quality and access controls, therefore, are becoming operational safety requirements, Bain said in a recent report on re-architecting the data platform for the AI era.
Bansal sees two common weaknesses in existing data environments: the divide between structured data held in warehouses and less-governed unstructured data stored elsewhere, and the lack of business context available to agents.
Companies may already have data dictionaries and catalogs, but an agent needs more than the definition of a customer or product to make a decision. It must understand that a customer may have subsidiaries in different regions, purchase several products and interact with the company through multiple channels.
“The traditional semantic layers don’t provide that context to our agents,” Bansal said. Much of that knowledge still resides in employees’ heads.
Warehouse and Lakehouse Capabilities Are Converging
To get data AI-ready, Bain advises most enterprises to extend their existing cloud data warehouse rather than replace it.
Platforms such as Snowflake, Google BigQuery and Amazon Redshift have added open table formats, machine learning integrations and vector-search capabilities. For organizations primarily running business intelligence, reporting and early-to-moderate generative AI applications, extending the warehouse can provide a faster and less complex route forward, Bain said.
A full lakehouse may be a better fit for organizations with substantial machine learning in production, real-time data requirements or teams that need to work from a common governed storage layer. Bain’s research finds that highly specialized, best-of-breed architectures may be appropriate for only 5% to 10% of organizations because of the integration, governance and financial management work involved.
“You don’t want the most sophisticated platform and architecture,” Bansal said, if it’s too complicated for the company’s needs or exceeds its ability to operate it.
Noel Yuhanna, vice president and principal analyst at Forrester, said he considers the lakehouse to be foundational for long-term AI-ready data architecture. A lakehouse can support SQL analytics, data science, machine learning and AI on a common platform, while open formats make it easier to use data without committing the entire environment to one vendor, he said. But organizations shouldn’t just dismantle and replace the systems they already have in place.
“You can’t really throw away existing stuff,” he said. “You’ve got to take what you have already in existence and see if your platform could actually meet the requirements.”
AI and the Semantic Layer
In a data architecture, the semantic layer sits between underlying data and the applications, analytics tools and AI systems that consume it. It establishes common definitions and provides the business context needed to interpret enterprise information.
For example, while an employee may be able to understand that different teams mean different numbers when they say “sales,” an agent needs more information to determine whether it means net sales minus cancellations, as a finance team may mean, or if it’s a gross sales number used to generate commissions for sales.
Josh Fecteau, chief data and analytics officer and CIO at Teradata, said this distinction helps explain why AI ROI is often easier to see at the individual level than at the enterprise level. An employee using a chatbot can add context or correct a misunderstanding. An enterprise agent operating independently must locate the right data and understand how it should be used.
“You have to have a fully described set of data and truth out of the gate,” he said.
That doesn’t mean all data must be physically consolidated. But the company needs consistent metadata, definitions and governance across its environment.
“If you have too many sources of truth, you have no source of truth,” Fecteau said.
Build the Foundation Through Business Use Cases
Installing a catalog, lakehouse or semantic platform won’t resolve disputes over data ownership.
CIOs can provide the technology to implement business processes, but department leaders must decide who owns the data underneath. They must also decide who owns what, and who is responsible for maintaining that data and those definitions as workflows evolve.
“You can have the best platform, but nobody will invest the resources behind it” unless ownership is clear, Bansal said.
Companies are also underinvesting in data engineering and architecture, he said. They may hire machine learning engineers but leave them cleaning data and constructing pipelines, limiting the return on expensive AI talent.
Bansal said he recommends attaching data-platform investments to specific AI use cases rather than asking a CFO to fund a broad initiative to improve data quality.
If a company wants to use AI for customer support, pricing or supply-chain management, it should identify the data required and then make sure that data is AI-ready.
“The value is in the customer support business case,” Bansal said. The semantic layer and data pipelines become necessary components of the investment rather than an open-ended platform program.
Teradata used a similar approach when it developed AI applications around its contracts. The company converted complex contract information into data that could be shared across agentic applications and built one use case on top of it.
Fecteau said that application produced about 100,000 hours of capacity savings. Teradata then reused the underlying knowledge to build additional applications, including a sales intelligence agent combining contract and customer data.
That return then compounds as multiple divisions and applications can benefit from work done to support that initial project.
Finding examples where you have fully described data sets that you can get value from today or tomorrow creates the proof points needed to justify further investment, Fecteau said. “It’s not just a proof of concept. It’s an actual use case that’s running in production, where you can say, ‘Well, here’s the proof.'”
Click Here For The Original Source.
