For many enterprise AI projects, the key constraint is not model size but whether the model can access relevant, current, reliable information that the organization is authorized to use. That is the central argument of Bapi Raju Ipperla’s 25 September 2026 article in The AI Journal: build the data foundation around the work AI must do, then choose and scale models against that foundation. It is useful advice, not a proven universal rule or a measured comparison of larger models with data platforms.
Why the data platform matters to enterprise AI
A model can generate a fluent answer without having the context needed to make that answer useful. In a business workflow, relevant context may live across operational systems, documents, event streams, and knowledge stores. Those sources may differ in format, freshness, ownership, and access rules. If the information is missing, stale, poorly indexed, or unauthorized, a more capable model does not by itself resolve the underlying problem.
As an Amazon Associate I earn from qualifying purchases.
A data platform can provide a governed way to make needed information available to AI workflows. In practice, that means connecting sources, managing identity and permissions, keeping information synchronized, and making the path from source to model observable. The goal is not to feed every system into every model. It is to provide the smallest useful set of current, authorized context for a particular task.
Free tools Windows power users keep installed
One-click scans. No signup required.
Ipperla’s article reports that McKinsey & Company found 88% of organizations used AI in at least one business function in 2025, while about one-third had begun scaling AI programmes across their enterprises. It separately reports that 7% had fully scaled AI organization-wide, without specifying the year for that analysis. The article also attributes to Gartner a finding that at least 50% of generative AI projects had been abandoned after proof of concept by the end of 2025, and a forecast that more than 40% of agentic AI projects would be cancelled by the end of 2027. These are figures as reported by the article; it supplies no underlying report titles, methods, or detailed denominators, so they should be treated as attributed context rather than independently verified comparisons.
#1 Best Overall
Start with the workflow, not the model
Before selecting a model or expanding a pilot, identify what the workflow must accomplish and what information it depends on. A customer-support assistant, for example, may need account records, current product documentation, and recent service events. A forecasting workflow may depend on different systems and different freshness requirements. The source map should be specific to the task rather than a generic inventory of everything the company stores.
- Select high-value workflows. Name the business process, its users, and the decision or task AI is expected to support.
- Trace required information. List the operational systems, documents, event streams, and knowledge stores the workflow needs. Distinguish essential sources from merely convenient ones.
- Record source conditions. For each source, document its owner, update cadence, known quality issues, and the access controls that apply.
- Define what good looks like. Establish how teams will judge retrieval quality, latency, freshness, failure rates, and the workflow’s business outcome before deployment.
This dependency map reveals whether the blocker is a model capability, a missing source, a synchronization delay, weak metadata, an access restriction, or a data-quality problem. It also gives teams a basis for deciding what to build first.
Design a governed context layer
A context layer is the controlled route by which a workflow retrieves and uses information from its sources. It should make relevant data discoverable and usable without bypassing existing rules. Identity and permissions need to be enforced before information reaches the model—not treated as a cleanup step after generation.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →When evaluating an implementation, assess whether it can:
- Apply identity, permissions, masking, consent, and regional restrictions at the point of access.
- Synchronize source changes and maintain indexes and metadata used for retrieval.
- Show where retrieved information came from and when it was last refreshed.
- Monitor freshness and surface failed or delayed updates.
- Support reusable foundations across workflows without granting each workflow unnecessary access.
These are evaluation criteria inferred from the article’s recommendations, not a product ranking or benchmark. The right architecture depends on the sources, security requirements, and workload; the article does not name a preferred vendor or platform.
Make freshness a workload requirement
Not every AI task needs real-time data. A workflow that summarizes a stable policy document may tolerate scheduled updates, while a decision that depends on recent operational events may not. Specify the acceptable age of information for each workflow and choose synchronization accordingly.
Streaming and event-driven architectures are relevant when new events materially affect decisions. They also add operational complexity, so they should solve a defined freshness need rather than serve as a default architecture. Measure whether the platform can meet the required latency and detect when a source or pipeline falls behind.
Operate retrieval as a data pipeline
Retrieval does not become reliable simply because documents or records have been placed in a search index. The information must be ingested, described with useful metadata, synchronized as sources change, indexed appropriately, and governed through permissions. Lineage and freshness monitoring help teams understand whether the answer was based on the right material at the right time.
When an AI answer is wrong, trace the full path before concluding that the model alone failed. Check what the system retrieved, the source’s freshness, transformations and APIs along the way, the permissions applied, and the model call itself. That evidence helps separate a reasoning error from an upstream data failure or an access-control issue.
Rank #4
Use observability to diagnose failures
Observability should cover the workflow end to end: model calls, retrieved information, APIs, transformations, permissions, and source freshness. Without those traces, teams may see a bad answer but lack the evidence to determine whether the cause was retrieval, stale source data, a failed integration, an access rule, or model reasoning.
Track measures tied to the workflow rather than model output alone. The article recommends assessing retrieval quality, latency, freshness, failure rates, and business outcomes. Those measures can reveal trade-offs: for example, whether fresher information improves a decision enough to justify the infrastructure and operating cost, or whether a slower update cadence is adequate.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →A practical 90-day sequence
Ipperla proposes the following sequence as a starting plan, not a guarantee or a timetable suited to every organization.
Best Value
| Period | Focus | Work to complete |
|---|---|---|
| Days 0–15 | Map dependencies | Identify three high-value workflows; trace their required sources; document ownership, freshness, permissions, and known quality issues. |
| Days 16–45 | Build a reusable context layer | Standardize access to core data, add streaming where freshness materially affects decisions, and establish platform-level identity, permission, and governance rules. |
| Days 46–90 | Prove trust before scale | Deploy one production workflow with end-to-end observability; measure retrieval quality, latency, freshness, failure rates, and business outcomes; use observed errors to improve the platform. |
The sequence deliberately narrows production validation to one workflow before scaling shared capabilities. If the workflow misses its targets, its traces and metrics should guide the next investment rather than an assumption that a larger model is the fix.
What company examples can—and cannot—show
The AI Journal article cites Uber in connection with event-driven and streaming architectures, Netflix for reusable internal data platforms, and LinkedIn for large-scale event-streaming infrastructure. These examples illustrate the article’s point that reusable data infrastructure can support multiple intelligent capabilities. The article provides no dates, implementation detail, measurements, or company engineering links for them, so they do not establish that another enterprise should copy a specific architecture or expect the same results.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
