Data Readiness & Architecture Modernization

Stop boiling the data lake before the first use case ships

The most expensive mistake in AI data readiness is the two-year "modernize everything first" program that delivers a pristine platform and zero business outcomes. The opposite mistake, wiring a model directly to whatever tables happen to exist, produces a demo that works once and a production system that hallucinates on stale, undocumented data no one owns. Executives need a middle path: a minimal viable data posture scoped to the decisions the AI will actually touch.

Minimal viable does not mean minimal quality. It means you resist the urge to govern all 4,000 tables and instead pick the 8 to 12 data products that feed your first two or three use cases, then hold those to a real standard. A CFO funding an AI forecasting assistant should ask exactly one question first: which data products does this depend on, and does each have a named owner and a freshness SLA. If the answer is no, the model's accuracy is a coin flip dressed up in a chart.

Data products, not datasets

A dataset is a table someone left lying around. A data product is a table with an owner, a service level, a contract with its consumers, and a quality gate that fails loudly. The shift from datasets to data products is the single highest-leverage move an executive can sponsor, because it moves accountability for data quality from "the data team, generally" to a named person for each product.

A data-product contract spec you can enforce

Every data product feeding an AI system should ship with a contract. The contract is not a document that lives in a wiki; it is a set of fields you check in the pipeline and review on a cadence. Below is a minimal spec that covers the fields that actually cause incidents when they are missing.

The consent-and-preferences row is deliberately in the same table as the rest. Privacy is not a separate workstream you bolt on later; it is a data product with an owner and a gate, exactly like the others.

Sequencing the posture without a two-year program

Work outward from the use case, not inward from the platform. The goal is a defensible posture in one quarter, not a finished architecture.

• Pick the two or three use cases with the clearest P&L link, then inventory only the data products they depend on. Ignore the rest for now.
• Assign a named owner and a written contract (owner, SLA, freshness, quality gate, consumer) to each of those products, and enforce the gate in the pipeline so bad data fails visibly.
• Adopt an append-only event ingestion pattern for new sources: capture raw events immutably first, transform downstream. This prevents the regret of lossy transforms you cannot replay.
• Stand up metadata and lineage as the control plane so you can answer, for any model output, exactly which sources and versions produced it. Lineage is what makes incidents diagnosable.
• Define PII and PHI safe zones with tokenization or masking at ingestion, so models train and infer on protected data only inside a controlled boundary with logged access.

Where data-readiness programs burn budget

• Modernizing the whole estate before shipping anything. Fix: scope readiness to the products your first use cases need, and let the rest wait for a real consumer.
• Data products with no owner. "The platform team owns it" means no one does. Fix: put a single named person on each product's contract.
• Transforming events on the way in and discarding the raw feed. Fix: land raw events immutably first, so you can reprocess when the model or schema changes.
• Treating lineage as a nice-to-have. Without it, every model incident becomes an archaeology project. Fix: make lineage a launch requirement, not a backlog item.
• Bolting privacy on after the model works. Fix: mask or tokenize PII at ingestion and treat consent as a first-class data product with a mandatory gate.

Data posture wins for the next 90 days

• Publish a one-page list of the 8 to 12 data products your first use cases depend on, each with a named owner.
• Write and enforce one real data contract, including a quality gate that fails the pipeline on violation.
• Turn on append-only capture for at least one high-value event source you were previously transforming on ingest.
• Stand up lineage for one end-to-end path so you can trace a single model output back to its sources.
• Define one PII safe zone with masking at ingestion and logged, role-scoped access for model workloads.

Related advisory guides

Go deeper on this theme with the detailed guides in this collection.