Most organizations that struggle with AI don't have a model problem. They have a data problem that AI makes visible. A language model can only reason over the context you give it, and a prediction model is only as good as the features it's trained on. If the underlying data is late, inconsistent or locked in silos, no amount of prompt engineering will fix it.
1. Start with a lakehouse on open formats
Open table formats such as Apache Iceberg or Hudi give you warehouse-style reliability — ACID transactions, schema evolution, time travel — on inexpensive object storage. Crucially, they let multiple engines (Spark for transformation, Trino for interactive queries, Python for ML) work on the same data without copying it.
2. Treat data quality as a product requirement
Define data contracts at the boundary between producers and the platform. Validate schemas, null rates, uniqueness and freshness on every load. Route failures to the team that owns the source, not to the analyst who notices the broken dashboard three days later.
- Freshness SLAs per critical dataset
- Automated tests on every transformation
- Column-level lineage from source to report
- Clear ownership for every domain
3. Make unstructured data first-class
Generative AI is mostly about documents, tickets, emails and knowledge bases. An AI-ready platform ingests these alongside structured data, tracks their permissions, and produces chunked, embedded representations as managed, versioned datasets — not as a side script someone runs on a laptop.
4. Build a semantic layer
When an AI assistant answers "what was revenue last quarter?", it should use the same definition of revenue as the finance dashboard. A governed semantic layer gives both humans and AI one source of truth for metrics, and dramatically improves the accuracy of natural-language analytics.
5. Govern access end to end
If a user can't see a document in the source system, the AI shouldn't be able to retrieve it for them either. Propagate access controls into your retrieval layer, mask PII where it isn't needed and log every access. This is what turns a promising pilot into something your security team will approve.
The fastest path to useful AI is usually a better data platform, not a bigger model.
None of this requires a multi-year program. Start with one high-value domain, build it properly end to end, and use it as the template for the next.