Share with your CDO
Brickworks, the Australian building products manufacturer, has flipped the usual sequence: instead of cleaning data before deploying AI, the company is using AI agents to accelerate the data quality work itself. Agents analyze similar records and recommend values for missing fields, with domain experts validating before anything gets written. The platform runs on Snowflake as the central repository, Boomi for connectivity, and domain-level quality dashboards that make accountability visible. General manager James Cosier describes it plainly as “AI in service of data quality” rather than the other way around.
What this means for your business
The CDO who is waiting for clean data before starting AI projects is solving the wrong sequencing problem. Brickworks’ approach reveals a self-reinforcing loop: AI accelerates data readiness, better data enables more capable AI, and the governance infrastructure built during that process (data glossaries, dictionaries, stewardship assignments) becomes the semantic layer that future AI-powered discovery actually runs on. Whether this story is about you depends on one question: does your current data quality program have an owner per domain, or does it live only inside the data team?
The part of this case study that deserves more scrutiny is the governance claim, not the technology one. Cosier is explicit that AI agents can find and recommend fixes faster than human teams, but that the accountability model, named stewards with clear domain ownership, is what determines whether those fixes stick. The recurring failure mode in enterprise data quality programs is technically correct remediation that degrades again within a quarter because no business owner felt responsible for the asset. Brickworks’ quality dashboards make that accountability visible and persistent, which is the structural move most organizations skip when they buy a data platform.
The documentation work Cosier calls “eating your vegetables” (data glossaries, field definitions, lineage records that engineers historically avoid) is quietly becoming the most strategically valuable artifact a data team can produce. As AI-powered data discovery tools mature, they consume exactly this metadata to answer business questions in natural language. Organizations that have invested in semantic documentation will compress the time-to-value on those tools significantly. The falsification condition here is straightforward: if natural language interfaces to enterprise data turn out to require less structured metadata than current architectures assume, this investment looks like over-engineering. Every indication from how large language models actually perform on enterprise queries points the other direction.
Based on reporting from State of Data & AI 2026: Scaling AI, originally published 2026-07-21 02:00:00.
