Next in Tech Ep. 197: Data Pipelines for AI

Eric Hanselman hosted me on S&P Global’s Next in Tech to argue that enterprise AI success depends on data pipeline quality more than model scale, and that the infrastructure patterns for getting it right already exist in the DevOps playbook.

The DevOps moment for data

Enterprise AI data infrastructure is in the same place software delivery was a decade ago. DevOps introduced continuous integration and delivery to replace brittle manual deployment, and data teams are now building pipeline automation for model training, evaluation, and governance. The cultural shift matters as much as the technical one. Organizations have to treat data delivery with the same rigor they eventually brought to code delivery.

Start small, iterate locally

I advocate starting with smaller models and localized datasets. The approach is the same one I have run since Chef and Amazon: prove the pattern works at small scale, measure what matters, then expand. Enterprises that try to solve the data pipeline problem at full scale first build fragile architectures and burn through budgets before they learn what actually works.

Pipeline quality as competitive advantage

The companies that win at enterprise AI are the ones that build the best data pipelines. Clean data in. Reliable inference out. Governance and cost controls at every stage. The data delivery layer is where the next generation of developer tools and infrastructure companies will be built.