Synopsis: The early generative AI build-out moved fast because the work that mattered most could be done in isolation — pick a model, write a prompt, plug in an API. That phase is over. The work now is harder and slower, because the bottleneck isn't the model. It's the state of the data that surrounds it. Recent industry research has 87% of IT leaders admitting that poor data readiness is the single biggest reason their AI projects aren't reaching production, and that gap isn't going to close by swapping in a smarter foundation model.
Shirshanka Das, co-founder and CTO of DataHub, sits down with Mike Vizard to make the case that the industry has been optimizing the wrong layer of the stack. Tuning prompts, evaluating new models and stitching agents together gets the headlines, but the messy, unglamorous work of curating organizational context — what the data means, who owns it, how fresh it is, what’s safe to expose — is what actually determines whether an AI system produces useful answers or confident hallucinations.
Das gets into the practical mechanics of fixing that. His argument is that context engineering can’t stay siloed inside individual AI teams; it has to be elevated into a shared, centralized layer that any agent in the organization can consult. That means crowdsourced metadata, governed semantic definitions and a living catalog of business context that travels with the data rather than living in a slide deck.
The broader thread is what enterprise AI ends up looking like once it’s treated as digital labor rather than novelty. Agents need durable memory, trustworthy context and clear accountability for the data they act on — and the organizations getting that infrastructure right are the ones whose AI initiatives will actually graduate from pilot to production.
