Databricks and Infosys say the gating factor for enterprise AI is no longer model quality but data readiness, with successful deployments now requiring 92% output precision as a baseline. Bavesh Patel, senior vice president at Databricks, and Rajan Padmanabhan, unit technology officer at Infosys, made the case in an April 27, 2026 episode of MIT Technology Review's Business Lab, produced in partnership with Infosys Topaz. Their argument: consumer AI has reached hundreds of millions of users, but enterprises are stalling because their data is fragmented, ungoverned, and stripped of context.
Padmanabhan put the bar plainly. "The successful customers, definitely for the precision to be more than 92% is not aspiration, that is a must-have," he said, framing the threshold as a precondition for AI to drive buy, sell, and recommendation decisions inside a business. Below that line, the outputs are not trustworthy enough to act on.
Patel said the precision problem traces back to data, not models. "The quality of that AI and how effective that AI is, is really dependent on information in your organization," he said. In most enterprises, that information sits locked inside proprietary SaaS applications, legacy systems, and disconnected dashboards, with no unified view of what exists, how fresh it is, or who can use it.
Key facts
- 01Infosys' Rajan Padmanabhan says enterprise AI customers now treat 92% output precision as a must-have, not an aspiration.
- 02Databricks' Bavesh Patel argues most enterprise data is locked in proprietary SaaS apps and disconnected systems, starving AI of context.
- 03Consumer AI has reached hundreds of millions of users, but Padmanabhan says enterprise adoption is stalling on precision.
- 04The discussion ran on MIT Technology Review's Business Lab on April 27, 2026, in partnership with Infosys Topaz.
- 05Patel frames the fix as moving data into open formats with cataloging, governance, and access controls before any AI project ships.
His blunter version: feed an AI system poor data and "you're actually going to end up having terrible AI." Patel argued the competitive edge in enterprise AI is not the foundation model a company picks but the proprietary data it can route into that model, plus the third-party data layered on top. Without consolidation into open formats and a working catalog, that edge stays theoretical.
“Successful customers treat 92% output precision as a must-have, not an aspiration, according to Infosys unit technology officer Rajan Padmanabhan.”— Jaeden Schafer
Padmanabhan drew a sharper line between consumer and enterprise AI. Consumer chat tools work because they were trained on the open internet, while enterprise AI needs internal context that is structured, unstructured, and user-generated all at once. "That context from your enterprise information, which is not only structured, both structured and unstructured and user-generated contents and all forms of data is going to be very, very critical," he said.
Patel pointed to ChatGPT as the contrast case. OpenAI aggregated the public web, synthesized it, and built transformer models on top. Most enterprises cannot even produce an inventory of their own data, let alone a connected graph of how datasets relate to one another. The first step, he said, is mapping the data estate, then putting cataloging, lineage, and governance around it.
Both executives pushed back on the instinct to launch AI pilots before the foundation is in place. Patel said customers are moving away from ad hoc experiments and toward projects tied directly to business metrics, with governance frameworks that decide what scales and what gets killed quickly. The deciding factor is whether a use case produces measurable value, not whether it demos well.
Padmanabhan described where this is heading. "What we are seeing as a new way of thinking is moving from a system of execution or a system of engagement to a system of action," he said, referring to AI agents that handle workflows and transactions rather than just answering questions. That shift raises the stakes on data quality further, because an agent acting on bad context produces bad outcomes at machine speed.
There is also an internal education gap. "We see this big opportunity just with AI literacy with business users, where they're very eager to understand how they should be thinking about AI," Patel said, noting that line-of-business teams want to know the building blocks underneath the hype before they commit budget. Training and enablement, in his telling, sit alongside infrastructure as a precondition for adoption.
The skeptical read is that this is a familiar pitch from a data platform vendor and a systems integrator: the data is messy, and they sell the cleanup. The 92% precision figure is a customer benchmark Padmanabhan cites rather than a published study, and the conversation runs as branded content with Infosys Topaz. Enterprises will still have to decide whether agentic workflows actually clear that bar in production or whether the threshold quietly slides.
Even with the caveats, the framing matches what large buyers are signaling in procurement. The bottleneck for AI in 2026 is not access to capable models, which is now a commodity discussion, but whether the data underneath them is structured, governed, and trusted enough to let an autonomous agent act. Vendors that solve the boring half of that problem are likely to capture more enterprise spend than the ones still racing benchmark scores.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
