Informatica World 2026: why your AI is an overconfident intern

This keynote from Informatica World 2026 was delivered by Pratik Parekh, President of Product Management, Data and AI Engineering at Informatica. The conference took place in Las Vegas in May 2026. The talk is about 45 minutes and is aimed at product and data leaders in enterprise organizations.

What it covers

Parekh opens with a provocation: enterprise AI in 2026 resembles a very smart intern with excellent credentials who has just arrived at a new company. The intern produces confident output, but the output is often wrong in ways that are hard to detect — because the intern does not know the organization’s taxonomy, history, or internal vocabulary. The analogy is direct and the keynote builds its entire argument from it.

The core claim is that AI agents already work at a technical level. The models are capable. What makes them fail in production is not model quality but data quality — specifically, the absence of clean, well-governed organizational context. When an agent does not know that “customer” means something different in your billing system versus your CRM, or that a product line was renamed two years ago, it produces plausible-sounding answers that are factually wrong about your business.

The keynote walks through Informatica’s response to this problem: an AI-ready data platform that prepares organizational data for agent consumption. The demonstrations include automated metadata enrichment, lineage tracking for AI outputs, and governance controls that allow enterprises to define what agents can and cannot access.

Who it is for

Primarily useful for PMs working on enterprise AI products or internal AI tooling. The keynote does not assume a technical audience — Parekh explains concepts clearly — but it assumes familiarity with enterprise data challenges. Consumer product PMs may find the context less directly applicable.

Key takeaways

1. An AI agent’s confidence is not a signal of correctness. Overconfidence in the absence of organizational context is a predictable failure mode, not a bug in the model.

2. Data preparation is a product problem, not just an engineering problem. PMs building AI features need to account for the quality and structure of the data the AI will actually use — not just the model’s benchmark performance.

3. Governance and access control are not obstacles to AI deployment; they are prerequisites. Teams that treat them as afterthoughts build AI features that fail in ways that erode user trust.

4. The gap between a demo and a production-ready AI feature is often a data gap. A model that performs well on clean, labeled examples may perform poorly on an organization’s actual data.

Worth watching if your team is planning to build AI features on top of internal enterprise data, or if you are evaluating why a previous AI initiative did not deliver expected results in production.