Anushrut Gupta is the applied AI lead at PromptQL, developing enterprise AI that can interpret messy organizational data, learn company-specific terminology, and produce answers through inspectable, controlled execution. His approach addresses a central obstacle to deploying AI inside businesses: important information is scattered across systems, while the definitions needed to interpret it frequently remain undocumented or change.
Gupta studied computer science at the Indian Institute of Technology Roorkee and earned an MBA from the Wharton School. In 2017, he interned with Intuit’s Innovation and Advanced Technology team, subsequently working at Intuit Research as a software engineer and product manager. He later became a senior product manager at Hasura, joined its AI Lab, and helped develop natural-language access to enterprise data into PromptQL.
- Data readiness cannot be a prerequisite. Gupta argues that companies cannot finish cleaning every database or documenting every business definition before deploying AI: schemas evolve, financial calendars vary, and terms such as customer acquisition cost or at-risk accounts mean different things across teams. His AI Engineer World’s Fair talk challenges static semantic layers and knowledge graphs that require exhaustive maintenance.
- An agentic semantic layer learns through correction. Gupta envisions an AI analyst that improves as colleagues explain ambiguous table names, reconcile customer identities, or clarify that department budgets are denominated in cents. Its understanding of company-specific language develops through practical work instead of depending on a supposedly complete, predefined business taxonomy.
- Deterministic execution separates planning from answers. A language model generates an editable plan in a domain-specific language, while a separate runtime queries enterprise systems, performs calculations, and returns results. Gupta’s account of the architecture emphasizes fine-grained authorization, execution outside the model context, and targeted model calls for classification or extraction. Customer credits and other consequential actions remain subject to human oversight.
- Shared-context AI workflows preserve continuity. Gupta has extended these ideas into persistent threads where people and AI collaborate on bug fixes, feature development, analytics, and investigations. In his description of replacing a conventional to-do list, work proceeds across concurrent shared contexts without repeatedly reconstructing background or handing tasks between disconnected tools.
