Perfect, standardized enterprise data will never arrive, so reliable AI has to learn the meaning of messy data while it is being used.
2
PromptQL has an LLM generate a domain-specific plan, then runs that plan in a deterministic runtime instead of asking the LLM to generate the final answer.
3
An agentic semantic layer learns business terms, table relationships, calculation rules, and corrections from user interactions, with versioned updates that can be rolled back.
Summary
Anushrut Gupta argues that waiting for perfect enterprise data is a dead end. Tables, columns, statuses, units, customer definitions, and financial periods change across systems, while business terms such as "active customer" and "at risk" depend on local knowledge. Gupta proposes an AI-maintained semantic layer that learns this context through corrections and use. PromptQL uses a language model to create a domain-specific plan for retrieving and computing data. A deterministic runtime executes that plan across databases and applications, so the language model does not generate the final result. In the demos, the system discovers incorrect statuses, identifies badly named tables, converts cents to dollars, learns a company-specific quarter, and carries those corrections into later queries. Gupta also shows a multi-step workflow spanning customer data, Zendesk tickets, sentiment analysis, and credits issued through an API. He presents this as a way to start with a useful but inexperienced analyst and improve its business understanding over time.
Perfect enterprise data is a permanent waiting condition
Gupta says teams keep preparing data so AI can understand its meanings and relationships, but the perfect-data goal will never be reached. Tables and columns have unclear names, nulls, old fields, inconsistent statuses, and different units across systems. One system may store revenue in cents while another uses dollars. He describes earlier attempts to solve this through standardization, Snowflake, Databricks, and master data management, followed by semantic layers for agents. Each approach still faces changing schemas, workflows, and business definitions. His conclusion is that reliable AI has to work with data as it exists and adapt when that data changes.
A definition such as customer acquisition cost hides many business decisions
A semantic layer entry might define customer acquisition cost as marketing spend divided by new customers. Gupta says that definition leaves several unresolved questions. Marketing spend could mean brand or performance spending. A new customer could be someone making a first purchase or a reactivated customer. The time period, failed trials, and seasonality may also change the calculation. He argues that manually entering every edge case is impossible because the relevant business context is often implicit. The same problem appears in a sales graph: a graph can connect deals, stages, close dates, and owners, but it cannot know what "at risk" means without the company's own definition.
Gupta says the central failure is that a general language model does not speak a company's business language. "GM" might mean gross margin in finance and general manager in HR. Teams can also define conversion, quarter, and active customer differently. Human analysts handle these ambiguities because they know the business, can query different data sources, and explain their reasoning. That explanation helps colleagues trust the answer. Gupta calls this accumulated company knowledge the missing piece in current AI systems. His proposed system should behave like a new analyst: capable on day one, open to correction, and able to learn the local definitions and working habits that experienced analysts acquire over time.
PromptQL separates planning from deterministic execution
PromptQL uses a foundation model to generate PromptQL plans, which are written in a domain-specific language for data retrieval, computation, aggregation, and semantics. The plan runs in a deterministic runtime on a distributed query engine that can access different data sources. The language model is not involved in the actual execution or in writing the final answer. Gupta says that giving results back to the model to generate an answer means accepting a possible hallucination. Separating the model's planning step from execution lets the system show the work and return the computed result directly.
The system can recover from messy data during a query
In the first demo, Gupta asks for the top five customers by revenue. PromptQL initially looks for a succeeded status, then notices that the underlying data uses paid and pending. It corrects the plan and returns the customers based on the actual statuses. A larger example asks it to identify organizations from user email domains, find the third-highest-revenue organization, inspect its latest 30 Zendesk tickets, summarize and classify their sentiment, then issue project credits based on that sentiment. The interface exposes intermediate reasoning, and Gupta says a user can edit a step, remove it, or replace it with another sequence. Those corrections become guidance for later work.
Interactions can update a versioned semantic layer
Gupta demonstrates a database whose tables are named "morg", "plug", and "zorp". PromptQL samples the rows, determines which table contains employees, which contains departments, and which is a junction table. The user then corrects another mistake: department budgets are stored in cents, so the system must divide by 100. Gupta says these hints can be used to suggest metadata improvements automatically from recent threads. The semantic layer records table context and units, and each update creates a versioned build. Users can apply a suggestion and fall back to an earlier build. A later question about suspicious anti-money-laundering amounts also teaches the system that the company's first quarter starts in February.
The claimed outcome is an analyst that becomes company-specific
Gupta contrasts day zero, when the AI does not know what an enterprise customer means, how to match customer IDs, or when the financial quarter starts, with day 30. In his example, the system has learned 47 business terms, mapped relationships across six systems, and discovered 12 calculation variants. He says the system reached 100% accuracy on complex tasks in that example, and attributes similar results to customer evaluations, including a Fortune 500 food chain and a high-growth fintech company. These are claims about PromptQL's demonstrations and customer reports. The talk's practical proposal is to deploy an initially inexperienced system, correct it during use, and let those corrections improve its semantic layer.
"Every single instance of your semantic layer is version controlled so a new build is created and you can always fall back to a previous build."14:33
Who should watch
You are trying to put an AI system over inconsistent tables, SaaS applications, or APIs and keep finding that business definitions are missing.
Your text-to-SQL, RAG, or tool-calling system produces plausible plans that are hard for users to inspect and correct.
You want an approach where corrections about units, table relationships, or company terminology become reusable metadata rather than one-off prompt edits.
