In a Data Science Connect expert panel, leaders from SAP, Turgon, Skyflow and AtScale debated what a single source of truth means when AI agents consume and create data. They covered semantic layers, permissions that travel with data, agent-generated data and evals against a company's own KPIs.
Key takeaways
- Truth lives in meaning. Palmer said agents need a semantic layer that defines how metrics are computed, not just access to a database.
- Context goes stale. Varma said catalogs and ontologies only help if the meaning they hold is current at the moment an agent acts.
- Permissions travel with data. Trivedi said semantics, relationships and access rules must be embedded in the data and enforced at every point of access.
- Agents create data too. Moktali warned that one agent's inference can become another agent's authorized input with no check in between.
- Measure on your own data. Panelists recommended evals against a company's own questions and KPIs to show whether agents get the right answer.
For decades, the single source of truth meant a database that people could query and trust. In the Data Science Connect expert panel webinar "The Single Source of Truth: Building Data Foundations for Agentic AI," four vendors argued that AI agents change that definition, and they disagreed on how far the change goes.
Preeti Vaidya, Executive Director, Data & AI at JPMorgan Chase, moderated. The panelists were Shubham Trivedi, Digital Solution Advisor at SAP; Ankur Varma, CEO and Co-Founder of Turgon; Amruta Moktali, Chief Product Officer at Skyflow; and Mark Palmer, Chief Strategy and Marketing Officer at AtScale. The webinar was sponsored by SAP, Turgon, Skyflow and AtScale.
How is the single source of truth changing for AI agents?
"I don't think that the source of truth is in the data," Palmer said. In the BI era, he said, human analysts corrected and reconciled data, and those fixes lived in their heads. LLMs have no such knowledge. He cited the open-source BIRD-Interact benchmark, where he said the top score for an LLM given raw data is 38% accuracy. His conclusion: what agents need is a single source of meaning, a semantic layer that defines how metrics are computed.
Moktali said agents scatter data into context windows, long-term memory, vector stores, tool calls and traces. The source of truth, she said, has moved from where data sits to where it is resolved and acted on, which is why Skyflow focuses on protecting data at runtime.
Varma said many companies have built catalogs and ontologies, but the key question is when they were last updated. Answers are only as good as the meaning at the moment of the prompt or action. Trivedi added that agents also need to know the business process: what to do with a purchase order, who approves it and whether a human must be involved. He said that doesn't mean the database is going away.
What makes data ready for agents to use?
Trivedi asked, "Would you trust an employee with no tribal knowledge to use data correctly from its description alone?" If not, an agent can't either. He said semantics such as definitions, units, hierarchies and relationships must travel with the data, identifiers must be consistent across systems, and permissions must be enforced at every point of access. An agent might need a product's selling price, for example, but not its cost.
Varma gave the example of revenue with and without tax. Both columns are valid, but a CFO and a head of sales use them differently, so policies on when to use each have to sit alongside the data, along with information on how fresh it is.
Trivedi also argued that batch ETL copies create stale, divergent data that can cause hallucinations. He called for event-based updates and avoiding copies.
Are AI agents becoming creators of enterprise data?
Moktali said data is increasingly multimodal and that agents now generate information other agents consume. "The inference generated by agent A is becoming authorized data that agent B uses," she said, often with nothing in between. She described companies whose support agents close tickets autonomously and then show the user a trace of what they did.
Palmer pushed back. He said no large company he talks to lets an LLM parse a document and update its system of record without human curation and validation. Varma said it is early but already happening, citing invoice data extracted and handed between agents for process automation. Trivedi suggested tiering human oversight by risk, starting with low-value, reversible tasks.
Varma also cautioned against holding AI to a higher bar than people. Teams should ask "accuracy in relation to what," he said, since human processes for tasks like fixing master data make mistakes and don't scale.
How can enterprises verify that agents get the right answer?
Palmer said the BIRD-Interact benchmark includes 600 questions across 22 databases, with flawed documentation and "golden SQL" answers. AtScale ships that framework with its semantic layer so customers can measure accuracy on their own data and questions, and he said some customers reach 100% accuracy on the questions they specify. Moktali said she sees more teams running evals tied to an agent's specific purpose.
Varma said Turgon's harness builds the context layer and tests it by having agents regenerate the business's existing KPIs, using the gaps as feedback. Logs, audit trails and ETL scripts, he said, are clues agents can use to learn a company's processes. Trivedi said SAP agents inherit permissions and business processes already defined in SAP systems. Palmer replied that not all of the world's data is in SAP.
Asked why agents guess at known facts, Palmer said there are often many versions of the truth. One AtScale customer, Blue Yonder, has 20 definitions of revenue, so getting the right answer is a matter of disambiguation.
What is the one rule for building a trusted data foundation?
In a closing round, Trivedi said governance, permissions and audit trails must be built in, not added later. Varma called for "living infrastructure" that keeps learning. Moktali said not to block data from agents but to make it trustworthy through tokenization and runtime controls. Palmer summed up his approach in four words: "Model, measure, govern, repeat."
Join the next conversation
Data Science Connect's 2027 schedule includes 24 free live webinars in four series: the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. Registration is coming soon; see the webinar schedule for upcoming sessions.
Frequently asked questions
What is a single source of truth for agentic AI?
Panelists said it is no longer just a database. It includes the semantics, business processes, policies and permissions agents need to use data correctly, kept current as the business changes.
Why do AI agents give wrong answers from enterprise data?
Palmer said LLMs given raw data score poorly on benchmarks because they guess among ambiguous definitions, such as which of many revenue measures to use.
How do you make enterprise data ready for AI agents?
Embed definitions, units, relationships and permissions with the data, keep identifiers consistent, update in real time instead of relying on batch copies, and protect sensitive values at runtime.
How can companies test whether AI agents are accurate?
Run evals on their own data: compare agent answers with known correct queries or have agents regenerate existing KPIs, then use the gaps as feedback.
Watch: The Single Source of Truth for Agentic AI | SAP, Turgon, Skyflow & AtScale | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
