In a Data Science Connect expert panel, leaders from Collate, Graphwise, Precisely and AtScale compared their views of the context layer for AI. They disagreed on whether metadata is enough, agreed that RAG and memory work together, and predicted a fight over who controls enterprise semantics.
Key takeaways
- Context is not one product. Palmer called it a constellation of technologies, while Srinivas described metadata, semantic and memory layers working together.
- Metadata maps, it doesn't compute. Palmer said metrics like revenue need a layer that computes the answer, not just describes where data lives.
- Check freshness before answering. Simson said agents need guardrails that look at quality and update metadata so stale data doesn't reach executives.
- RAG and memory are complementary. Retrieval supplies task-relevant facts; long-term memory holds preferences and policies learned over time.
- Let experts fix the model of the business. Blumauer said subject matter experts need an interface to correct ontology definitions directly.
Every vendor now says AI needs context. In the Data Science Connect expert panel webinar "The Context Layer: The Emerging Stack for Context-Aware AI," four of them explained what they mean by it, and where they think the others' approaches fall short.
Supriya Bhatia, an advisory board member at Data Science Connect, moderated. The panelists were Suresh Srinivas, Co-Founder and CEO of Collate; Andreas Blumauer, Senior VP Growth at Graphwise; Jeremy Simson, Chief Architect of SaaS at Precisely; and Mark Palmer, Chief Marketing Strategy Officer at AtScale. Brian Mink, co-founder and president of Data Science Connect, opened the session.
What is the context layer in AI architecture?
Srinivas said foundation models are maturing and intelligence is becoming a commodity, but intelligence without understanding a company's data won't create business value. Collate, the company behind the open-source Open Metadata project, describes three layers: data context from a unified metadata graph covering ownership, quality and compliance; a semantic layer built on ontology standards such as RDF and OWL; and a memory layer built from collaboration between users and AI.
Palmer rejected the term as a category. "Context management space is not a space," he said. "It's a very large constellation of different technologies." Metadata, in his view, "is a map of the information that exists in the enterprise. It doesn't actually compute the answer." AtScale's semantic layer computes metrics such as revenue so the number AI returns is correct.
Simson said Precisely prefers to leave data where it lives, because that is where permissions, logging and auditing already exist. He described a customer whose CFO was about to take chatbot answers into a board meeting without knowing a pipeline had failed the night before and the data was wrong. Agents need guardrails that check freshness and quality metadata before answering, he said.
Blumauer described RDF-based knowledge graphs as a "semantic backbone," a more expressive semantic layer that can link structured and unstructured data and capture procedural knowledge.
How do you reduce AI hallucinations with context?
Simson said Precisely combines retrieval for search, memory at the user, role and company levels, and a rule that the AI always returns to the authoritative source rather than relying on cached or remembered data. Srinivas said Collate's own published research found a 7x improvement in AI-generated SQL when a semantic layer was added.
Palmer argued that metadata-based approaches reach 60% to 80% accuracy, which isn't enough for a revenue figure on a CFO's dashboard. He said the more interesting hallucination work is in managing system prompts, tracing which prompts worked, intercepting multi-step agent calculations and using LLMs as judges.
Blumauer described translating natural-language questions into SPARQL queries against a knowledge graph to retrieve data deterministically, for customers who need 100% accuracy. Subject matter experts also need a way to fix errors. Few economists at a large bank write Python, he said, so they need an interface where they can edit the ontology's definitions and relationships directly.
When should you use RAG and when should you use memory?
Srinivas said it isn't a choice. Retrieval supplies the business information relevant to a task, while long-term memory holds preferences, policies and brand guidelines learned over time. Without memory, he said, working with AI is "like you have a brand new employee that has no context of what has happened in the past."
Blumauer recommended connecting the memory system to the same knowledge graph as retrieval, so both refer to the same entities. Simson said it depends on the application. Memory is essential in an internal Precisely system where about 26 agents run a development cycle from a Jira ticket to a feature-flagged deployment, but less useful elsewhere. He urged teams to test constantly and to make LLMs cite their sources.
What role does orchestration play in the context layer?
Srinivas said orchestration decides which context to use, when, in what order and what to do as a fallback, and that it requires extensive testing and evals. Blumauer described governance as code inside the workflow, using the SHACL standard to validate data before it enters an AI workflow.
Simson warned that some large providers expect all data to sit inside their ecosystems, which isn't realistic for most enterprises. He advised starting small and deciding where data should live.
Palmer said the more important orchestration question is how to keep agents on track as they generate their own workflows. Business rules, such as how aggressively to treat a customer at risk, often aren't written down anywhere.
Where is context-aware AI headed?
Blumauer and Srinivas both expected formal semantics to become central, with Srinivas adding that guardrails, governance and auditability will matter more as agents gain access to enterprise systems. Palmer predicted fights among large platform providers over who controls semantics, backed the Open Semantic Interchange (OSI) standard and said customers will increasingly demand that context be measured. Simson expected pricing pressure on tokens, more use of open-source models running locally and a greater focus on governance. "AI is just another user on your network who has access to your data and is going to make decisions for you," he said.
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 context layer in AI?
The panel described it as the data context, semantics, memory, governance and access controls that help an AI system understand a company's data and business, though Palmer said the term covers many different technologies.
Is metadata enough to stop AI hallucinations?
Panelists agreed metadata alone is insufficient. Semantic layers that compute metrics, knowledge graphs with ontologies, quality signals and authoritative sources are also needed.
Should enterprises use RAG or memory for AI context?
Both. Srinivas said retrieval supplies relevant business information, while long-term memory keeps preferences and policies so users don't have to repeat them.
Where is context-aware AI heading in the next 18 months?
Panelists expected more formal semantics, stronger governance and auditability, competition among platform providers over semantics, and customers demanding that context be measured.
Watch: The Context Layer: Building Context-Aware AI with Collate, Graphwise, Precisely & AtScale on YouTube
Event: Webinars · More coverage: Webinars insights
