Seven vendors had 10 minutes each at Data Science Connect's GenAI Demo Day Q3. Their demos covered conversational agents, logical data layers, verified metrics, confidence signaling, fresh data pipelines, the Agent2Agent protocol and location data delivered through MCP servers.
Key takeaways
- Agents are quick to stand up. Snowflake's Kaitlyn Wells set up a marketing agent with two tools in under 10 minutes.
- Context drives accuracy. Denodo and Zenlytic both lean on business metadata and verified metric definitions to keep answers trustworthy.
- Signal uncertainty. PromptQL's Rob Dominguez said showing how accurate an answer is from day one matters more than starting at 90% accuracy.
- Stale data gives stale answers. Fivetran's Brad Moore said AI data must be clean, compliant and fresh.
- A2A and MCP are complementary. MCP connects an agent to tools; A2A connects agents to each other, Google's Holt Skinner said.
Seven vendors took 10 minutes each at Data Science Connect's GenAI Demo Day Q3, a virtual event hosted by DSC co-founder and president Brian Mink and headlined by Snowflake and Denodo, with sponsors Zenlytic, PromptQL, Fivetran, Google Cloud and Precisely. A shared theme ran through the demos: AI answers are only as reliable as the business context and data behind them. Here is what each presenter showed.
Snowflake: a conversational agent over structured and unstructured data
Kaitlyn Wells, AI/ML Architect on Snowflake's Applied Field Engineering team, demonstrated Snowflake Intelligence, the company's conversational assistant. Acting as a business analyst at a marketing agency, she asked a "marketing genie" agent about 2024 ad performance by platform. The agent used Cortex Analyst to generate SQL, then charted cost per acquisition by month on request.
She then queried a 65-page HubSpot research report on how marketers use AI. Cortex Search returned recommendations with citations to the source. Behind the scenes, Wells showed the agent itself: two connected tools, model selection left on auto, and traces for debugging latency or tool choice. Setting it up took under 10 minutes, she said, and everything shown is also available through Snowflake's agents API for use in tools such as Slack or Teams.
Denodo: a logical data fabric and deep query
Chris Walters, a principal sales engineer at Denodo, said LLMs trained on public internet data lack an organization's own context. Denodo's answer is a logical data layer that connects cloud, on-premises and API sources without copying data into one place. Denodo stores only metadata, he said, including the business descriptions that improve AI accuracy, which its assistant can draft with an LLM.
In a banking demo, a chatbot built on the Denodo AI SDK turned a question about outstanding account balances into a query against the right view. He also showed a new deep query capability for "how" and "why" questions. Asked how a bank could improve profitability by 50%, it ran a multi-step, multi-source analysis and produced a report with charts and recommendations. Such runs take several minutes, so he showed one prepared in advance.
Zenlytic: verified metrics and what-if analysis
Paul Blankley, CTO and co-founder of Zenlytic, showed Zoe, the company's AI data analyst. Asked by voice for year-to-date revenue, orders and profit margin by product, Zoe reused existing definitions from the semantic model, marking them with a green check as verified, and flagged profit margin percentage as a new calculation she had derived. A "promote to verified" option adds such a definition to the model, so it can grow as people ask questions.
Zoe also ran a scenario: moving 80% of ad spend from the worst-performing channel to the best one and estimating the effect on customer acquisition cost. The goal, Blankley said, is "the trust of a semantic layer, but the depth of SQL."
PromptQL: confidence signals and continuous learning
Rob Dominguez, engineering manager at PromptQL, said the biggest barrier to adoption is that "even when AI doesn't know, it's still going to answer." Confidently wrong answers erode trust, he said, and much business context is tribal knowledge never captured up front.
PromptQL builds a company-specific domain language on top of foundation models from any provider, rather than training its own. It shows a plain-language query plan, with the underlying Python and SQL for technical users. Two features stood out in his talk: native confidence signaling, and continuous learning from user corrections, such as telling it to look at a different table. "The key to long-term adoption isn't starting with 90% accuracy. It's signaling how accurate something is from day one," Dominguez said.
Fivetran: fresh, clean, compliant data
"Fivetran is not an AI company. We are a data company," said Brad Moore, Senior Sales Engineer, Enterprise. Most companies start AI at the tool level, he said, but business value comes from starting with the business problem and working back to the data. That data must be clean, compliant and fresh. "If you're working an AI model off of data that's a month old, you're going to get month old results," he said.
Moore highlighted Fivetran's ability to land data in Iceberg or Delta format in cloud object storage, where structured and unstructured data can be combined, and described a Snowflake Summit lab that turned a raw Postgres database of Southern California wineries into a natural-language app.
Google Cloud: the Agent2Agent protocol
Holt Skinner, Developer Advocate for Google Cloud AI, explained the Agent2Agent (A2A) protocol, an open standard for agents built on different frameworks and hosts to work together. Agents advertise capabilities through an "agent card," negotiate formats, share task status and collaborate. Asked whether A2A competes with MCP, he said "A2A actually loves MCP": MCP connects an agent to its tools, while A2A connects agents to one another.
A2A works with any framework or model, he said, and the project is owned by the Linux Foundation. He pointed to built-in support in Google's Agent Development Kit, IBM's BeeAI, LangGraph and Pydantic, and warned that copycat documentation sites appear in searches. The official site is a2a-protocol.org.
Precisely: location data through MCP servers
Nikhil Parekh, a senior solutions engineer at Precisely, demonstrated the company's MCP servers inside Claude. Without the server, Claude asked for more details to complete an address. With it, Claude returned a full address with ZIP+4 and coordinates. Follow-up questions chained property attributes, neighborhood and school data, wildfire and crime risk, and emergency dispatch (PSAP) information. Point-level location accuracy matters, Parekh said, offering the example of a fire truck sent to the wrong end of a street. He said Precisely offers more than 44 APIs, from geocoding to phone and email validation.
Catch the next AI Demo Day
AI Demo Days are one of four series on Data Science Connect's 2027 schedule of 24 free live webinars, alongside the AI × Enterprise Panel Series, Agentic AI Showcases and the Plug & Play Series. See the full lineup on the webinars page.
Frequently asked questions
What is the difference between A2A and MCP?
According to Google's Holt Skinner, MCP is how an agent connects to its tools, APIs and resources, while the Agent2Agent protocol lets independent agents communicate, delegate tasks and manage shared workflows. Sophisticated systems often use both.
What is confidence signaling in AI?
PromptQL's term for an AI system telling users how sure it is of each answer, so they know when to trust it and when to verify or correct it.
How do AI data analysts keep metrics consistent?
In Zenlytic's demo, its analyst Zoe reused verified definitions from the semantic model, flagged new calculations as unverified and let admins promote them to verified.
What are Precisely MCP servers?
MCP servers that let AI assistants such as Claude call Precisely's location and data APIs, including address autocomplete, property attributes and wildfire and crime risk.
Watch: GenAI Demo Day Q3 on YouTube
Event: Webinars · More coverage: Webinars insights
