At Data Science Connect's AI Demo Day Q2, six vendors showed working enterprise AI in 10 minutes each. Mistral AI and SAP demoed agents, while Rhino, Precisely, Skyflow and SecuPi showed how MCP servers can connect AI assistants to enterprise data without exposing what should stay protected.
Key takeaways
- Agents are moving to the cloud. Mistral's teleport feature moved a coding agent's work from a laptop to a cloud workspace.
- Skills follow paths, agents find them. SAP's Trivedi separated rule-based Joule skills from goal-driven agents.
- AI can analyze data it can't see. Rhino's federated platform returned only weights and aggregate statistics to Claude.
- Missing context invites hallucination. Precisely's MCP server adds location data that first-party records lack.
- Protect data in flight, not just at rest. Skyflow tokenizes PII in prompts, and SecuPi enforces cell-level policies on AI queries.
Six vendors had about 10 minutes each to show working enterprise AI at Data Science Connect's AI Demo Day Q2, a virtual event with Mistral AI as headlining sponsor and SAP, Rhino Federated Computing, Precisely, Skyflow and SecuPi as sponsors. A common thread ran through most of the demos: connecting AI assistants such as Claude to enterprise systems through MCP servers while keeping sensitive data under control.
Presenters were Aaron Delp, Head of Technical Marketing at Mistral AI; Shubham Trivedi, Digital Solution Advisor at SAP; David Saluki, a forward deployed engineer for AI and ML at Rhino Federated Computing; Ben Tuck, Product Marketing Manager at Precisely; Sam Sternberg, Global Head of Solutions Architecture at Skyflow; and Noam Dror, Chief Information Security Officer at SecuPi. Here's what each demonstrated.
Mistral AI: What can agentic work mode, OCR and voice models do?
Delp set out to run "four demos in about nine minutes," all built on releases from the previous 30 days. In Mistral's chat interface, a new work mode for agents pulled instructions from Mistral's documentation site. In Mistral Vibe, which he called Mistral's coding agent, a feature called teleport moved an agent's work from his laptop into a cloud workspace, where it fixed a disordered priority column in a support desk app and opened a GitHub pull request.
He also ran OCR on technical documents, loaded the results into a database behind an MCP connector, and searched a parts library in plain language. An agentic workflow reviewed a drug regulatory submission's modules in one pass and reported gaps and recommendations. Finally, he recorded his voice, transcribed it and played it back in a different synthetic voice, named Marie, set to sound excited. He said batch API calls can handle up to a million files at a time.
SAP: What is the difference between an AI skill and an AI agent?
Trivedi showed Joule Studio, part of SAP Build on SAP's Business Technology Platform. A Joule skill fits rule-based, repetitive work where the path is known, such as creating a ticket, he said. An agent plans, reasons, acts on its own and can work with other agents when the path is unknown. "Skills do what you tell them to do. Agents figure out what to do," he said.
His demo followed a billing dispute in SAP's enterprise service management. Agents classified a customer email, proposed three solutions and, after a human reviewed the case and chose options, consulted cash collection and financial accounting agents, drafted a customer email and settled the dispute. Each step was visible in a reasoning tab for auditing. Trivedi credited SAP's knowledge graph with grounding the agents in business context.
Rhino Federated Computing: How can AI analyze data it can't see?
Saluki said Rhino separates the orchestration plane from the compute plane, so analytics, model training and agents run across distributed data "without the data ever leaving its source." Through the Rhino MCP server in Claude Desktop, he profiled a lung cancer data set held at a site he didn't own, compared two sites' cohorts, plotted results and fit a Cox proportional hazards model. Only model weights and aggregate statistics travel back, he said.
Comparisons like this usually take weeks or months of approvals, Saluki said. When he asked for the raw data, Claude replied that no tool allowed it, "and that's actually by design."
Precisely: How does location data reduce AI hallucinations?
Tuck said first-party data alone leaves gaps that AI fills by making things up. On location questions, he said, "best case you are going to get no answer and worst case you're going to get one that's just completely fabricated." Precisely's MCP server connects AI tools to its data enrichment sets and location services, including geocoding, address verification, risk factors such as wildfire, demographics, parcels and buildings.
In Claude, he geocoded an incomplete New York address, then generated a site selection report that combined Precisely data with a sample consumer spend file. He described the PreciselyID as a persistent identifier that links an address to related data sets, and said the MCP server is free to use with a Precisely API key.
Skyflow: How can you protect PII without breaking AI?
Sternberg said "a small fraction of data drives almost all of the regulatory and breach risk." Skyflow detects sensitive data across 69 entity types and replaces it with context-preserving tokens, so a model can follow a person through a document without seeing their name.
He showed three uses: protecting a document before it reaches a model, tokenizing a customer's name and Social Security details in a support chat before the prompt went to OpenAI and restoring them in the reply, and indexing enterprise files into a knowledge base so that a question about an employee's salary is answered on tokenized data. "You don't need to choose between protecting your data and keeping it usable," he said.
SecuPi: How do you control what AI can access in a database?
Dror said AI lacks deterministic access controls, and not every value can be tokenized; a two-letter state code has few possible substitutes. SecuPi applies policies down to the column, row and cell level, logs the original user rather than a service account, and assigns data risk scores. "You don't need an army of developers," he said.
In his Claude demo, a query through SecuPi's MCP server returned encrypted email addresses. Changing one policy, so that users with GDPR training see data in the clear, made them readable. Both queries appeared in an audit log naming the real user.
Watch the next AI Demo Day
AI Demo Days are one of four series in Data Science Connect's 2027 schedule of 24 free live webinars. See the webinar schedule for upcoming demo sessions.
Frequently asked questions
What is the difference between an AI skill and an AI agent?
In SAP's framing, a skill runs a known, predefined set of steps, while an agent is given a goal and plans and reasons its way to the next steps.
What is federated AI?
An approach in which analysis and model training run where the data lives, with only model weights and aggregate statistics shared back, as Rhino demonstrated.
How can you protect PII sent to an LLM?
Skyflow's approach replaces sensitive values with context-preserving tokens before a prompt reaches the model and restores them in the response for authorized users.
How do MCP servers connect AI to enterprise data?
In these demos, MCP servers let Claude query location data, federated data sets and databases through governed tools rather than direct access.
Watch: AI Demo Day Q2: Enterprise AI Demos from Mistral AI, SAP, Rhino, Skyflow, Precisely & SecuPi on YouTube
Event: Webinars · More coverage: Webinars insights
