At Data Science Connect's AI Demo Day, four vendors ran 10-minute live demos. Mistral AI showed document OCR and voice models, SAS detected AI-generated forgeries, Fivetran served one dataset to seven query engines, and Precisely fed location data to Claude through an MCP server.
Key takeaways
- Confidence scores keep humans in the loop. Mistral's OCR demo flagged low-confidence handwritten fields for review.
- Fighting forgeries takes AI. SAS combined forensic computer vision, OCR and prompt-injection checks to flag manipulated documents.
- LLMs summarize well but estimate poorly. SAS replaced LLM authenticity estimates with deterministic models.
- One copy can serve every engine. Fivetran showed seven query engines returning matching answers from a single open-format dataset.
- AI needs external data. Precisely argued that internal data alone leaves gaps that lead to hallucinations and missed ROI.
Four vendors had about 10 minutes each to show working enterprise AI at Data Science Connect's AI Demo Day, a virtual event sponsored by Mistral AI, the headlining sponsor, along with SAS, Fivetran and Precisely. The demos covered document and voice AI, detecting AI-generated forgeries, giving many AI agents one copy of the data, and feeding external location data to AI tools.
Presenters were Jenn Cunningham, Product Marketing Manager at Mistral AI; Robert Blanchard, Principal Data Scientist at SAS; Kelly Kohlheffer, Senior Global Director, Partner Sales Engineering at Fivetran; and Ben Tuck, Product Marketing Manager at Precisely. Here's what each demonstrated.
Mistral AI: How does document AI handle handwritten forms?
Cunningham described Mistral as a frontier AI lab building LLMs for the enterprise, with general-purpose models plus specialist models for OCR and audio. She worked in Mistral's Studio platform, which she called one platform to build, deploy and govern agentic AI systems, and said everything shown is also available through the API.
Her first demo ran OCR on a fictional insurance annuity application filled out mostly by hand. The output reported a 95% confidence score for the document and a minimum of 16% for individual fields, color-coded so that low-confidence fields could go to a person for review.
A colleague's workflow in Mistral's Vibe platform, which she said has work and code modes, applied the same idea to a photographed patient intake form. Anything under a 90% confidence score was routed for human review before the file was approved and indexed.
She also showed text-to-speech with a cloned voice in English and Spanish, and speech-to-text transcription of a recording made in a London pub during an England-Argentina match. She pointed to factory floors and noisy airfields as use cases, along with subtitles and translation.
SAS: How can AI detect synthetic forgeries?
Blanchard said people can no longer reliably tell real images from manipulated ones. He cited forgery increases of 244% from 2023 to 2024 and 300% from 2024 to 2025, and said customers had come to SAS overwhelmed by synthetic forgeries of documents such as bank statements, IDs and receipts.
His agentic pipeline, built in SAS Intelligent Decisioning, starts with a dual-stream computer vision model. One stream reads an image semantically. The other applies forensic transformations, such as error level analysis and noise residual analysis, that expose traits of generated or edited images. Real photos carry noise, he said, while synthetic images are very clean.
A classification model then routes documents to OCR. Before extracted text reaches an LLM, the pipeline checks it for indirect prompt injection, hidden text meant to make an agent do something it shouldn't. If an attack is found, the LLM is bypassed. One test ID in the demo was flagged for hidden admin-override text.
Blanchard said early versions asked the LLM to estimate authenticity, and the estimates were unreliable. The team replaced them with deterministic models, combined by a global adjudicator. Results feed a dashboard for a human auditor. Explainability is switched off for images judged real. "Compute costs money," he said. "So why spend the compute to know why something is real?"
He closed by arguing that "trust can no longer be passive." It has to be evaluated by models, processes and human judgment.
Fivetran: How can every AI agent use one copy of the data?
"I don't believe that we are really the primary data consumer anymore," Kohlheffer said. Agents are becoming the primary consumers of data, often several at once for a single person, and they act on confusing answers instead of asking follow-up questions. So the data underneath has to be right first. Kohlheffer noted that Fivetran and dbt Labs had merged a month or two earlier.
The demo used Claude Code in VS Code with a skill and an MCP server. Fivetran's managed data lake service, Kohlheffer explained, moves data from any of more than 750 sources into a customer-owned bucket in S3, GCS or ADLS, writing Parquet files with both Iceberg and Delta metadata on top.
The agent queried that single dataset through Snowflake, Databricks, Athena, BigQuery, DuckDB and Microsoft Fabric, then added ClickHouse live as a seventh engine, and all returned matching answers. The tally: "seven engines, five catalogs, three open formats, one copy." Kohlheffer said 11 engines have been connected this way, in under an hour combined.
Precisely: Why does AI need external location data?
Tuck said many AI projects fall short of ROI because they rely only on internal data. Without external data on properties, demographics, boundaries and risk, AI tools return no answer or hallucinate, he said.
Precisely links its datasets with unique IDs and extends them through its DataLink partner program, he said, citing Dun & Bradstreet as an example. An MCP server lets business users ask questions in natural language without GIS tools or spatial joins.
In Claude, Tuck showed a report for a fictional telecom company weighing fiber expansion in the Dallas-Fort Worth area. It ranked ZIP codes using demographics, population growth forecasts and carrier coverage, and folded in building permit data from DataLink partner Shovels, which he said covers more than 160 million permits.
See the next AI Demo Day
AI Demo Days are one of four series in Data Science Connect's 2027 lineup of 24 free live webinars, alongside the AI × Enterprise Panel Series, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule for upcoming demo sessions.
Frequently asked questions
How can AI detect fake or manipulated documents?
SAS's demo used a computer vision model with forensic transformations such as error level analysis and noise residual analysis, plus OCR checks, with results reviewed by a human auditor.
What is an indirect prompt injection attack?
As SAS's Robert Blanchard described it, hidden text in a document that tries to make an AI agent do something it shouldn't, such as downloading a malicious payload.
Can multiple query engines share one copy of data?
In Fivetran's demo, data written as Parquet with Iceberg and Delta metadata was queried by seven engines, including Snowflake, Databricks and BigQuery, with matching results.
How do AI tools use location data?
Precisely's demo connected Claude to an MCP server that returns property, demographic, boundary and risk data from natural-language questions, without GIS tools.
Watch: AI Demo Day: Enterprise AI Demos from Mistral AI, Fivetran, SAS & Precisely | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
