AI assistants give confidently wrong business numbers when they guess at your data. In our Data to AI webinar we showed how a modeling layer, a query layer instead of raw data, and a receipt with every number change that, with benchmark results on NetSuite data and a live demo on marketing data.
On 5 November, Petr Nemeth (Dataddo's founder) and Radim Marek ran a live session on Data to AI, our way of connecting business data to Claude, ChatGPT and other AI assistants. About 60 minutes: how it works, a benchmark, a live demo and audience Q&A. Here is what was covered.
Watch the full recording, or jump to the four short clips linked below.
The problem: the confidently wrong number
Ask an AI assistant a question about your business data and it will answer. The risk is not an error message. It is a clean, well-argued number that is wrong: the wrong currency, a position summed as if it were a count, a business rule the model never saw.
Most assistants connected to a database write their own SQL, guess what the columns mean, and hand you the result. You have to trust the query.
Radim explains it in 40 seconds: agents that write their own SQL can misread columns and metadata, pick the wrong currency or make a business-logic error, and still sound sure of themselves.
Clip (0:40): What is a confidently wrong answer?
How Data to AI is different
Data to AI sits on the same connectors as the rest of the Dataddo platform, and adds a modeling layer and an MCP server on top. Three design choices came up repeatedly:
- The assistant never receives the raw dataset. It gets a query layer. Data leaves the platform only as the answer to a question, and fields marked as sensitive never reach the model.
- The assistant does not write SQL. It picks from a defined set of operations built on metadata, so it cannot invent an aggregation that makes no sense (you cannot sum a ranking position, for example).
- Every number comes with a receipt. Lineage, freshness, completeness and constraints travel with the data, so the assistant knows where a figure came from and how far to trust it. When it cannot answer safely, it says so instead of guessing.
The receipt idea is the one that stuck with people. In this clip Radim explains why a number should arrive with its lineage, freshness and constraints, so the assistant can show where it came from instead of guessing:
Clip (0:47): Every number comes with a receipt
Any model that supports the MCP protocol works, including models hosted in Europe, which matters for teams with data-sovereignty requirements.
The benchmark
We ran 21 questions, three times each, against a NetSuite company group built from real-world data (200,000 transactions, three currencies), using three Claude models and three ways of connecting the data: a raw SQL connector, CSV exports, and Data to AI. The answer key was fixed before the run. Full methodology is in the benchmark whitepaper.
What Petr highlighted:
- A smaller model on Data to AI matched a frontier model on raw SQL. Haiku 5.5 on Data to AI had 87.3% good outcomes (correct answer or correct decline). Opus 5.5 on raw SQL had 81.0%.
- That is about 24x cheaper per question ($0.0045 versus $0.108). The 24x compares two different models on purpose: it is what a team pays for the model and connection combination it actually chooses. The accuracy gap between the two is small on 63 answers each, so read it as "matches", not "beats".
- Fewer confidently wrong answers. With Data to AI, Sonnet 5 gave 58% fewer confidently wrong answers vs with raw SQL ,
- Sensitive data stayed put. Asked for a sales rep's email address, Data to AI disclosed it 0 of 9 times. Raw SQL disclosed it 9 of 9.
The practical takeaway from Petr: with well-organized data and the right context, you do not need a frontier model to get a solid answer.
The live demo
"The whole magic that Data to AI can bring to you is actually where you would employ more than one source." - Radim Marek
Radim ran the demo live, unscripted, on a synthetic company ("Fernhill Outdoors") with Facebook Ads, Google Analytics and Google Search Console connected.
- Cross-checking sources. Claude was asked to verify what the ad platform claimed against GA4 e-commerce data, then to check GA4 for missing traffic using Search Console. Its answer: GA4 lost about 60% of its traffic on two days (Oct 6 and Nov 3), roughly 4% of the 30 days, along with what may be behind the drop. The model layer also knows that Search Console data lags by about three days, so it refuses to compare the most recent days instead of reporting a false gap.
Clip (0:33): Is GA4 missing any traffic? Ask your data. (Sample data; Claude took about two minutes to answer and the wait is cut.)
- A budget decision. Asked where the next 5,000 euros should go across Google and Meta, Claude answered with returns calculated from the actual traffic and spend data.
- A shareable result. Claude then generated an HTML report, ready to send to colleagues or paste into a team chat. Radim's framing: not a replacement for a warehouse or dashboard, but an answer a minute after the question instead of a ticket and a wait.
Radim also noted that in demos on customers' own marketing data, the same approach has surfaced Google Ads spend landing on broken pages and the largest campaign budgets going to the weakest campaigns.
How you set it up
For existing Dataddo customers on an Anywhere or Enterprise plan, Data to AI is already available as another destination for a flow. There is also a dedicated, simpler interface:
- Choose your AI client (Claude, ChatGPT or a generic MCP connection).
- Create a context: pick the connectors and data you want the assistant to see.
- Add the Dataddo data access connector in your AI client and approve the tools. Keep the query tool enabled, since it is how data reaches the model.
Governance sits in the same place. Each person connects through their own account with an MCP client role, you control who sees which context, and audit trails, usage, freshness and data quality checks are all visible per context. A free trial lets you test it on your own data.
Questions from the audience
- How do you involve the business in a technical exercise? The aim is to cut the chain of ETL, warehouse, metric layer and dashboard for the questions that do not need one, and make connecting business data straightforward. Dashboards still have their place.
- Does it help overloaded BI teams? Yes, that is a target use case. People already "vibe-code" reports; a governed interface lets them do it on trusted data while the analytics team focuses on the work that needs it. Radim on why the enterprises we already work with see room for this:
Clip (0:46): BI team buried in vibe-coded reports?
- Best tool for report structure with Shopify? Add a Shopify context. Relations and definitions ship with the data, so you ask questions directly instead of designing a report first.
- Can you stop the model from guessing? That is what the semantic model is for: rules such as "positions cannot be summed", and no free-form SQL, so every answer arrives with its lineage.
Try it
Start a free trial and connect your own data. Or watch the full webinar recording.