AI data analysis and MCP: why a wrong answer looks like a right one

AI data analysis and MCP: why a wrong answer looks like a right one

The average of 10% and 30% is not always 20%

Two stores convert 10% and 30% of their visitors. Average the two and you get 20%. The chain’s real conversion is 25%. More on that below.

The current story is that AI replaces the reporting stack. Upload a sales export to Claude or ChatGPT, ask a question, and get an answer with charts and interactive dashboards. No BI project and no BI tool needed, and no analyst needed anymore to interpret the data.

Part of that is true. Retail and restaurant chains are already doing it, and for a one-off question the analysis can be excellent. One restaurant chain replaced their Power BI reporting with reports built in Claude in two days.

The part that isn't true is harder to see, because a wrong answer from an AI looks exactly like a right one. This post is about where and why that happens. It also covers the building blocks needed for a cost-efficient, reliable reporting and analytics environment that uses the analysis capabilities of modern AI tools like Claude and ChatGPT.

The first report is easy. The 100th is not

Sending data files back and forth with point-to-point integrations from business databases to Claude or ChatGPT is inefficient, even if the process is fully automated. Data warehouses were invented to make data clean, trustworthy, reliable and easy to query. Semantic layers were developed on top so that organisations have one version of the truth for metric definitions and similar concepts. User access and data security governance tools and processes were added to ensure responsibility and auditability, and to make sure sensitive data is not leaked outside. Data warehouses have always helped businesses tackle data management issues that are very hard to solve "on the fly", like:

  • Yesterday's labour data file had a new site ID that does not match the site IDs in the POS system

  • Sales include VAT in one source and exclude it in another

  • A late return changes last week's result after the report went out

  • Like-for-like growth measurement needs the same set of stores and comparable periods

  • Stock turn calculation needs past inventory snapshots stored somewhere

  • Two customer cohorts need to be compared on their campaign conversion KPIs

Connecting the AI directly to the data sources automates the fetching, but it does not easily solve the issues above. An AI will write a convincing explanation from an incomplete data export, and a well-written answer makes the mistake harder to spot.

Take an ambitious cross-sell analysis of your past three years of detailed transaction data. If Claude sends it as an SQL query to your operational POS database, it might jam your cash register operations for the rest of the day, because an operational POS system is not meant to handle such heavy queries. This is one of the reasons data warehouses were invented in the first place: to shield operational business systems from heavy external queries.

Two stores, two conversion rate tales

One store records 10 transactions from 100 visitors. Another records 90 from 300. Their conversion rates are 10% and 30%.

The chain's conversion is 25%: 100 transactions divided by 400 visitors. Average the two percentages and you get 20%, which is wrong and looks completely reasonable. The same trap sits in, for example, margin percentages and average basket values. Unique customers can't be added up across stores. Month-end stock values should not be summed into an annual figure.

The fix is a shared set of definitions, usually called a semantic layer. It says what net sales means, whether returns are in it, and how conversion is calculated at every aggregation level. The calculation then runs in the same predictable way every time, before the AI even sees a number. The AI can explain the result. It should not invent the business definitions or the arithmetic required to come up with the correct numbers.

What an MCP connector does

Connectors are the current answer to the data access challenge. MCP is the standard that lets an assistant like Claude use tools and data, and most analytics vendors are building one if they have not already released one. Zoined MCP is being piloted, with a first official release planned for Q4 2026.

A connector is a short list of abilities. Zoined MCP v1 lets an AI assistant:

  • find the metrics and dimensions available

  • choose reporting periods, comparisons and aggregation levels

  • request results with filtering, sorting, trends and changes

None of that makes a number correct. What decides the answer is what sits behind the connector: connected sources, history and shared definitions. A point-to-point connector without the supporting data warehouse and semantic layer might give you the wrong "20%" answer. Zoined MCP gives you the correct "25%" answer, and fast, because the data warehouse infrastructure is optimised for fast ad hoc querying that combines multiple sources, also in near real time.

This is the part of the AI story that gets skipped. The demo shows the question and the answer. It doesn't show who decided what net sales means, or how to make sure the same question consistently returns an answer with the same logic and definitions.

What are token costs?

Tokens are the units an AI model reads and writes, and you pay for both. Sending thousands of raw records to the model costs far more than sending the totals it needs. What follows is a model with stated assumptions, to make the topic more concrete.

Take a retail chain with 20 stores and 600 analysis sessions a month: one per store on 22 working days, plus 160 for central teams. Prices are Claude Sonnet 5.5 API list rates in November 2026: $2 per million input tokens and $10 per million output tokens.

Per month

Raw records sent to the model

Prepared results sent to the model

Input tokens per session

400,000

20,000

Output tokens per session

8,000

8,000

Input tokens

240 million

12 million

Output tokens

4.8 million

4.8 million

Model cost (USD)

$528

$72

The 400,000 figure is 10,000 detailed records at about 40 tokens each. The 20,000 figure assumes summarised data where feasible, exact drill-down results and pre-calculated metrics.

Two caveats. The left column is the heavy case, where raw records go straight into the conversation (the AI's context window). A careful setup that has the AI run code over files would use fewer tokens. The second caveat is that $528 a month will not decide anything for a 20-store chain. The reason to care is the previous section. A wrong number costs more than a token bill.

Where we stand

We sell the thing this post argues for, so discount accordingly. Zoined is a data warehouse and semantic layer for retail and restaurant chains. We also help enrich our customers' data with industry-specific solutions like AI-based sales forecasting, segment and cohort management, and benchmarking tools. Our MCP connector is currently in the pilot phase. If you want to try it, get in touch.

Product

Verticals

Pricing

Resources

Partners

COMPANY INFO & INVOICING

Zoined Oy (2440998-6)


Tallberginkatu 2A

00180, Helsinki

Finland

Invoicing address:

17623F@scan.netvisor.fi
Maventa 003724409986
OVD: FI4912373000127370


VAT number:
FI24409986

© 2026 Zoined Oy. All rights reserved.

COMPANY INFO & INVOICING

Zoined Oy (2440998-6)


Tallberginkatu 2A

00180, Helsinki

Finland

Invoicing address:

17623F@scan.netvisor.fi
Maventa 003724409986
OVD: FI4912373000127370


VAT number:
FI24409986

© 2026 Zoined Oy.

All rights reserved.

COMPANY INFO & INVOICING

Zoined Oy (2440998-6)


Tallberginkatu 2A

00180, Helsinki

Finland

Invoicing address:

17623F@scan.netvisor.fi
Maventa 003724409986
OVD: FI4912373000127370


VAT number:
FI24409986

© 2026 Zoined Oy. All rights reserved.