AtScale for AI

When the finance team talks about “revenue,” the sales team talks about “revenue,” and even the AI assistant talks about “revenue,” those three figures may actually come from three different sets of algorithms. AtScale for AI addresses precisely this kind of data verification challenge in enterprises: it links the meaning of metrics, calculation rules, and validation processes, ensuring that AI has a solid foundation to rely on when querying existing business data.

It is better suited for companies that already have a data warehouse, BI reports, and a data team. When evaluating it, first determine whether your metrics can be clearly modeled and whether the AI can reliably access these metrics within its authorized scope. This article is based on data as of October 17, 2026Compiled from official documentation, including selection analysis and recommended validation procedures; no deployment or paid testing has been conducted.

What is AtScale for AI?

AtScale for AI is a product from AtScale launched on October 15, 2026The enterprise AI capabilities suite that has been released includes question clarification, business metric calculation, AI-assisted semantic modeling, and answer evaluation. Please pay special attention to the delivery status:Ava and ACE are now available; CoModel and Canary are still coming soon....You shouldn't assume that all features are already available for use.Official Announcement

To understand this, we can start with the “semantic layer.” Enterprise databases store records such as orders, customers, products, and dates; the semantic layer organizes these records into business concepts such as “net revenue,” “active customers,” and “year-over-year growth,” while clearly defining relationships and algorithms. This way, business users can ask questions using familiar terminology, while the data team is responsible for maintaining reusable interpretation rules.AtScale Help Center: Introduction to the Semantic Layer

For example, “sales in the East China region this quarter” raises at least three questions: Should the data be based on order creation dates or revenue recognition dates? Are refunds deducted? Does “East China” refer to the customer’s location or the store’s location? Such discrepancies are difficult to resolve solely through more articulate AI responses. AtScale’s product approach is to first map the question to a well-defined business definition and then perform the calculation.

The Company and Product Foundation Behind It

AtScale is a provider of enterprise semantic layers. The company’s history traces the origins of its product back to 2013; the team has a background in Yahoo’s data team and has long addressed issues related to metric consistency across BI tools and large-scale data analysis. Consequently, AtScale for AI extends AI workflows based on the existing semantic layer, rather than being a consumer-grade Q&A service built from scratch.Company Profile

The base platform supports data platforms such as Snowflake, Databricks, and Google BigQuery, and also connects to analytics tools such as Power BI, Excel, and Tableau. AI is integrated via the AtScale MCP Server, allowing BI and AI to reuse the same set of metric definitions. Specific data platform versions, deployment methods, and client support should still be verified on a case-by-case basis.Official Integration Guide

See Design Center Official Interface Guide, to learn about the semantic modeling canvas of the underlying platform; this page does not indicate that CoModel or Canary has been officially released.

What are the responsibilities of each of the four components?

1. Ava: First, determine exactly which metric the user is asking about.

Ava stands for Ask, Verify, Answer, and focuses on aligning natural language questions with governed metrics. When faced with a term like “revenue,” which can have multiple meanings, it clarifies the intended meaning before performing calculations, thereby reducing instances where the system provides a seemingly correct answer despite using the wrong metric.Status: Available.

2. ACE: Enforce existing metric rules

ACE, which stands for AI Computation Engine, is responsible for implementing business logic and computational logic while balancing query performance and computational cost.Status: Available.

3. CoModel: Assists in building and maintaining semantic models

CoModel is designed to translate business context into metrics, dimensions, and relationships to assist with model maintenance. Proposed changes still require manual review.Status: Coming soon.

4. Canary: Verifying Answers Based on Real-World Business Problems

The Canary project plans to use a BIRD-Interact-based evaluation framework, along with the company's own questions and verified answers, to identify errors and track regressions.Status: Coming soon.The above four items and their status are subject to the official announcement.Components and Usability Guidelines

Why is ACE particularly important for complex metrics?

Simple summation is relatively easy to understand, but “inventory balances,” “gross profit margins,” and “cumulative revenue through this month” involve different calculation rules. Inventory can typically be summed across warehouses, but daily inventory levels cannot simply be added together to arrive at the end-of-month inventory; ratios also often require first aggregating the numerator and denominator before recalculating. AtScale’s technical documentation lists multi-stage calculations, semi-additive balances, and financial calendars as key areas of focus for ACE processing, and explains that calculations are pushed down to the enterprise’s existing data platform.AtScale for AI: Technical Overview

The implication for model selection is this: during a demonstration, don’t just look at “sales totals by region.” Instead, identify the metrics most likely to be disputed within the company and verify whether the aggregation granularity, time boundaries, and filter criteria remain consistent throughout.

What practical value does it offer?

The following is an application analysis based on the product's mechanics; it does not represent the actual test results from this article.

• Reduce discrepancies between financial statements and AI. If BI and AI use the same set of audited metrics, this can reduce the likelihood that each tool will reinterpret the metrics. However, this is contingent on the metric definitions themselves being correct and the retrieval process actually passing through that semantic layer.
• Make the data inquiry process easier to verify. In addition to the AI’s response, you should also save a record of which metrics were used, what filter criteria were applied, what time range was selected, and the details of the query execution. Only then can business users determine whether the two answers are addressing the same question.
• Turn repeated fixes into reusable rules. When the team discovers that the definition of “active customer” is missing, incorporating it into the governance model makes it easier to reuse across tools than simply modifying a single chat prompt.
• Ensure that cost estimates have a clear target. It is important to consider not only the pricing at the semantic layer itself, but also to measure query latency, data warehouse computational overhead, and model invocation costs. The optimization gains claimed by the product should not be directly treated as realized savings in a company’s procurement budget.

What are some suitable use cases?

Financial and Operational Analysis

Suitable for scenarios requiring strict consistency in metrics, such as revenue recognition, budget execution, and regional profit margins. We recommend first selecting a metric that has already been approved by the finance department, and then comparing the results from BI tools, databases, and AI responses using the same metrics. Figures related to official financial statements, settlements, or decision-making should still follow the existing approval process.

Retail, Inventory, and Supply Chain

It is suitable for comparing sales, inventory, and delivery data by store, region, product category, and time period. Pay special attention to inventory snapshots, cumulative totals across periods, returns processing, and hierarchical summaries—and avoid mistaking “being able to look it up” for “being calculated correctly.”

AI Data Assistant for Business Teams

This is suitable for providing employees within an existing permission framework with a natural language query interface. The official documentation includes connection instructions for Claude, ChatGPT, Databricks, and Google Gemini, among others; however, the integration methods and prerequisites vary slightly across different clients.AI Application Documentation

Cross-Tool Metrics Governance

If a single company uses Excel, Power BI, Tableau, and AI assistants simultaneously, maintaining consistent business definitions is valuable in and of itself. Conversely, if an individual is merely analyzing a few CSV files from time to time, the costs associated with modeling, deployment, and operations on an enterprise-grade semantic platform typically require particularly careful evaluation.

Getting Started: Recommended Pilot Routes

The following are implementation recommendations based on the official onboarding process; they are not records of completed tests.

Step 1: Prepare a set of business questions with a narrow scope and clear answers

Select a business domain, such as order analysis. Prepare 20–50 questions and have the metrics owner verify the answers; the questions should cover edge cases such as refunds, cross-month transactions, currencies, zero values, null values, and synonyms. This number is a recommendation for the pilot program and does not represent an official product threshold.

At the same time, clearly specify the person responsible for each metric, its definition, data source, authorized users, and update schedule. Without this information, no matter how quickly the model generates results, it will be difficult to determine whether the results meet business requirements.

Step 2: Import Data and Build a Model

The data team confirms the deployment plan and data connections, builds the model in Design Center, and checks the dimensions, metrics, relationships, and time logic. The official canvas displays the model structure, and the text editor allows you to edit SML; the model validator also displays structural errors, data validation prompts, and warnings.Design Center Documentation

Step 3: Configure the MCP and AI Client

Administrators must first enable the AtScale MCP Server, then configure the connection according to the documentation for the corresponding client. Taking Claude as an example, the official requirements include having MCP enabled, holding the appropriate administrative role, and using the instance’s MCP address and specified authentication settings. Authentication keys should only be handled by authorized administrators during the trusted configuration process and must not be included in chat logs, public repositories, or article examples.MCP Management Documentation,Claude Integration Documentation

Step 4: Verify the complete workflow from metric discovery to audit queries

The official MCP tool recommends first discovering the model, then exploring the fields, confirming the fields, querying the results, and finally checking the actual query sent to the warehouse. This indicates that effective validation should cover both “what definitions were used” and “what was ultimately calculated”; it is not sufficient to merely check whether the response text is coherent.MCP Tool Reference

You can use the following test questions to check the key points:

• “What was last quarter’s revenue? Please first confirm which definition of revenue is being used.”
• “List net sales by region, and specify the reporting date and return processing rules.”
• “Report the end-of-month inventory, rather than adding up the daily inventory figures for the month.”
• “Compare these results with the existing BI reports and list the different filter criteria.”

These are examples of test questions; they do not imply that the product will necessarily respond in the specified format.

Step 5: Establish Entry Requirements

Include accuracy, handling of unanswered questions, permission isolation, latency, and cost in the acceptance criteria. Rerun the question set after replacing the large model, modifying metric definitions, or adjusting the data table. If Canary has not yet been released, you can first perform these checks using your team’s existing testing processes and then evaluate whether to adopt its automated evaluation features at a later stage.

How does it compare to dbt, Looker, and Snowflake?

The following comparison is based on the positioning described in each manufacturer’s official documentation; it is not a performance ranking, nor should it be used to conclude that any particular solution is more accurate in all scenarios.

Basis for Comparison:dbt MetricFlow Documentation,Looker Conversation Analysis Documentation,Snowflake Cortex Analyst API Documentation.

A pragmatic approach is to conduct a pilot using the same set of real-world business problems, the same set of correct answers, and the same scope of permissions, and then compare development hours with total costs. Do not assume that just because a product uses terms like “semantic layer” or “MCP,” it offers the full range of capabilities provided by another product.

Pricing and Trials: What Do We Know So Far?

AtScale's current official pricing is as follows: DSO (Deployed Semantic Object)Based on this, measurement is performed using semantic elements such as metrics and dimensions that are in production and available for use. The pricing page lists two editions—Standard and Enterprise—indicating that there are no charges based on user seats or query counts; objects that are under development or not yet available to users are not included in the same production deployment metrics.This page does not list a standard price for placing an order directly; you should request a quote and confirm the scope of the contract.Official Pricing Information

When making a purchase, we recommend verifying the following additional details: which items are subject to billing; how duplicate deployments are calculated; how adjustments are made for usage exceeding the agreed-upon scope; whether all AI components are included; the trial period and scope of technical support; and who is responsible for covering the costs of the repository and model services. The items listed above are for reference purposes only and do not imply that any specific additional charges are imposed by the official provider.

Historical references to “starting at $2,500 per month” from a 2024 article can still be found online. This figure should not be taken as the current pricing for the AtScale for AI suite in 2026.Historical Pricing Articles

Restrictions You Should Be Aware Of Before Use

1. The new kits have not all been delivered yet. The evaluation and procurement contracts should separately confirm the availability status and scope of commitment for Ava, ACE, CoModel, and Canary.
2. This requires companies to make their own investments in modeling and governance. The semantic layer cannot automatically determine whether business rules are valid. Errors in source data, conflicting metric definitions, and omissions in permission configurations all require handling by a designated person.
3. MCP queries have clearly defined boundaries. The current tool requires queries to use the `SELECT` statement; it does not support CTEs, self-joins, or subqueries, and limits each query to a single model. These are limitations of the MCP interface layer and should not be mistaken for the full range of SQL capabilities available across the entire AtScale platform.Query Limits
4. Compatibility should be verified on a product-by-product basis. The official known issues page currently states that Databricks Genie Code connections are not supported; “Support for Databricks” should not be interpreted as support for all of its AI products.Known Issues
5. Just because the calculation is performed in the warehouse doesn't mean that no data leaves the warehouse. It is still necessary to verify what query results, prompts, or logs the AI client receives, as well as the corresponding data retention and access policies. These are deployment review considerations; security conclusions cannot be drawn based solely on the “on-premises computing” marketing claim.
6. When discussing accuracy rates, it is essential to consider the full scope of the data. The official website contains results for 410 business-related questions as well as results for 317 answerable questions and adjustments to the evaluation environment. Since the data scope and methodologies differ, they cannot be compared directly; moreover, test performance should not be interpreted as a guarantee of accuracy for any given business problem.Official AI Demo Instructions,Explanation of the Official Benchmark Report

In a nutshell

What makes AtScale for AI noteworthy is that it enables enterprise AI decision-making to be based on maintainable, quantifiable, and verifiable business metrics. For teams that already have a data warehouse, cross-tool analytics needs, and a foundation for metrics governance, it’s worth adding to their pilot list; if they’re only dealing with scattered files, or hope that AI will automatically understand the entire business without any rules, they should first assess whether the investment aligns with their needs.

The most valuable next step is to approach vendors with a controversial metric, a set of questions with confirmed answers, and clear budget constraints to request demonstrations and quotes, and then verify the capabilities they have already delivered.

Official Portal and Reading Materials

• AtScale Official Website
• AtScale for AI Announcement
• Official AI Technology Overview
• AI Connect and MCP Documentation
• Official Integration List
• Pricing and Quote Request

Date of data verification: October 17, 2026. Features, compatibility, and commercial terms are subject to change; procurement and implementation are subject to the vendor’s latest documentation and contract. This document does not claim to have completed installation, integration, performance testing, or accuracy testing.