EBOOK
AtScale for Agentic AI: A Field Guide
How to give AI agents certified numbers over MCP today, with the SQL behind every answer kept on record.
On the BIRD-Interact business questions that have a correct answer, Claude Sonnet 5 got 46% right writing its own SQL. With an auto-generated AtScale semantic layer, it got 93%. That sounds good until an agent starts chaining steps. At 93% per step, a ten-step workflow is a coin flip, and a 100-step workflow gets every step right 0.07% of the time. This field guide shows how to put agents to work on certified numbers over MCP, with each answer’s SQL on file for your auditors.
What you’ll learn:
- How two wrong lines in a 600-line overnight reorder list get approved, and why an agent writing its own SQL against hundreds of tables has to guess what each number means and can guess differently on every run.
- What the BIRD-Interact benchmark shows about LLMs doing business math, and what compounding errors do to multi-step agent workflows.
- How the accuracy attestation layer sits between your agents and the warehouse, and the four MCP tool calls that take a question from
list_modelstoget_outbound_queriesand an audit record. - What agents get from AtScale (the MCP server, ACE, Canary and semantic modeling), worked through an example with inventory, exchange rates and customers who move regions.
- An eight-point agent-ready checklist, nine industry agents with the metric each one has to get right, and what a 30-day AtScale agentic workshop covers.
Two wrong numbers hide among 598 right ones
A retailer’s agent builds tomorrow’s reorder list overnight. The replenishment buyer checks it when she gets in.
The agent adds up inventory across weeks and orders 810 units short on line 209: 430 instead of 1,240. Then it counts a transfer twice and orders 780 too many on line 214: 2,340 instead of 1,560. Both lines look like the other 598, and she approves all 600.
The model usually isn’t the problem. An agent writing its own SQL against a few hundred tables has to guess what each number means, and it can guess differently every run.
Why 93% accuracy isn’t enough for AI agents
On 317 answerable BIRD-Interact questions:
- 46%: Claude Sonnet 5 on its own, writing SQL against the raw schema
- 93%: The same model with a semantic layer AtScale generated automatically
- 0.07%: The chance a 100-step workflow gets every step right at 93% per step
Getting one answer in 14 wrong is fine when a person checks every number. An agent that chains 100 steps carries each miss into the next one.
Preliminary results. The odds assume independent errors at each step. See atscale.com/bird.
Who this eBook is for
- Data and analytics leaders: Learn how to define each metric once and serve it to every agent, chat tool and dashboard.
- AI and platform engineers: Find the six MCP tools, how OAuth sign-in carries permissions, and the SQL record each call returns.
- Finance and operations: See which numbers an agent acted on, and pull the exact SQL that produced each one.
Read it before your agents work the night shift
See how agents get certified numbers from ACE over MCP, and how each answer leaves its SQL on file for your auditors.
Frequently asked questions
Why do AI agents return wrong numbers on business data?
An agent writing its own SQL against hundreds of tables has to guess what each number means, such as which tables to join and how to aggregate. It can guess differently on every run, so its answers aren’t consistent even when the model is capable.
Why isn’t 93% accuracy good enough for AI agents?
Errors compound when an agent chains steps. At 93% accuracy per step, a ten-step workflow gets every step right about 48% of the time, and a 100-step workflow about 0.07% of the time, assuming independent errors.
How do AI agents get certified numbers over MCP?
Through the AtScale MCP server, agents query metrics that are defined once in the AtScale semantic layer instead of writing their own SQL. In preliminary BIRD-Interact results on 317 answerable questions, Claude Sonnet 5 got 46% right writing its own SQL and 93% with an automatically generated AtScale semantic layer.
How can you audit what an AI agent did?
Every request an agent makes through AtScale keeps the SQL that produced the answer on record. Teams can check which numbers an agent acted on and pull the exact SQL behind each one using the get_outbound_queries MCP tool.