MEET ATSCALE / SKILLS AND SEMANTIC LAYERS
Skills got Claude to 95%. Then the data changed.
Skills make AI better at analytics. Each one is a hand-written piece of a semantic layer, and keeping hundreds of them right takes a team, tooling and testing most companies don’t have.
WHAT ANTHROPIC FOUND
From 21% to 95%, then back to 65% in a month.
In June, Anthropic’s data team published how they got Claude to answer analytics questions accurately. Without skills, Claude didn’t get past 21% on their evals. The team wrote metric definitions and reference files by hand, one business area at a time, and accuracy rose above 95%.
FIG. 1 · ACCURACY AT ANTHROPIC, BEFORE AND AFTER SKILLS
Their data model changed every day, and the skills that described it didn’t. In their words, without active maintenance, skill docs are “wrong within weeks.” Most companies don’t have a team to keep that up.
THE PROBLEM WITH SKILLS ALONE
Where skills break down.
Handcrafted skills don't scale.
An enterprise has dozens of domains: finance, where most teams start, then sales, marketing, support, manufacturing, product and research. Each has its own metrics to define, calculate, test and keep current. Writing each metric by hand, one domain at a time, means building a semantic layer from scratch. It works for a few metrics in one area and stalls long before it covers the business.
Skills go stale as the data changes.
Tables, columns and business rules change every week. Nothing updates a skill when they do, and nothing tells you when an answer starts to drift.
A skill can't tell when a question is vague.
“How are comps doing?” leaves out which stores count, which calendar to use and which period to compare. A skill can define same-store sales correctly and still answer the wrong version of the question.
Complex metrics need an engine built to compute them.
Same-store sales has to be calculated separately for each store, each region and the total, on the fiscal calendar. A skill can hold tested SQL for the questions someone thought of, but a new filter, level or time period means new SQL, written by hand or by the model. A computation engine applies the metric’s rules to any question, the same way every time.
Skills belong to one model and one tool.
Most enterprises run more than one AI model, plus Excel, Power BI and their own agents. Skills written for one model have to be rebuilt or copied for the rest, and the copies drift apart.
WHAT ATSCALE ADDS
AtScale is the semantic layer those skills rebuild by hand.
AtScale defines every metric once and calculates it the same way everywhere. With Canary, coming this fall, it’ll also test each metric on a schedule, so accuracy doesn’t drift. On the BIRD-Interact benchmark, adding AtScale made an LLM twice as accurate.
| Problem with skills alone | What AtScale does |
|---|---|
| Doesn’t scale across domains | One semantic model covers every domain, from finance to support. AtScale Semantic Modeling uses AI to draft each domain’s metric definitions from your data and query history, and a data engineer reviews and approves each one. |
| Goes stale | Definitions are version-controlled in Git and reviewed like code. AtScale Canary, arriving this fall, will run your questions on a schedule and flag any answer that changes. |
| Can’t spot a vague question | AtScale Ava (Ask, Verify, Answer) asks which metric you meant before anything is calculated. |
| Complex metrics | AtScale ACE (the AI Computation Engine) is a semantic computation engine built for complex metrics. It applies each metric’s rules to any question and shows the SQL it used. |
| Tied to one model and tool | One semantic model serves every AI model, agent and BI tool, through MCP, SQL and Excel. |
For how each product works, see What AtScale does and The path to 100% AI accuracy.
WHERE SKILLS FIT
Write skills for what’s unique to your business and use AtScale for the rest.
Skills are still useful for your own processes and terms. AtScale handles the parts every company needs and few have a team to build: defining metrics, clarifying questions, calculating answers and testing them.
Book a workshop.
- Bring us your most controversial metric and the business questions your team and your agents ask about it every week.
- We benchmark your AI on them, with AtScale and without it.
- You keep the semantic model we built to run the test.