The certify-and-govern layer for AI-ready data

The enabler of high-accuracy AI access to corporate data.

Everyone is building AI on their corporate data. Almost no one has made that data correct for AI to read. DataStele certifies your metrics, metadata, and semantics once — so every AI client, across every warehouse, gets the golden number instead of a confident guess.

Request early access Read the story ↓

Databricks · Snowflake · BigQuery — one certified source of truth

175 CE · Luoyang · Han Dynasty

熹平石經

The Xiping Stone Classics — the world's first act of data governance.

For centuries, the Confucian classics survived only as hand-copied scrolls. After generations of transcription, no two copies fully agreed. Schools disputed single characters, errors compounded silently, and a scholar's career could turn on which version he had learned.

The scholar Cai Yong petitioned Emperor Ling to end the drift. Over eight years, the corrected, authoritative text of the seven classics was carved into forty-six stone steles and erected before the Imperial Academy.

When the steles were unveiled, history records that more than a thousand carriages a day crowded the streets — scholars coming to copy from the stone and correct their own texts against it.

One certified source. Carved once. Read by everyone. Trusted completely.

Eighteen centuries later, the same problem.

Then	Now
Countless hand-copies of the classics, each subtly wrong.	Countless definitions of "revenue," cryptic columns, undocumented tables.
Scholars argued over the authoritative character.	Three teams report three numbers; the AI confidently returns a fourth.
Carve the corrected text in stone — once.	Certify metrics, metadata & semantics in DataStele — once.
Every scholar verifies against the stone.	Every AI client reads the golden number, not a guess.

The problem, today

The plumbing works. The water is dirty.

Snowflake and Databricks shipped the pipes — governed gateways, semantic-layer scaffolding, certification fields. But a governed pipe pointed at cryptic columns and uncertified metrics produces confidently wrong answers with a perfect audit trail.

0.92 → 0.42

Text-to-SQL execution accuracy collapses when the schema is opaque — a ~58% silent-error rate.

~21%

State-of-the-art AI agents on enterprise-realistic data — versus 91% on the academic benchmark.

Silent

The worst failures: valid SQL matching the wrong business definition. No error. No signal.

"The hardest part of a data catalog isn't standing it up — it's filling it in." — the empty-container problem, in one line

What DataStele is

A certify-and-monitor control plane
that makes data readable by AI.

The MCP server is just the delivery surface. The product is two functions.

The Certify Gate

Nothing reaches the AI as authoritative until a human has signed off. Three artifacts certify through one draft → certified → deprecated workflow:

Metrics — the golden number, signed off by the governance team.
Metadata — descriptions and coded-value decodes, signed off by a data steward.
Access — classification clearances via Okta and RBAC, on every query.

Uncertified content is withheld, masked, or flagged — never served silently as truth.

The Monitoring Plane

A gate is only credible if its state is visible. DataStele makes governance provable, not just present:

Coverage — what's certified vs. draft vs. uncurated, and the gap report.
Audit — every query, its decision, and which certified artifact backed the answer.
Drift — when upstream changes, or raw SQL recomputes a certified metric.

Read-only reporting over events already recorded — the stele you can inspect.

How it works

Certify once. Every AI client reads the canon.

Certify

Define golden metrics and enrich metadata once. AI drafts, a human signs off. The definition is carved into the registry.

Govern

Okta authenticates the caller; classification-driven RBAC decides what they may read. PII, finance, compliance, and risk are gated uniformly.

Serve

AI clients call query_metric and get the golden number — compiled server-side. They cannot alter the computation.

Watch

Coverage, audit, and drift stream to the monitoring plane. When reality diverges from the canon, you see it.

Warehouse-agnostic

One connector interface. Databricks, Snowflake, BigQuery — the core never knows which is underneath. One certification spans them all.

Two-tier auth

Okta OIDC authenticates identity; classification-driven RBAC grants clearance to data classes, not tables. Tag a column once, every role's access is decided.

Golden-metric registry

Corporate definitions live in one native registry with a certification state machine — not scattered across dashboards that drift.

The certification agents can't absorb

An agent can discover what a column is. It can never certify which of five revenue definitions your CFO approved. That judgment is the moat.

The shift

From builder to governor.

As AI makes query generation cheap, the scarce and valuable work moves from building and generating to governing and validating. Fewer people write the 400th dashboard; more people own the definitions that dashboards — and now autonomous agents — depend on.

DataStele is the operating system for that shift. The place where business meaning is decided once, signed off by a human, and read correctly by every AI that follows.

Make your data readable to AI.

Carve your definitions once. Serve the golden number to every client, forever.

Request early access