AI Features

AI is the foundation of Measured, not an add-on. Measured uses AI to help you author rules faster, explain failures clearly, detect drift, and, through a native MCP server, let your own LLMs and agents operate the platform directly. Trustworthy data in, trustworthy AI out.

MCP, connect your own LLMs

Measured ships a native Model Context Protocol server. Point Claude, your in-house copilots, or autonomous agents at Measured and they can:

  • Read data-quality status, health scores, and recent failures
  • Author and edit rules on assets and assemble rule groups
  • Investigate failures in natural language and propose fixes
  • Trigger validation runs and report results back into a conversation

It's AI managing data quality for AI, the same models that consume your data can help keep it clean. On Enterprise, the MCP server and all AI features can run against an on-prem or self-hosted LLM via the Privacy Agent, so nothing leaves your network.

Ask AI

A contextual assistant is available throughout the app. Ask AI knows your current org, workspace, and team context, links to the relevant docs, and helps you author rules, interpret results, and find your way around, so nobody has to read the manual.

AI rule generation

Instead of manually configuring expectation parameters, describe what good data looks like in plain English:

"The orders table should always have at least 1,000 rows, the order_id column should be unique and never null, the status column should only contain values like 'pending', 'shipped', 'delivered', 'cancelled', and the total_amount should be positive and less than 1 million."

Measured generates a complete set of rules from this description:

[
  { "rule_type": "expect_table_row_count_to_be_between", "kwargs": { "min_value": 1000 } },
  { "rule_type": "expect_column_values_to_be_unique", "kwargs": { "column": "order_id" } },
  { "rule_type": "expect_column_values_to_not_be_null", "kwargs": { "column": "order_id" } },
  { "rule_type": "expect_column_values_to_be_in_set",
    "kwargs": { "column": "status", "value_set": ["pending", "shipped", "delivered", "cancelled"] } },
  { "rule_type": "expect_column_values_to_be_between",
    "kwargs": { "column": "total_amount", "min_value": 0, "max_value": 1000000 } }
]

Review, edit, and save. The AI drafts; you approve.

Anomaly explanations

When a checkpoint run fails, click Explain next to any failing expectation to get a plain-English explanation:

"15 rows had a total_amount value greater than 1,000,000. The largest unexpected value was 9,847,320 in row with order_id=88421, which was created on 2026-05-23. This may indicate a data entry error or a currency unit mismatch (e.g., values stored in cents rather than dollars)."

Explanations include:

  • Plain-English description of the failure
  • The observed unexpected values (up to 10 rows)
  • Likely causes
  • Suggested remediation steps

Rule suggestions from a profile

After connecting a data source and profiling an asset, click Suggest rules to have Measured automatically draft a starter set of rules on the asset based on:

  • Column data types
  • Column names (e.g. columns named id, email, created_at get type-appropriate rules)
  • Nullable constraints from the schema

This gives you a solid baseline to refine rather than starting from scratch.

Configuration

Using Anthropic API (default)

AI features are available on all plans. On the Free plan, bring your own Anthropic API key. On Enterprise and above, credits are included.

Set your API key in the Measured environment:

ANTHROPIC_API_KEY=sk-ant-...

All AI features will use the claude-3-5-haiku model by default. Override with:

ANTHROPIC_MODEL=claude-3-5-sonnet-20241022

Using Privacy Agent (on-premises)

See the Privacy Agent documentation to run AI features without sending data to Anthropic.

Disabling AI features

To disable AI features entirely (removes the AI rule generator and Explain buttons from the UI):

ANTHROPIC_API_KEY=  # unset or empty

Token usage and cost

AI rule generation uses approximately 500–2,000 input tokens and 200–500 output tokens per request, depending on the complexity of the description and the number of expectations generated. Anomaly explanations use 1,000–3,000 tokens.

At claude-3-5-haiku pricing, typical usage across a team of 10 costs less than $5/month.

To avoid unexpected costs, set a monthly spend cap in your Anthropic console.