Indegene says its data-quality agent cut rule-writing from 30 hours to 4

Indegene, a life-sciences services company, says an LLM-powered agent has cut the time to write data-quality rules for a dataset from about 30 hours to about 4. The claim is in a post published on 4 November 2026 on the AWS Machine Learning Blog, co-written by two AWS solutions architects and Indegene's Shravan K S. It describes a Data Quality Agent inside Indegene's NEXT Unified Data Platform.

What the agent does

The agent drafts data-quality rules for a table, then generates SQL and Python for the ones a person approves. According to the post, it builds context from six inputs: catalog metadata, a data profile, anomaly results against historical baselines, existing rules, an external knowledge base of public standards, and an internal one holding glossaries, SOPs and past reviewer feedback.

The business rules are the point. The post gives examples such as market share summing to 100% within a territory and a prescriber's product eligibility matching their specialty, which null and type checks cannot catch. Users can also describe a check in plain English and get SQL back.

Stack and controls

• Model and runtime: Claude Sonnet 4.5 on Amazon Bedrock, with the agent built on Strands Agents and run on Amazon Bedrock AgentCore, per the post.
• Review: each rule gets a confidence score. Only high-confidence, low-risk rules are approved automatically; critical and business rules always go to a subject-matter expert.
• Validation: generated code gets syntax and schema checks and, where possible, a test run on sample data. A second LLM reviews rules written by the first, to reduce hallucinations.
• Audit: every decision is stored in a central, versioned rule repository.

The reported numbers

The post reports onboarding a new table now takes 2-4 hours instead of 2-5 working days. Its quality table estimates rule-suggestion accuracy at 80-90%, rule acceptance (approved without changes) at 60-75%, and a reviewer override or edit rate of 20-35%. SQL passes validation first time 90-97% of the time when grounded in the schema and 80-90% for custom SQL.

On cost, the post says up to 90% lower input-token cost on cached context with Bedrock prompt caching, and about 45-51% less compute from profiling metadata first. All of these sit under a footnote: "Estimated from Indegene's production deployments".

What the post leaves out

It gives no dollar cost per dataset or per month, no sample size, and no measured error rate for rules that were approved and later turned out wrong. The 30-to-4-hour comparison does not say who did the original work or how it was timed. Nothing describes a failure, a rollback or what Indegene would change. As with any vendor story, it is a design to learn from rather than a benchmark.

The control design is the reusable part. A 20-35% edit rate and a stated hallucination range mean the humans in the loop are doing real work, which is a reason to budget reviewer time rather than assume it disappears.