The agent patterns we actually deploy for data teams. Some reason over your data to reach a judgment; others execute the work once the judgment is made.
Translate plain-language business questions into governed SQL against your semantic layer, returning the result set alongside the query and lineage behind it.
e.g. Answers “why did EMEA margin drop last quarter” with the query, the tables and the drivers.
Correlate pipeline failures with upstream lineage, recent deploys and log patterns to isolate the actual cause rather than the first symptom.
e.g. Traces a failed dbt model back to a schema change in an ingested source three hops upstream.
Continuously evaluate freshness, volume, nullability and distribution drift, then judge which anomalies are genuinely worth an alert.
e.g. Flags a 40% row-count drop as a real incident while suppressing an expected month-end dip.
Convert legacy Oracle, Teradata, Netezza and SQL Server logic into target-platform SQL, with automated row-level validation of every conversion.
e.g. Rewrites 4,000 Teradata stored procedures into Snowflake SQL with a per-object diff report.
Review warehouse spend against query patterns to identify runaway queries, idle compute and mis-sized clusters, with concrete right-sizing recommendations.
e.g. Identifies a nightly full scan that a partition filter reduces to 4% of its original cost.
Propose dimensional models and incremental strategies from real query patterns rather than guesswork, with the trade-offs stated explicitly.
e.g. Recommends a star schema and grain after analysing six months of BI query history.
Classify sensitive columns across the estate, then propose masking policies, access rules and retention windows aligned to your compliance posture.
e.g. Detects unmasked patient identifiers in a staging schema and drafts the masking policy.
Generate table and column descriptions from actual usage and lineage, then keep them current as the underlying models change.
e.g. Documents 800 undocumented columns and flags 30 whose meaning has drifted from their name.
Build, backtest and monitor demand, revenue and capacity models, surfacing when a forecast has degraded enough to need retraining.
e.g. Detects that a demand model has drifted past tolerance after a pricing change.
Assemble evidence for ISO 27001, SOC 2, HIPAA and GDPR reviews by mapping controls to access logs, lineage and retention policy in force.
e.g. Produces an access-review pack showing who queried regulated tables in the audit period.
Run scheduling, retries, backfills and dependency resolution across your existing orchestrator, escalating only what genuinely needs a human.
e.g. Backfills 90 days after a late-arriving source, in dependency order, without manual triggers.
Onboard new sources by inferring schemas, mapping fields to existing models and generating the load and test scaffolding for review.
e.g. Onboards a new SaaS API and produces staging models plus tests within a single sprint.
Deduplicate and route incidents with lineage and blast-radius context attached, so on-call sees impact rather than a bare failure message.
e.g. Groups 40 downstream failures into one incident naming the single upstream cause.
Evaluate access requests against role definitions and data classification, provisioning what policy permits and routing the rest to an approver.
e.g. Grants read on a curated mart automatically, escalating any request touching PII.
Build, validate and distribute scheduled BI outputs, holding delivery when the underlying data fails its freshness or quality checks.
e.g. Blocks the Monday exec dashboard when an upstream feed has not landed, and says why.
Getting this wrong is the most common reason agent projects stall. The two categories need different guardrails, different evaluation and a different level of autonomy.
They investigate, weigh evidence and produce an answer a human will act on. The risk is a confident wrong answer, so they are evaluated against expected results and always cite their working.
Guardrail: every output traceable to a query, a table and a row.
They take action inside your systems — running jobs, provisioning access, delivering reports. The risk is an action that is hard to undo, so they run with approval gates and deterministic fallbacks.
Guardrail: writes, spend and destructive operations route to a named approver.
Tell us the workflow and we will tell you which pattern fits, what your data needs first, and what it takes to get it into production.