DocuSign pairs Claude Code with dbt tools and reports 50% productivity gain
DocuSign has piloted an AI workflow built on Claude Code, dbt v2, the dbt MCP server and dbt Agent Skills, and reports a 50 percent productivity gain from requirement to a validated, review-ready dbt asset. The results were shared by Bishal Gupta at dbt Summit 2026 and summarized by dbt Labs.
- DocuSign's Bishal Gupta reported a piloted 50 percent productivity gain from requirement to a review-ready dbt asset.
- Claude Code orchestrates the workflow, with dbt Agent Skills, the dbt MCP server and dbt v2 handling conventions, metadata and validation.
- Work is split into discovery, dbt development and orchestration phases so failures are retried individually.
- Guardrails include branch isolation, mandatory human review and a 1,000-row cap on test runs.
- Governance of teams' custom skills via a master skills file is described as early and still evolving.
Context switching, not SQL, was the bottleneck
According to dbt Labs' write-up of the session, Gupta, who worked on trusted data products and analytics engineering modernization at DocuSign, started from a simple observation: SQL generation was never really the bottleneck. Context switching was. Engineers moved between a Jira requirement, dbt conventions that lived mostly in people's heads, and repeated validation, with no clean handoff between stages of the work.
His answer was to make Claude Code the orchestrator of the workflow and give it three supporting pieces:
- dbt Agent Skills supply project conventions and naming patterns.
- The dbt MCP server supplies structured metadata and lineage.
- dbt v2 validates compiled SQL before a pull request goes out.
Phases instead of one large agent
Gupta did not build a single large agent to handle everything. He split the workflow into three phases: discovery, dbt development, and orchestration. The design means a failure in one phase, such as a failed compile or a skipped unit test, is retried on its own instead of restarting the whole pipeline.
In a live demo, Gupta asked the workflow to build a new quarterly aggregation model. It pulled in the relevant dbt Agent Skills and MCP context, generated the model and its YAML, then compiled and validated the result against dbt v2. It finished by producing a pull request and a matching orchestration job. Branch isolation was in place, and human review was still required before anything shipped.
Results and guardrails
Gupta reported a piloted 50 percent productivity gain from requirement to a validated, review-ready dbt asset. He also reported more consistent naming, testing, and semantic patterns across the team, which let reviewers focus on engineering judgment instead of style fixes. His framing, as relayed by dbt Labs: AI proposes, and engineers stay accountable for what actually merges.
Questions from the audience centered on guardrails. Gupta said test runs are capped at 1,000 rows so an agent cannot accidentally trigger a billion-row build. He also described an early, still-evolving approach to governing which teams' custom skills get layered onto the shared workflow, using a master skills file.
What it means for data teams
The account points to a pattern for teams that want to use coding agents on analytics work without loosening control. Conventions are written down as skills rather than left in people's heads, metadata comes from a structured source, validation runs before review, and a human approves the merge. The 50 percent figure comes from a pilot and is Gupta's own report, and the governance of shared skills across teams is, by his description, still taking shape.
dbt Labs says every keynote and breakout from dbt Summit 2026 is available on demand.
