How We Built an Autonomous Development Workflow at Nordic Corporate Bank

Value of the Information & Strength of the Argument

The talk provides valuable insights into a real-world implementation of an autonomous development workflow, offering concrete metrics and a detailed description of the system architecture. The speaker’s argument is persuasive, supported by specific examples and a live demo. However, the presentation is largely anecdotal and lacks independent verification or a critical examination of potential drawbacks, such as security risks or the long-term impact on code quality. The speaker does acknowledge some challenges, such as the need for human oversight and the new bottleneck of reviewing PRs, but the overall tone is optimistic and promotional.

Scientific Rigor, Source Quality, Title Accuracy

The talk is based on the speaker’s direct experience, which lends it authenticity, but it does not cite external sources or academic research. The title accurately reflects the content. The speaker mentions a report by DNB Carnegie on agentic workflows, but does not provide a specific reference. The talk is a personal account rather than a rigorous scientific study, and the lack of independent verification limits its scientific rigor.

The title accurately reflects the content, which details the construction and operation of an autonomous development workflow at a bank.

7/10

The talk provides a detailed, first-hand account of a real-world implementation of an autonomous development workflow in a banking context. The speaker presents concrete metrics (e.g., 572 PRs in March 2026 vs 141 a year earlier) and describes the system architecture. However, the presentation is largely anecdotal, lacks independent verification, and does not provide a formal evaluation of risks or limitations.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction and context: the speaker introduces himself and the bank, and explains the team size.
• 05:00The inspiration from a conference talk and the decision to build an AI agent.
• 10:00Introduction of Nils Skog, the AI agent, and its integration with Azure DevOps.
• 15:00Example of a product owner interacting with Nils Skog, showing the autonomous workflow.
• 20:00Live demo: the agent implements a search feature, showing the monitoring dashboard and PR creation.
• 25:00Discussion of the impact on productivity, with metrics showing a significant increase in PRs.
• 30:00Lessons learned: the need for structure, the distinction between coding and developing, and the new bottleneck of PR review.
• 35:00Security considerations: the agent's lack of access to customer data and the human approval step.
• 40:00Future plans: expanding the agent's role to handle more bank operations and become a full 'employee'.
• 45:00Under the hood: the architecture of the system, including triggers, API gateway, PostgreSQL, and container app jobs.

• NDC AI Conference — The talk was recorded at NDC AI in Oslo, Norway.
• NDC Conferences — The conference organizer, providing information about upcoming events.

• NDC AI Conference — The conference where the talk was presented, supporting the context of the presentation.