DORA: The ROI of AI-Assisted Software Development

A joint report from Google Cloud and DORA defines AI as an 'amplifier.' Organizations with solid internal platforms, deployment pipelines, and team capabilities gain increased speed and delivery power from AI, while organizations with weak foundations end up shouldering greater technical debt and verification costs.

Right after adoption, a J-curve forms. Early productivity can actually drop as learning new workflows, the burden of verifying AI output (the Verification Tax), and bottlenecks in testing, approval, and deployment that fail to keep pace with faster code generation all overlap.

• With a mature IDP, automated guardrails, and AI-readable internal data, ROI grows quickly.
• Relying on manual testing, bureaucratic approval processes, and fragmented data can cause AI to accelerate maintenance costs and technical debt.
• The report cited productivity gains of 35–40% on Greenfield work, versus under 10% improvement on complex Brownfield code.
• Inference costs have fallen 280x from November 2022 to October 2024, and the real burden has shifted from model costs to verification, governance, and training costs.

ROI is calculated along three axes: Headcount Reinvestment Capacity, Extra Feature Deployment Revenue, and Downtime Impact. In a sample based on 500 people, the report presented a Year 1 investment of $8.4M (combining hard costs of $5.1M and J-curve costs of $3.3M), a Year 1 return of $11.6M, an ROI of 39%, and a payback period of about 8 months, while also mentioning a 3-year average ROI of 727% based on actual customer data. From Year 2 onward, compounding effects are expected as the shift moves from coding assistants to autonomous agents.

The execution roadmap has two stages. First, use CapEx to build the Context Layer — namely the IDP and AI-accessible data. Then, use OpEx to invest in the verification and training capabilities needed to develop engineers into high-level orchestrators of AI agents. The key performance metric is not commit counts but Experiment Frequency, and uncertainty should be addressed by dividing scenarios into conservative, realistic, and optimistic to persuade the CFO.

The appendix reinterprets METR's RCT on experienced open-source developers — where AI usage increased task time by 19% — through the lens of the J-curve. The message is that in environments lacking trustworthy guardrails, internal data, and an IDP, the verification burden can outweigh AI's speed gains.