Baseline Results & DORA 2025 Benchmark

DORA 2025 benchmark comparison

Prolaio is at or above benchmark on all five DORA headline metrics. Reliance (86% vs. 65%) is the biggest gap above global, suggesting engineers here lean on AI more heavily than the industry average. Code quality perception (77% vs. 59%) is also notably ahead.

Trust in AI-generated output

Trust tracks the global average at the top end (23% vs. 24% high trust). Prolaio has less low trust than global (23% vs. 30%), meaning the distribution clusters more toward the middle. More than half the org sits at "somewhat": they use AI but don't fully believe what it gives them. Open-text responses confirm this with repeated mentions of rework, hallucinations, and the need to carefully review AI output.

Extended benchmarks

Productivity: Prolaio's largest bucket is split between "moderately" and "slightly" (both 36%), while DORA's largest is "slightly" (41%). Prolaio engineers are somewhat more confident in their productivity gains. 5% report a decrease vs. 5% globally.

Code quality: 77% report improvement (vs. 59% DORA), but the shape differs: Prolaio has a heavier "slightly improved" bucket (50% vs. 31% DORA) and fewer "no impact" responses (18% vs. 30%). Only 5% report any worsening vs. 10% globally.

Task reliance: where we lead, where we lag

Prolaio is above the global benchmark on most tasks, often significantly. The two exceptions are code review (-11 pp) and security analysis (-19 pp), where Prolaio engineers use AI less than the industry average. Code review is particularly notable because multiple open-text responses flag review burden as a growing concern, yet engineers aren't turning to AI to help with review itself. Security analysis may reflect the regulated environment (21 CFR Part 11, SOC2) where engineers are cautious about AI in compliance-adjacent tasks.

AI reliance by task

Top 3: Writing new code, debugging, writing documentation. These are the "path of least resistance" tasks where AI provides the most obvious value.

Bottom 3: Creating specifications, code review, security analysis. These are higher-judgment activities where engineers either don't trust AI or haven't found effective workflows yet.

Individual vs. team gap

Both gaps are moderate and positive: people feel AI helps them personally more than it helps their team. This is expected for a baseline. The gap should narrow as AI moves from personal habit to team workflow. If it widens in future PIs, adoption is staying superficial.

Perceived impact

86% report AI increased their productivity (matches DORA global at >80%). 77% report improved code quality (notably above the 59% DORA benchmark). However, the open-text responses paint a more nuanced picture: several engineers flagged concerns about code quality degradation in MRs, AI-generated "slop," and the need for careful review.

Capability enablers

Policy clarity (3.36) has a wide spread: 4 people say "extremely clear" while 1 says "not at all clear" and 5 say "slightly clear." The policy exists but hasn't reached everyone evenly.

Experimentation support (4.14) is the strongest enabler score. The org is doing well at encouraging experimentation. This is a strength to maintain.

Internal data access (2.41) is the biggest gap and the most actionable finding. Engineers overwhelmingly report that AI tools can't see their codebase, Jira, Confluence, BigQuery, or Slack. This is where MCP, RAG, and codebase-aware tooling investments would register.

Shadow AI

KR status: not met. 2 engineers are still regularly using personal or unlicensed tools. The question is whether our licensed tools don't cover a workflow these engineers need, or whether they just haven't migrated.

Open-text themes

Key takeaways for PI planning

1. Adoption is real and broad. 95% of engineers use AI to some degree. The median engineer interacts with AI 4 hours per day. This is not a fringe activity.
2. Trust is the bottleneck. Of the three adoption legs, trust is the weakest (3.05/5) while reliance (3.41) and reflexive use (3.32) are both above the DORA benchmark. Engineers use AI a lot but don't fully trust what it gives them.
3. Internal data access is the biggest enabler gap. At 2.41/5, this is the lowest score in the survey. AI tools are working without context. Investing in MCP, codebase-aware tooling, and RAG would directly address both the trust and the data access scores.
4. Code review is an emerging pain point. Code review ranks near the bottom for AI reliance (2.05/5), and multiple open-text responses flag that AI-generated MRs are harder to review: more files changed, less clarity on intent, "code slop."
5. Shadow AI is nearly eliminated but not zero. 2 engineers still regularly use personal/unlicensed tools. The KR target of 100% is not yet met.
6. Response rate needs work. 52% is below the 75% minimum. For PI10, the pulse needs to be pushed harder.