The Agentic Divide

In Wave 1 of our benchmark, 12% of Decision Makers qualified as Runners, the most AI-mature tier in Georgian’s Crawl, Walk, Run framework. In Wave 3, that share has risen to 22%.

In our view, the gap in AI maturity between companies looks less like a curve and more like a divide.

The three findings below describe:

• How Runners appear to operate differently across stack, practice and governance.
• How Runners are using agents in parallel and how spend on AI agents is growing faster than budgets.
• How the binding constraint on AI-driven engineering appears to be shifting away from productivity and reliability towards verification.

Georgian’s Crawl, Walk, Run framework segments Decision Makers into four tiers (Crawler, Walker, Jogger, Runner) based on how AI is operationalized inside the business. Runners are those who report AI projects in production at scale with significant budget commitment. For this analysis we combine Crawlers and Walkers into a single comparison group to create a large enough dataset for statistical analysis.

Across the dimensions where Wave 3 data shows the clearest Runner-vs-rest separation, four themes emerge: stack, practice, governance, and investment & ROI.

Stack. Runners appear to have invested in a more mature model and infrastructure stack.

Runners are more likely than Crawlers and Walkers to report a combination of custom or fine-tuned models in production, large reasoning models (LRMs) deployed and LLM observability tooling deployed.

These models and tools are associated with the multi-step reasoning, complex planning and long-horizon agentic execution that supports agentic AI.

Practice. Runners are 5× more likely to have Agentic AI live in production.

62% of Runners report having agentic AI live and expanding in production, against 13% of Crawlers/Walkers. Among Runners using automated coding tools, 53% report parallel-agent workflows (where most or all engineers are running multiple agents on distinct tasks concurrently) against 7% of Crawlers and Walkers.

Perhaps thanks to these agentic and infrastructure advantages, Runners are able to automate qualitatively harder problems using AI. 42% of Runners describe their internal AI workflows as mostly or very complex (multi-step processes spanning multiple tools or teams) against 10% among Crawlers and Walkers.

Governance. Runners appear to have built more guardrails which allow for greater agentic autonomy.

Across six categories of agentic guardrails (sandboxed environments, allowlists, human approval gates, policy-as-code, role-based permissions and monitoring or audit logs), Runners report an average adoption rate of 59%, against 37% among Crawlers and Walkers.

This investment in agentic guardrails may contribute to Runners having greater comfort with higher levels of agentic autonomy. 38% of Runners report comfort letting agentic AI operate with full or near-full autonomy in production, against 13% among Crawlers and Walkers. In our view, the guardrails come first: the eval, observability and policy-as-code investment is what lets Runners hand more decisions to AI without per-action human approval.

Investment and ROI. The median Runner reports AI accounting for 3× more budget share.

Runners report a 3× bigger share of their IT budget being dedicated to AI than Crawlers and Walkers. This budget is, in our view, what allows for greater investment in agentic infrastructure and guardrails.

In return, Runners are more likely to be expecting revenue from AI projects than efficiency gains alone. Runners are 16 points more likely to cite new revenue, rather than efficiency alone, as a top-three reason for investing in AI.

Based on Wave 3 responses, the cost of operating like a Runner appears to be rising. The data also suggests increased spend could surface as a budget question before it shows up in ROI tracking as revenue.

63% of Tech Decision Makers report engineers now running multiple AI agents simultaneously: 80% among Runners against 45% among Crawlers and Walkers. Meanwhile, almost half of Tech Decision Makers expect flat or declining entry-level headcount over the next 12 months.

Per-seat budgeting assumes cost rises with headcount. Agentic coding tooling cost tends to rise with usage, a different cost shape entirely.

The investment in coding agents is showing clear benefits.

Speed to production has accelerated across the board:

• 71% of Tech Decision Makers report moving AI features from pilot to production in under six months, and 25% in under three.

But the self-reported revenue impact is uneven:

• Runners report 70% positive AI impact on new revenue, against 39% among Crawlers and Walkers.

Most Tech Decision Makers cannot yet quantify what they are getting back. 45% report actively measuring AI ROI against financial outcomes (23% against revenue, 22% against cost savings) while 24% see benefits but have not translated them into specific revenue or cost impact. A further 15% have a financial framework but do not track against it, and 14% do not know their AI ROI at all.

In Wave 2 (June 2025), our data showed AI delivering strong gains on developer productivity metrics but lagging well behind on software reliability metrics, a 35-point gap between the two. The Wave 3 data suggests that gap has closed substantially: reliability has caught up to within 10 points of productivity.

71% of Tech Decision Makers now report AI yielding positive impact on developer productivity metrics, up 13 points from Wave 2. Reliability metrics moved further still: the average across four reliability metrics rose from 23% in June 2025 to 60% in May 2026, a 37-point increase.

In our view, while the data paints a rosy picture for productivity and reliability, it also raises a question about whether these are still the right metrics. Thanks to the extensive use of coding agents described in Finding 2, developer productivity and reliability appear to be moving toward table-stakes.

Nahim Nasser, Head of Engineering at Georgian’s AI Lab, points to verifiability and trust as the next inflection point. As agentic AI speeds up software development and a new wave of goal-oriented vibe-coding tools comes to market, the question shifts from how fast engineers can produce code to how confidently teams can verify what was produced. In practice, that means eval suites, runtime monitoring of agent behaviour, and human review workflows for high-stakes diffs.

This connects back to Finding 1: Runners already report the LLM observability tooling and the guardrail scaffolding that kind of verification depends on.

Runners appear to be operating AI differently, not at higher volume on the same stack, but with a different stack, a different practice and a different governance posture (Finding 1). That different way of working has a cost shape the per-seat budget model was not built for (Finding 2). And it leans on the eval, observability and guardrail layer that, in our view, will increasingly separate teams that can verify agentic output at scale from teams that cannot (Finding 3).