Real-world agentic AI adoption lags capability, study finds

Agentic AI has begun entering software engineering workflows, yet the gap between what these systems can do in theory and what companies are actually deploying in practice remains substantial, according to new research.

Study overview#

Researchers conducted interviews with 16 practitioners across 12 companies to assess how agentic AI adoption is progressing in the field. Using a six-level maturity framework to analyze the findings, they discovered that most production deployments cluster at the lower maturity tiers—Levels 1 through 3—while experimental capabilities at four companies extend beyond what's currently integrated into live systems.

Key barriers to deployment#

Four recurring obstacles emerged across the organizations studied:

• Context management: Difficulty maintaining proper context across agentic processes
• Performance on proprietary content: Systems struggle when applied to company-specific or closed-source material
• Non-determinism and qualification: Unpredictable outputs complicate verification and validation
• Data confidentiality: Concerns about sensitive information exposure during agentic operations

The core problem: verification gap#

The researchers framed these barriers as a capability-deployment verification gap—the mismatch between what agentic AI systems are capable of and what organizations feel confident using in production. This gap is structured by two interrelated dimensions: information asymmetry (uncertainty about how and why systems behave as they do) and qualification absence (lack of standardized ways to validate agentic systems for specific tasks).

Implications#

The findings suggest that agentic AI's transition from experimental tool to reliable production component hinges on solving not just technical challenges, but also establishing clearer qualification and verification frameworks that can bridge the trust gap between capability and deployment.