Recoleta
GitHub
Source note

Agentic AI in Industry: Adoption Level and Deployment Barriers

Authors
Spyridon Alvanakis Apostolou, Jan Bosch, Helena Holmström Olsson
Source
arXiv
Published
Collection
Software Intelligence
Agentic AISoftware EngineeringIndustrial AdoptionMulti Agent SystemsHuman AI Interaction
Open arXivMarkdown source

Summary

This interview study finds that industry use of agentic AI in software engineering is still mostly at assistant or task-agent use. The main blocker is verification: companies can prototype stronger agents, but they cannot qualify their outputs for active workflows without human review.

Problem

  • It studies how companies adopt agentic AI in real software development workflows, where evidence is still limited.
  • The problem matters because unreliable AI output, weak traceability, data leakage risk, and poor fit to proprietary code can block use in safety-regulated and large legacy systems.

Approach

  • The authors ran semi-structured interviews with 16 practitioners at 12 companies across small, medium, and large organizations.
  • They assigned each company to a 6-level agentic AI maturity scale, with Level 0 for unsupported individual use and Level 5 for self-healing systems.
  • They compared cases across company size, regulation, active tools, SDLC tasks, reported limits, and experimental deployments.
  • They used two local LLMs, gpt-oss-20b and Qwen3-14B, to check structured interview summaries; 11 of 62 suggested additions were accepted after manual review.

Results

  • Current production maturity was low: 7 of 12 companies were at Level 1 AI Assistants, 4 were at Level 2 Task Agents, 1 was at Level 3 Collaborative AI, and 0 were at Levels 0, 4, or 5.
  • Regulation did not prevent all progress: among regulated companies, 5 were Level 1 and 3 were Level 2; among unregulated companies, 2 were Level 1, 1 was Level 2, and 1 was Level 3.
  • Four companies, C6, C7, C8, and C12, had experimental capabilities above their production maturity level, but could not move them into active workflows because output verification depended on human review.
  • The strongest reported deployment barriers were context-window limits for large and fragmented codebases, weak performance on proprietary languages and protocols, non-deterministic output that conflicts with qualification rules, and data confidentiality limits for cloud LLMs.
  • C7 reported that multi-agent workflows inside a copilot environment reduced bug-resolution turnaround from days or weeks to hours, but end-to-end agentic pipelines were still excluded from active development workflows.
  • The study reports no benchmark-style model accuracy results; its quantitative claims come from interview counts, maturity assignments, and the 16-interview, 12-company sample.

Link