Back to the library

Key question New

AI-processed · Author-approved

Why does the same AI tool get used, bypassed, or repurposed across routine diagnostic work?

Recurrent AI Assemblage Forms in Radiological Interpretation Routines
Mehdi Berrahou · International Conference on Information Systems (ICIS) · 2026
Illustration for: Why does the same AI tool get used, bypassed, or repurposed across routine diagnostic work?

Loads Spotify's player; Spotify may set cookies. This device remembers your choice.

Key findings

  • The same AI artifact does not stabilize into a single workflow configuration, but instead recurs across three forms: sustained second-look verification, desynchronized bypassing, and aligned upstream triage.
  • Maintaining the prescribed second-look verification requires invisible compensatory labor from professionals, including re-entering closed files and absorbing task interruptions caused by processing delays.
  • Tool non-use stems from infrastructural latency and mismatches between the algorithm's training data and complex clinical cases, rather than outright resistance or distrust.
  • Clinicians occasionally mobilize the system upstream as a rapid triage filter when case complexity is low and technical performance aligns, despite official policies discouraging this practice.
  • Case-by-case adaptations occur with mutual visual awareness but minimal verbal discussion among peers, allowing divergent usage patterns to persist without departmental reconciliation.

What it means for you

CIO / IT Executive

Allocate budget to upgrade local server infrastructure and reduce algorithmic processing latency, directly targeting the technical delays that currently force radiologists to bypass the tool or perform invisible compensatory labor.

IT Manager

Conduct a Monday morning audit of network bottlenecks and file-closure protocols to eliminate the need for radiologists to manually re-enter closed patient files during second-look verifications.

Business Strategist

Redesign vendor ROI and workflow models to account for the actual tripartite usage patterns (second-look, upstream triage, and bypassing) rather than assuming a uniform, prescribed reading mode.

Researcher

Initiate an observational study focusing on how clinicians achieve mutual visual awareness during unarticulated, case-by-case workflow adaptations in shared reading rooms.

Policymaker

Revise clinical governance frameworks to formally evaluate and reconcile the discrepancy between official policies discouraging upstream triage and its actual practical utility in low-complexity cases.

Summary

This qualitative study investigates how an AI-based pulmonary nodule detection tool is integrated into clinical interpretation routines at a cancer center over an eighteen-month period. The researcher triangulated 23 semi-structured interviews across vendor staff and hospital clinicians, direct workplace observations, technical documents, and a structured diary kept by a radiologist across reading sessions.

Problem

AI deployment in specialized work environments frequently produces unintended consequences, including non-use, extra verification labor, and workflow fragmentation. Current management and clinical frameworks typically view AI adoption as a uniform reading mode or an individual trust decision, failing to explain how the same tool can be enacted in fundamentally different ways across daily tasks.

Transcript
Host: Welcome to A.I.S. Insights — powered by Living Knowledge. I'm Anna Ivy Summers.

Expert: And I'm Alex Ian Sutherland. Today we are examining a study titled "Recurrent AI Assemblage Forms in Radiological Interpretation Routines" by Mehdi Berrahou, presented at the International Conference on Information Systems in 2026.

Host: The study explores how an AI tool for pulmonary nodule detection was integrated into clinical routines in a cancer center over eighteen months. Alex, what is the core organizational problem the author addresses?

Expert: When organizations deploy AI, management often assumes workers will settle into a single, uniform workflow, framing non-use simply as a matter of trust or individual resistance. In medical imaging, deployments are typically designed as a second-look verification step. Yet in practice, AI integration frequently produces unintended friction, added verification work, and workflow fragmentation. The study investigates why the exact same software can be enacted in fundamentally different ways across daily tasks.

Host: How did the researcher capture these operational dynamics?

Expert: Mehdi Berrahou conducted an eighteen-month qualitative investigation combining twenty-three semi-structured interviews across vendor staff, radiologists, and clinical administrators, alongside direct workplace observations and technical documentation. To capture situated actions that interviews often miss, the study also utilized a structured diary kept by one radiologist over seven reading sessions, followed by an explicitation interview.

Host: What did the findings show regarding how professionals actually interact with the AI system?

Expert: The study found that instead of stabilizing into one standard routine, the tool repeatedly manifested in three distinct forms. The first is the sustained form, where radiologists use the tool as prescribed for second-look verification. However, keeping this sequence functional required invisible compensatory labor. Because processing on vendor servers took ten to fifteen minutes, clinicians had to interrupt their workflow, move to another patient, and later reload closed cases to review the AI's output.

Host: That introduces substantial cognitive friction. What were the other two forms?

Expert: The second is the desynchronized form, which resulted in bypassing the tool. This non-use stemmed from infrastructural and clinical misalignments. For instance, the AI was trained on general screening cohorts without cancer history, causing high false-positive rates on cancer patients with scarred lung tissue. When combined with server delays that caused results to arrive after report dictation, clinicians regularly completed examinations without the tool.

Host: And the third form?

Expert: That is the aligned form, where the AI is mobilized upstream as a triage filter. In straightforward cases with timely processing, clinicians leveraged the tool's high negative predictive value and immediate summary display to quickly rule out abnormalities. Although leadership and vendor guidelines discouraged this practice due to concerns about anchoring bias, clinicians occasionally enacted it to maintain reading tempo.

Host: Interestingly, the study points out that while clinicians observed colleagues adapting their usage, they rarely talked about it.

Expert: Exactly. Radiologists operated in shared physical workspaces and visually noticed their peers using alternative workflows, such as dictating before results arrived. However, these adjustments were rarely verbalized due to time pressure and professional sensitivities around automation. Consequently, divergent practices persisted without departmental alignment.

Host: What are the key takeaways for business leaders and technology managers deploying AI?

Expert: First, evaluate AI integration beyond standalone technical accuracy or user trust metrics. Actionability depends heavily on operational infrastructure, such as network latency and synchronization with legacy systems. Second, account for hidden compensatory costs; requiring workers to verify delayed outputs can cause task fragmentation that negates efficiency gains. Finally, organizations need active forums for reflective communication. Involving middle management to facilitate structured feedback between end-users, IT, and vendors allows localized adaptations to be addressed constructively rather than persisting as uncoordinated workarounds.

Host: That provides a very clear, practical perspective on AI implementation. Thank you for breaking down the study, Alex.

Expert: My pleasure, Anna.

Host: And thank you to our listeners. Join us next time on A.I.S. Insights — powered by Living Knowledge.
Artificial Intelligence & Automation Information Systems & Management Human-Computer Interaction Data Analytics & Big Data Digital Transformation & Strategy Cybersecurity & Privacy Cloud & Infrastructure E-Commerce & Digital Platforms Operations & Supply Chain Sustainability & Green IS

artificial intelligence, organizational routines, sociomaterial assemblage, radiology, healthcare, qualitative research