--- title: "When AI Says \"Expected\": Automation Bias and the SUSAR That Slips Through" description: "Discover how automation bias in AI-supported SAE and SUSAR workflows can lead to missed expedited reporting obligations and increased pharmacovigilance risk." url: https://qbdgroup.com/en/blog/automation-bias-ai-susar-reporting-risk type: "Blog post" language: en published: 2026-07-08 author: "Anam Ahmad" category: "Pharmacovigilance" publisher: "QbD Group" citation: "QbD Group, \"When AI Says \"Expected\": Automation Bias and the SUSAR That Slips Through\", https://qbdgroup.com/en/blog/automation-bias-ai-susar-reporting-risk" --- # When AI Says "Expected": Automation Bias and the SUSAR That Slips Through > Discover how automation bias in AI-supported SAE and SUSAR workflows can lead to missed expedited reporting obligations and increased pharmacovigilance risk. **Artificial intelligence** is rapidly becoming part of pharmacovigilance operations. From case intake and triage to SAE processing and signal detection, AI-driven tools promise greater efficiency and consistency across increasingly complex safety workflows. However, efficiency introduces a different type of risk. **Automation bias** — the tendency to over-rely on system-generated outputs — can influence safety-critical decisions in ways that are difficult to detect until a reporting obligation is missed or a regulatory question is raised. In **SAE** and **SUSAR** workflows, the consequences are not theoretical. Over-reliance on automated classifications can contribute to: - incorrect seriousness assessments - missed expedited reporting requirements - incomplete medical evaluation - inaccurate expectedness determinations The risk becomes greatest precisely where uncertainty is highest.

In This Blog Post

- what automation bias means in pharmacovigilance operations - how an apparently reasonable AI classification can result in a missed SUSAR - why accountability remains with the sponsor under UK and EU regulations - practical governance measures to keep humans meaningfully in the loop
## Understanding Automation Bias in Pharmacovigilance Automation bias occurs when reviewers place **excessive trust in system-generated recommendations**, particularly when those recommendations are presented with confidence scores, structured outputs, and seemingly objective reasoning. Under operational pressure, an automated classification can quickly become the default conclusion. The reviewer accepts the recommendation, the case moves forward, and the independent medical judgment that the regulatory framework assumes a qualified professional will apply is gradually displaced. In pharmacovigilance, this creates a unique challenge. The objective is not simply to process cases efficiently. It is to **ensure that safety-critical decisions remain medically and scientifically defensible**. ## A Scenario Worth Sitting With Consider a Phase II oncology-naïve study in a rare inflammatory condition. The investigational product has a limited safety database and a sparse Reference Safety Information (RSI) document. A case arrives from a study site: a patient experiences sudden-onset severe headache, visual disturbance, and elevated blood pressure requiring unplanned hospitalisation. An AI-supported triage tool, trained predominantly on data from larger late-phase programmes with well-characterised adverse event profiles, identifies headache terminology already referenced within the RSI. The system classifies the event as: - expected - non-expedited - standard reporting pathway The confidence score appears high. The reviewing safety specialist accepts the recommendation without applying independent clinical judgement. The case proceeds through the standard 15-day workflow. ## The SUSAR That Slips Through The clinical reality is very different. The combination of acute headache, neurological symptoms, and hypertension may indicate **posterior reversible encephalopathy syndrome (PRES)**, a rare but potentially life-threatening condition not characterised within the RSI. Under **ICH E2A** principles, the case may qualify as: - unexpected - serious - potentially life-threatening As a result, it **may meet the criteria for expedited 7-day SUSAR reporting**. At this point, the regulatory reporting clock has already been running since sponsor awareness of the event. What appeared to be an efficient classification decision has now become a compliance risk. ## Why This Is Not an AI Failure It is tempting to describe this situation as a model failure. In reality, the technology performed exactly as designed. The system identified patterns, compared terminology against available training data, and produced a statistically plausible output. The problem is not that the AI generated an answer. The problem is that a confidently presented answer displaced the clinical reasoning that pharmacovigilance frameworks rely upon. This distinction matters. Early-phase development programmes often have: - limited safety databases - evolving risk profiles - restricted RSI content - higher uncertainty regarding unexpected events These are precisely the environments where historical training data is least representative and where human medical judgement provides the greatest value. ## Accountability Does Not Move One of the most important misconceptions surrounding AI adoption is the assumption that automation somehow transfers responsibility. It does not. Under both UK and EU pharmacovigilance expectations, **regulatory accountability remains with the sponsor**. Reporting timelines begin at the point of sponsor awareness. A missed SUSAR remains a sponsor responsibility regardless of: - which AI system generated the classification - which CRO processed the case - which vendor supplied the technology AI changes the nature of pharmacovigilance risk. It does not transfer ownership of that risk. ## Designing Safety Governance Around Human Oversight The solution is not to eliminate AI-supported triage. **The solution is to design governance models that ensure humans remain meaningfully involved in safety-critical decisions.** Many organisations are now adopting operating models where: - AI supports prioritisation rather than final determination - predefined escalation thresholds trigger mandatory review - confidence scores influence oversight requirements - medically qualified professionals retain final decision-making authority This approach aligns closely with emerging industry expectations for responsible AI deployment in regulated environments. The same principles appear throughout current guidance and industry recommendations: - define a clear intended use for every AI application - maintain medically qualified oversight for safety-critical decisions - avoid fully autonomous decision-making in pharmacovigilance workflows - train staff to recognise automation bias as well as system functionality Organisations that address human factors explicitly tend to achieve the greatest value from AI adoption while maintaining regulatory confidence. ## The Most Dangerous AI Output Is Often the Most Convincing One The organisations that adopt AI most successfully are not those that assume the technology will always be right. They are the ones that assume **it will occasionally be confidently wrong**. Automation bias emerges when confidence replaces curiosity. AI provides pharmacovigilance professionals with powerful tools to improve efficiency, consistency, and scalability. However, those tools do not replace the need for clinical reasoning. The strongest safety organisations are therefore not those that remove humans from the process, but those that **design systems where technology and expert judgement reinforce one another**. --- Source: https://qbdgroup.com/en/blog/automation-bias-ai-susar-reporting-risk — © QbD Group. Quote freely with attribution and a link back.