LinkedIn·Tuesday, 18 August 2026·8d ago

Human-in-the-loop sounds reassuring. Until you calculate how many humans you actually need. ⚖️ Many early AI governance models rely on…

Vivanti

1,003 followers

Human-in-the-loop sounds reassuring. Until you calculate how many humans you actually need. ⚖️ Many early AI governance models rely on intensive human review. That makes sense in a pilot. At scale, it can become the point of failure. If every output requires someone to validate each statement, compare it with approved sources, assess medical nuance, check local requirements and document the decision, the process may be controlled - but too slow to deliver the benefit AI was introduced for. The alternative is not weaker oversight. It is more deliberate oversight. 🛡️ Teams need to decide what can be anchored to pre-approved sources, what can be checked automatically, what requires specialist escalation and where human judgement is non-negotiable. That means different controls for different jobs: An internal practice simulation does not carry the same risk as patient-facing medical information. A draft localisation check is not equivalent to a final MLR decision. A content-insight tool does not require the same oversight as a system influencing clinical action. This is also why the latest Medicines and Healthcare products Regulatory Agency AI Airlock work is interesting. The Airlock focuses on AI-enabled medical devices, so it is not a blueprint for every pharmaceutical AI use case. But the underlying principle is relevant: controls should reflect intended use, evidence requirements and risk. At Vivanti, we see this across very different AI use cases - from training and medical content to content operations and analytics. That breadth makes one point clear: There is no credible universal 'human-in-the-loop' design. The reviewer, evidence standard, escalation route and acceptable latency have to be defined for the specific job. ⏱️ Governance fails when it is too weak to manage risk. It also fails when it makes responsible use operationally impossible. Reflection: Has your AI governance model been tested against the volume of human review it will create at scale? #ResponsibleAI #AIinPharma #MHRA #AIGovernance #LifeSciencesInnovation

♥ 2💬 1↻ 1

View on LinkedIn