Intelligent Health.tech Issue 35 | Page 26

F E A T U R E
AI projects can stall without strong foundations
Striking the right balance
Despite Agentic AI’ s potential advantages, some healthcare-sector adopters remain understandably wary. There’ s concern, for example, that trusts may accidentally unleash an AI system they can’ t control. Recent disclosures from OpenAI, Anthropic and Meta show that this is not merely a theoretical risk. During cybersecurity evaluations, their AI agents reached real external systems and took actions outside the intended boundaries of the tests that resulted in unauthorised access to third-party services.
These incidents exposed weaknesses in network isolation and oversight, underlining why containment, permissions and configuration controls are core governance requirements, not merely technical details. With these measures in place, an agent’ s access to systems and data can be tightly controlled so it focuses only on information relevant to its intended task and does not expose sensitive data to unauthorised staff.
Strict controls and solid foundations
Agentic AI systems can trigger workflows, connect multiple data sources or automate operational tasks – all of which can help reduce administrative burden. However, NHS Trusts need to define where an agent enters a process, which data it can access and what it may automate. It’ s also important to define when a human needs to provide oversight or take over completely.
Those decisions need to be supported by strong governance and reliable data. Information should be consolidated so the system and its users can work from a dependable source of truth. For NHS trusts, these controls are also essential to maintaining confidentiality and meeting their legal and ethical obligations to protect sensitive information. Without those foundations, trusts risk scaling inefficiencies and operational problems rather than resolving them.
Learning starts when the agent goes live
Employees intending to use new AI systems will need formal training and education but
26 www. intelligenthealth. tech