Healthcare & Life Sciences in the Age of Agency

Executive summary

• Agents are already in daily use. Regulators, health systems and pharmaceutical companies have moved from AI that drafts and summarises to AI that carries out multi-step work. Examples include FDA's agentic tools for review staff, health-system agents built inside Epic's electronic health record, and more than 12,000 agents created by clinicians in Singapore's public system.
• The value is in workflow, and so is the risk. Most early gains sit in administrative and operational work: chart review, scheduling, prior authorisation and medical enquiries. That work borders clinical judgement, which is where regulators are now looking closely.
• Governance is converging across markets, but with different timelines. The FDA–EMA joint principles, Singapore's AIHGle 2.0, the TGA's software review and the amended EU AI Act all point to the same demands: lifecycle monitoring, a clearly defined context of use, and named accountability. Boards should build to that common standard now and not wait for local deadlines.

What's changing

Regulators are using agents themselves. In December 2025 the US FDA gave staff agentic AI capabilities. These are voluntary tools for multi-step work across premarket reviews, post-market surveillance, inspections and compliance, and they build on the generative assistant Elsa, launched in June 2025 ^[1]. The agency that reviews industry submissions is now using the same class of technology, which is likely to raise its expectations of sponsors' own AI practice.

Health systems are building agents inside the clinical record. At HIMSS26, Epic previewed Agent Factory, a platform for health systems to build, customise and monitor their own agents within the electronic health record ^[3]. Early deployments include:

• At Advocate Health, a pharmacy agent that summarises patient history to support medication-order verification, and infusion agents that handle scheduling and patient communication ^[2].
• At ECU Health, a transfer-centre agent that the system reports saves about 20 hours of chart review a week ^[2].

Epic also reports customer results from its revenue-cycle and patient-facing tools, including faster prior-authorisation submissions ^[3]. These are vendor-reported figures and should be treated as such.

Asia-Pacific is building agents at system level. Singapore's national health-technology agency, Synapxe, runs AgentSea, a sector-wide platform on which public healthcare professionals have created more than 12,000 agents. Use cases include pre-consultation patient summaries. Ministers describe adoption as early-stage ^[4].

Pharma is moving from pilots into regulated workflows. Daiichi Sankyo has used agentic AI for medical enquiries from patients and healthcare practitioners in Europe and Canada. It integrated the system with legacy platforms and spent around nine months on legal review before launch. The company plans to extend it to protocol writing and market-access dossiers ^[5]. More broadly, a Deloitte survey of 100 US health-system and health-plan technology executives found that 61% are already building or implementing agentic AI, and 85% plan to increase investment over the next two to three years ^[6].

Through the Age of Agency lens

Age of Agency: delegation becomes a clinical governance question. When AI only answered questions, a clinician stood between every output and the patient. Agents remove some of those checkpoints. They pull records, draft documents, message patients and queue tasks. FDA now asks directly whether autonomous multi-step action, use of external tools and fewer chances for human review need extra oversight ^[10].

For health executives, the practical question is no longer which tools to buy. It is which decisions the organisation is prepared to delegate, to which agent, under whose authority, and with what fallback when the agent is wrong. Organisations that treat that as an operating-model decision, rather than an IT project, will compete better for scarce clinicians and for payer and regulator trust. BCG's advice to concentrate on a small number of high-value use cases rather than many pilots reflects the same discipline ^[7].

Agentrification: who gains and who loses as tasks move. The first tasks to move to agents are the coordination work around care: chart review for transfers, discharge preparation, scheduling, prior authorisation and medical-information responses ^[2][3][5]. That shifts value in three ways:

• Roles change shape. Transfer nurses, case managers, pharmacists and medical writers spend less time assembling information and more time on judgement. Without deliberate role redesign, organisations may simply add volume to these roles. BCG's estimate that 70% of the effort lies in people and processes is a sensible planning assumption ^[7].
• Power shifts to platform owners. When agents are built inside the record system, the vendor that owns the workflow gains leverage. Health systems need their own evaluation capability and must keep their agent logic and data portable.
• Aged care is exposed first. Australian aged care already uses digital scribes, multilingual communication tools and assessment aids to free frontline staff for direct care ^[15]. The same sector faces sharp concerns about companion chatbots and monitoring of vulnerable residents ^[15]. Providers that agentrify thoughtfully can ease workforce pressure. Those that automate the relationship itself risk regulatory and reputational damage.

UBInt: shared agent infrastructure as public-health capability. Singapore's model, where a common platform (Tandem) and a sector-wide agent environment (AgentSea) are available to every public institution, treats capable AI as shared infrastructure, not as a privilege of individual hospitals ^[4]. That is the Universal Basic Intelligence argument applied to a health system. It narrows the gap between well-funded academic centres and smaller providers, and it creates one place for governance. Other jurisdictions in the region, including New Zealand's national system and Australian state health services, should ask whether shared agent infrastructure would serve equity and safety better than each institution procuring on its own.

Risks and governance

Key risks

• Function creep into regulated territory. An agent that summarises notes is usually administrative. An agent that suggests diagnoses or orders tests without clinician oversight may be a medical device. The UK MHRA clarified this in July 2026, noting that disclaimers do not override contradictory marketing claims ^[14]. The TGA has flagged that some digital scribes may already meet the medical-device definition ^[13].
• Drift and silent degradation. Agent behaviour changes as models update and data shifts. Point-in-time validation is not enough ^[9][13].
• Accountability gaps. Australian law assigns offences to "a person", which is hard to apply where AI replaces human involvement ^[13].
• Vulnerable populations. Older people and patient-facing conversational tools carry higher risk of manipulation and over-reliance ^[10][15].

Regulation and guidance

• United States. FDA's August 2026 discussion paper on generative AI-enabled devices proposes judging risk by how independently the software acts and how severe the harm from a wrong output would be. It considers competency-based premarket benchmarking and asks specifically about agentic systems. Comments close on 19 October 2026 ^[10].
• Europe and the US jointly. In January 2026 the FDA and EMA published ten non-binding principles for AI across the medicines lifecycle. They include a risk-based approach, a clear context of use, evaluation of the human–AI interaction, and lifecycle management including checks for data drift ^[9].
• European Union. The Digital Omnibus on AI came into effect on 27 July 2026. It moves the application date for high-risk AI embedded in regulated products, which include medical devices, to 2 August 2028 ^[11][12]. This is a delay, not an exemption.
• Singapore. AIHGle 2.0, issued by MOH and HSA in March 2026, sets documented responsibilities for developers, deployers and users. It covers both clinical and clinical-operations AI and expects post-deployment monitoring and incident reporting ^[4][8].
• Australia. The TGA's review of medical device software, including AI, recommends urgent work on digital scribes and on adaptive and generative AI, and signals more compliance action ^[13]. The Aged Care Quality and Safety Commission's own AI policy, which keeps final decisions with staff and maintains a register of AI uses, is a useful governance benchmark ^[16].
• Global. WHO's guidance on large multi-modal models remains the reference point for ethics and governance, and it notes that their broad capabilities are not yet proven ^[17].

Questions for boards and leadership

1. Which decisions and actions have we delegated to AI agents today, and is there a single register that names who is accountable for each one?
2. For each agent, have we defined its context of use clearly enough to know when it has moved from administrative work into territory regulated as a medical device in each market where we operate?
3. How do we detect drift or degraded performance after go-live, and who has the authority to suspend an agent?
4. Where agents remove human checkpoints, what has replaced them, and has clinical governance approved that change?
5. Are we redesigning roles so that time freed by agents goes to care and judgement, or are we just adding volume?
6. How dependent are we on one platform vendor's agent ecosystem, and can we evaluate and move our agents independently?
7. For aged care and patient-facing services, what lines will we not cross in automating relationships with vulnerable people?

How KIG helps

• KIG Advisory. We help boards and executive teams decide what to delegate to agents, redesign operating models and roles around those decisions, and build governance that works across the different regimes in Asia-Pacific, the US and Europe.
• KIG Research. We apply the Age of Agency, Agentrification and UBInt frameworks to healthcare and life sciences, turning regulatory change and deployment evidence into practical decision tools for leadership teams.
• KIG Ventures. We draw on what we learn from building and backing AI companies, including how agents behave in production, where integration fails, and what makes governance workable, and bring that experience to our advisory work.