From Rule-Based to Goal-Oriented Automation
Traditional RPA tools are deterministic. They rely on highly structured, predefined rules and workflows to mimic repetitive tasks typically carried out by human operators – logging into systems, transferring data between spreadsheets, or generating reports. While they can significantly improve throughput and reduce human error in narrow domains, they are inherently limited in their flexibility. A small change in a user interface or data format can cause an RPA bot to fail. They are fragile, difficult to scale, and offer little capacity to deal with unstructured data or edge cases.
AI agents, on the other hand, represent a leap forward in autonomy and intelligence. These systems combine language models (typically large foundation models like GPT or Claude), planning algorithms, retrieval-augmented generation (RAG), memory components, and integration frameworks into modular “agents” that can interpret goals, plan multi-step actions, execute those actions via tools or APIs, and adapt based on feedback or unexpected results. Crucially, AI agents are not programmed to follow a single rigid workflow – instead, they are given objectives, such as “triage patient referral emails and add relevant information to the EPR,” and figure out the best sequence of steps to complete that goal based on current context.
Technically, these agents operate using components like:
- Language model core – This is often a transformer-based LLM capable of understanding user queries, generating summaries, and writing code or data transformations.
- Planner module – Determines the sequence of actions needed to reach a goal. May include task decomposition and tree-of-thought reasoning.
- Tool abstraction layer – Interfaces with external systems – databases, EPRs, calendar APIs, SNOMED-CT services – via function-calling or API-wrapping plugins.
- Long-term memory – Stores context from past interactions or prior task outputs, enabling learning across sessions.
- State tracker – Keeps track of what the agent has already done and what remains, similar to a finite state machine but with greater flexibility.
This architecture means AI agents can ingest both structured and unstructured data (e.g. PDFs, free text, emails), cross-reference against clinical knowledge bases or internal policy documents, and then act on live systems in a way that is traceable and auditable. Compared to static RPA bots, AI agents are more resilient to change, more contextually aware, and more capable of intelligent interaction with both systems and humans.
Emerging Use Cases Within the NHS
Several high-value domains within the NHS are already primed for the integration of AI agents. One clear area is referral management. While RPA might be used to monitor an inbox and copy referral PDFs into a folder, an AI agent can extract relevant patient information from the document, check against local service eligibility criteria, assess urgency using contextual data (such as comorbidities or past admissions), and then populate the appropriate fields in an EPR or scheduling system. It can even generate a draft response back to the referrer, ask clarifying questions if information is missing, or escalate to clinical review if flagged as urgent.
Another promising area is outpatient follow-up triage. AI agents can review discharge summaries, identify if follow-up is required, compare against waiting list backlogs and prioritisation rules, and automatically suggest follow-up timeframes or alternative community care pathways. Unlike rigid rules-based engines, they can account for nuance – such as a patient’s recent unplanned admissions or flagged safeguarding concerns – by drawing on multiple, often siloed, data sources.
AI agents are also being trialled for digital document processing in pathology and radiology. They can read scanned images of lab results, extract key markers, link to historical patient trajectories, and surface potential anomalies or inconsistencies. Rather than replacing clinicians, they act as intelligent assistants – pre-populating draft summaries, highlighting gaps, or flagging unexpected trends for review. Because these agents can explain their reasoning – e.g. “flagged due to elevated creatinine and prior kidney disease noted in EHR” – they can support human oversight while accelerating throughput.
On the operational side, AI agents are well-suited to complex reporting tasks like CQUIN audits, FOI responses, or incident review summaries. Given a data dump and a set of compliance rules, an agent can synthesise structured reports, extract trends, and even draft written commentary. This reduces time spent by clinical governance and information teams, freeing up capacity to act on insights rather than just compile them.
Practical Considerations for Implementation
Introducing AI agents into a healthcare setting requires careful architectural planning and governance. Unlike RPA deployments, which are often peripheral to core clinical systems, AI agents require deeper integration with internal APIs, audit trails, and data governance policies. This starts with secure deployment – often in cloud-based container environments that support access control, sandboxing, and real-time monitoring.
Agents should be deployed using the principle of least privilege, ensuring they can only access the systems and actions necessary for their defined scope. Tool access (e.g. ability to update patient records or query sensitive databases) should be strictly managed via API gateways with full logging. This ensures every decision, data pull, or action taken by an agent can be reviewed and, if necessary, traced back to its origin.
Human-in-the-loop design remains critical, especially in clinical contexts. Agents can be configured to operate in “review mode,” where they draft outputs for human sign-off before anything is actioned – much like an assistant preparing case notes or email responses. As confidence builds and performance is validated, these thresholds can be adjusted. Additionally, prompt chaining and memory components must be carefully controlled to avoid context bleed or data leakage between tasks.
Model fine-tuning and custom instruction sets – especially when dealing with NHS-specific terminology, coding frameworks like SNOMED-CT, or local service names – can greatly improve performance. The use of RAG (retrieval-augmented generation) frameworks allows agents to dynamically pull in updated policy documents or clinical guidelines without needing to retrain the core model.
Charting a Path Forward
For NHS organisations looking to move beyond RPA and into intelligent automation, AI agents offer a flexible, scalable, and future-proof path forward. But their introduction should be driven by service design, not just technology enthusiasm. It’s essential to begin with problems that are both high-impact and containable – where agent performance can be evaluated, risks managed, and benefits clearly measured.
Digital leaders should create multidisciplinary Tiger Teams – clinical, technical, governance – to identify these opportunity areas, map out data flows, and define what “good” looks like. Partnering with specialist consultancies and development teams experienced in agent-based systems can accelerate delivery and reduce risk.
At 6B, we work with NHS clients to design and build intelligent agent-based workflows that integrate safely and securely with live systems. Our expertise spans not just the technology, but the organisational change, governance, and ethical frameworks required to ensure these tools enhance care rather than complicate it.
The shift from task automation to goal-driven autonomy isn’t just a technological leap – it’s a redefinition of what digital working means in healthcare. With AI agents, the NHS has an opportunity to offload repetitive work, empower frontline staff, and build more adaptive, intelligent services that respond in real time to the needs of patients and providers alike.