Healthcare workflows often combine fixed rules, discriminative models, generated content, interfaces, software bots, agents, and human judgment. This guide shows how the pieces relate and how to evaluate each one.
AI systems infer from inputs how to produce predictions, content, recommendations, or decisions.1
Automation uses technology to carry out steps in a process and can reduce human intervention through predetermined logic embodied in software or machines.2
A product name rarely reveals the method used in every step.
Rules-based, AI-assisted, agentic, and hybrid can all describe different system designs.
Use the simple line for first exposure. The detailed line closely follows the authoritative definition linked in its footnote.
Automation carries out steps in a process. AI infers outputs from inputs. Discriminative AI predicts or classifies, and generative AI creates content. An AI agent acts toward a goal. Agentic AI coordinates one or more agents across tasks and tools.
Each footnote opens the primary source used for the detailed definition.| Term | Simple definition | Detailed, source-based definition |
|---|---|---|
| AI system | An AI system infers from inputs how to produce outputs such as predictions, content, recommendations, or decisions. | An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment.1 |
| Automation | Automation uses technology to carry out steps in a process. | NHS England describes automation as a wide range of technologies that reduce human intervention in processes. Human intervention is reduced by predetermining decision criteria, subprocess relationships, and related actions and embodying those predeterminations in software or machines.2 |
| Discriminative AI | Discriminative AI uses relationships in data to make predictions or assign information to predefined categories. | In a WHO technical brief, discriminative models analyze relationships between variables to make predictions. Classification models learn patterns and assign data to predefined categories.3 |
| Generative AI | Generative AI creates new content, such as text, images, or video. | “Generative AI is a category of AI techniques in which algorithms are trained on data sets that can be used to generate new content, such as text, images or video.”4 |
| AI agent | An AI agent perceives and acts on its environment to achieve a goal. | AI agents are systems that can perceive and act on their environment, often autonomously, to achieve specific goals and adapt their behaviour in response to changing inputs or contexts.6 |
| Agentic AI | Agentic AI uses one or more coordinated AI agents to pursue goals across multiple steps. | Agentic AI refers to systems composed of one or multiple coordinated AI agents that can break down tasks, collaborate, use external tools and pursue goals over extended periods with limited human supervision.6 |
The horizontal axis describes how a result is produced. The vertical axis describes how the next action is selected. These dimensions separate a fixed multistep workflow from an agentic system.
Discriminative and generative describe the output a system produces. Agentic describes how a system works toward a goal across steps.
Uses relationships in data to make predictions or assign information to predefined categories.3
Creates new content, such as text, images, or video.4
Uses one or more coordinated agents to break down tasks, use tools, and pursue goals across steps.6
| Property | Fixed-rule automation | Discriminative AI | Generative AI | Agentic AI |
|---|---|---|---|---|
| Core behavior | Applies a specified rule, formula, mapping, or sequence | Estimates, classifies, ranks, recommends, evaluates, or analyzes | Creates new content | Selects and adapts actions toward a goal |
| Created through | Process mapping, logic design, configuration, and testing | Task definition, data preparation, model training, validation, and threshold selection | Pretrained model, instructions, approved sources, evaluations, and review controls | Goals, models, tools, state, permissions, approval gates, and stop conditions |
| Common failure | Outdated rule, missed exception, broken route | Performance drift, poor calibration, subgroup disparity, workflow misuse | False, incomplete, biased, or unsupported content | Excessive authority, unsafe tool call, compounding error, weak recovery |
| Main monitoring | Rule ownership, exception logs, policy and code-table updates | Current performance, data shift, thresholds, outcomes, subgroups | Draft quality, source use, recurring errors, model or prompt changes | End-to-end behavior, permissions, tool use, approvals, logs, recovery |
| Human role | Designs rules and handles exceptions | Defines use, reviews performance, interprets or acts on output | Verifies content against evidence and professional judgment | Sets authority, monitors behavior, approves consequential actions |
A healthcare application combines methods, models, capabilities, workflow technology, and human decision points. These terms form a stack of distinct layers.
A method family that develops models through exposure to training data. Neural networks and deep learning belong within machine learning.1
A field concerned with machines capable of processing, analyzing, and generating human language.10
In the generative-AI context, a broadly trained model that can be adapted for varied downstream tasks.11
A class of language models that uses deep-learning algorithms and very large textual data sets.10
The workflow feels like one feature to the nurse and patient. Its internal components require different owners, evidence, and controls.
New input enters the portal.
A configured rule sends the message to a clinical queue.
A trained classifier estimates topic and urgency.
An LLM drafts a reply for review.
A person checks the message, chart, and draft before sending.
Many products combine AI with infrastructure and automation that perform reliable, predefined work. Classify the underlying component and its function.
Applies maintained conditions, formulas, mappings, or decision tables.
Repeats configured, rules-based actions across applications. An HHS/FDA privacy assessment documents an RPA service with no AI tools or technologies recorded.1213
Provides a defined access point or exchanges information between systems. FHIR is an HL7 standard for exchanging healthcare information electronically.1514
Converts images of typed, handwritten, or printed text into machine-encoded text.16
Summarizes data, applies a fixed formula, or presents a result. A learned model may supply one of its inputs.
Performs programmed physical actions. AI can add perception, adaptation, or planning.
Both require disciplined design and ongoing ownership. Their maintenance burden appears in different places.
An agent may be technically capable of using a tool while organizational policy limits its authority. Healthcare governance needs both views.
Government of Canada guidance treats permissions, human checkpoints, logs, stop mechanisms, and recovery planning as controls for agentic systems.6
The same screen can contain several components. Inventory each one, identify its consequence, and apply the control that fits its behavior.
Fixed rule, learned prediction, generated content, or human judgment.
A person, a fixed workflow, or an AI agent.
Delay, administrative rework, financial loss, privacy exposure, or patient harm.
| Component | Evidence to request | Operational control | Owner questions |
|---|---|---|---|
| Rule or workflow | Logic, rule source, exception paths, test cases, change history | Named owner, review date, exception queue, regression testing89 | Which policy or code table drives it? Who updates it? |
| RPA or interface | Process map, data mapping, access design, failure logs | Credential controls, connection monitoring, failure alert, reconciliation81415 | Which screen, field, endpoint, or credential creates fragility? |
| Discriminative AI | Intended use, data description, evaluation results, affected groups, thresholds | Performance monitoring, change review, threshold governance, override7 | For which patients, settings, and decisions was it evaluated? |
| Generative AI | Approved sources, evaluation set, known failure patterns, reviewer design | Source review, content checks, human review, recurring-error analysis5 | What evidence should support the draft? Who can correct or reject it? |
| Agentic AI | Goal, tools, identity, permissions, test trajectories, logs, recovery plan | Least privilege, approval gates, stop rules, monitoring, recovery6 | Which actions are permitted? Where must the agent stop for approval? |
| Human review | Reviewer role, timing, information, workload, intervention authority | Clear accountability, source access, override, and correction56 | Can the reviewer see the evidence and change the result before consequence? |
