After the medical copilot comes a choice of more clearly bounded responsibilities, not necessarily a fully autonomous doctor. A useful framework separates retrieving information, making contextual recommendations, preparing drafts, executing authorised actions and coordinating several tasks. Each transition introduces new evidence, permission and recovery requirements.
The five stages below are an editorial framework, not an official regulatory scale, maturity certification or product ranking. They describe what software does in a particular workflow. One product may occupy several stages at once.
The autonomy ladder
| Stage | Role | Additional information or control needed | Appropriate human question |
|---|---|---|---|
| 1 | Retrieve and explain clinical information | Relevant sources and clear limits | Does this evidence support the claim? |
| 2 | Make a contextual recommendation | Sufficient reliable patient and setting information | Does this apply to this situation? |
| 3 | Prepare a clinical or administrative draft | Correct identity, destination and intended action | Is this the version and action I intend to approve? |
| 4 | Execute a bounded authorised action | Enforced permissions and external confirmation | What happened, and what remains unresolved? |
| 5 | Coordinate several tasks | Dependency management, monitoring and safe handback | Can the whole pathway be understood and interrupted appropriately? |
Risk does not rise in a simple straight line. A misleading reference answer can be consequential, while a tightly constrained administrative action may be relatively bounded. The framework separates authority rather than assigning a universal safety score.
Stage one: information retrieval and clinical reference
At this stage, a clinician asks a question and receives information. The system does not acquire authority merely because its explanation is persuasive. Its value depends on source selection, interpretation, relevance and the ability to expose uncertainty.
As described for September 2026, Ask-iatroX occupies a clinical-reference role with linked sources. Its methodology describes grounding and checking processes. Those features do not make reference use risk-free or establish that every answer is correct.
A useful evaluation would examine whether important qualifications survive summarisation and whether unsupported claims are avoided. It should not only ask whether a related citation appears.
Stage two: contextual recommendations
The task changes when the system recommends what should happen for a particular patient. It now needs to know which information is current, which is uncertain and which is missing. Local services and the patient's preferences may also matter.
Consider a fictional follow-up decision. A general guideline is relevant, but the patient has already agreed a different plan after a specialist discussion. A contextual system should not silently override that information merely because the general recommendation is easier to retrieve.
The MHRA intended-purpose guidance, checked on 15 October 2026, is relevant to defining the exact function and users. Calling the product a copilot does not settle its regulatory status or the evidence required.
Stage three: drafting for review
Preparing a referral, record entry or message can reduce manual work while leaving approval with an authorised person. The draft should be clearly distinguishable from a completed action.
The review needs to show the patient or encounter, relevant information, destination and consequences. If the draft changes after approval, the system must know whether the earlier approval still applies. A general instruction to "review carefully" is less useful than an interface that makes the consequential elements inspectable.
A fictional patient message can be medically plausible but inconsistent with the plan actually agreed. The reviewer needs to detect that difference before it is sent, not merely confirm that the sentences are grammatical.
Stage four: bounded execution
Execution means changing something outside the conversation. The software may submit an approved document or perform another permitted task. Its authority should be enforced by the system rather than inferred from its confidence.
The NCSC's prompt-injection analysis, published in December 2025, supports limiting the consequences of untrusted content influencing a model. An external letter should not be able to grant the agent permission to disclose information or use a new tool.
Completion also needs evidence from the receiving system. Submission, acceptance and clinical resolution are different states. The agent should not claim the last merely because the first occurred.
Stage five: coordination across a workflow
At this stage, several tasks depend on one another. The system may need to wait for a response, reconsider a plan after new information or hand back when an assumption no longer holds.
In a fictional pathway, a referral is accepted but the patient later reports that they cannot attend. A coordinator must preserve both the completed referral and the unresolved change in circumstances. It cannot treat the original sequence as permanently correct.
This stage requires a record of dependencies and outstanding work. Another authorised person should be able to take over without guessing which actions have occurred. A conversational summary may help, but it is not a replacement for operational state.
More context does not necessarily mean more autonomy
The OpenAI Epic plugin documentation, checked on 15 October 2026, provides a useful counterexample to a simplistic ladder. It describes organisational access to authorised patient information while remaining read-only.
A system can therefore become more contextually useful without being given write permissions. That may be the appropriate design for its intended use. The next stage is not automatically the correct product roadmap.
Likewise, a narrowly bounded execution tool may have less need for broad clinical reasoning than a sophisticated recommendation system. Buyers should inspect the actual function rather than infer autonomy from model capability.
Evaluate the transition, not just the destination
Before expanding authority, test what changes. Moving from reference to recommendation requires assessment of patient-context handling. Moving from drafting to execution requires permission and completion tests. Moving into coordination requires interruption, dependency and recovery tests.
A staged evaluation should include appropriate refusals and handbacks as successful behaviour. A system that declines to act outside scope has not necessarily failed the task; it may have preserved the intended boundary.
The English NHS clinical-risk standards distinguish manufacture from deployment and use. Local configuration can change the practical risk even when the underlying product remains the same.
What clinicians need to practise
A clinician supervising stage-three drafts needs different operational understanding from someone overseeing stage-five coordination. Both need sufficient clinical knowledge to recognise when the proposed work is inappropriate.
A proposed educational exercise can ask the learner to identify the current stage, the missing information and the next approval boundary. iatroX's September 2026 Tutor and simulation formats can support reasoning practice, without constituting product-specific certification or proof that deskilling has been prevented.
The useful endpoint is not maximum autonomy. It is a clear allocation of work in which information, permission and responsibility remain aligned.
Frequently asked questions
Is this five-stage ladder an official regulatory framework?
No: it is an editorial tool for describing functions and authority. Regulatory assessment depends on the intended purpose and applicable jurisdiction.
Is stage five always better than stage one?
No: the appropriate level depends on the task and evidence. Greater autonomy can add unnecessary complexity or risk.
Can a product be at several stages simultaneously?
Yes: it may retrieve information in one function, prepare drafts in another and execute a separate bounded task. Each function needs its own assessment.