Using AI in medical writing used to mean giving a large language model a prompt and reviewing its response.

AI agents are changing that interaction.

Instead of simply generating text, an AI agent can work through multiple steps, such as retrieving information, analyzing source documents, drafting content, cross-referencing data and checking outputs.

For medical writers, this raises a different question: How do you communicate effectively with an AI agent when working on clinical and regulatory content?

The answer goes beyond writing better prompts. Working with AI agents increasingly means defining the objective, providing the right context and sources, establishing constraints and deciding when human judgment is required.

From prompting to supervising AI agents

Consider a simple instruction:

“Write the safety section of this clinical study report.”

An LLM can generate a response. But important questions remain.

Which data should it use? Which documents are authoritative? What terminology should it follow? What happens if two sources conflict? Can it infer missing information? Which statements need to be verified?

These questions are fundamental in regulatory medical writing.

An agentic workflow can potentially retrieve relevant information, analyze sources, execute intermediate tasks, generate content and perform defined checks before returning the result to the medical writer.

The interaction therefore evolves:

Prompting → Instructing → Collaborating → Supervising

With prompting, the medical writer requests an output. With instruction, the writer also defines how the task should be performed. Collaboration introduces an iterative exchange between writer and agent. Supervision goes further: the medical writer establishes what the agent can do, what it must verify and when it should escalate an issue.

The medical writer does not disappear from the workflow. Their role shifts toward directing and reviewing it.

Why context matters more than the perfect prompt

This shift is closely related to an emerging concept in AI: context engineering.

Recent research suggests that this shift is already taking place in healthcare. In 2026, researchers writing in Nature Medicine described healthcare professionals as potential “context engineers”, responsible for shaping the data, task, tools and rules surrounding AI interactions.

Research published in BMJ Digital Health & AI similarly argues that effective medical AI depends not simply on better prompts, but on systematically determining what information an AI system receives and how that context is managed.

Prompt engineering focuses primarily on how an instruction is written. Context engineering focuses on what information, rules and resources the AI has available when completing the task.

Recent research in medical AI describes context engineering as systematically determining what information should be retrieved, how it should be structured and how it should be managed throughout an AI interaction.

For medical writers, that context might include:

• Sources: protocols, SAPs, TLFs, datasets and approved reference documents.
• Task: drafting, summarizing, comparing, cross-referencing or validating.
• Document: CSR, patient narrative, investigator brochure or another regulatory document.
• Rules: templates, terminology, style requirements and workflow constraints.
• Review: what the agent can complete and what requires human judgment.

More information is not necessarily better. Relevant, structured and controlled context matters more than simply giving an AI system as much information as possible.

How should medical writers communicate with AI agents?

A practical framework can make human-agent interaction more reliable.

1. Define the objective

Tell the agent exactly what it needs to accomplish.

For example:

“Review this patient narrative for consistency with the approved source data and flag discrepancies for medical writer review.”

This establishes both the task and the expected outcome.

2. Define the sources

Specify which information the agent is allowed to use.

Depending on the workflow, these could include approved protocols and amendments, SAPs, validated datasets, TLFs, templates or other controlled documents.

If two authoritative sources conflict, the agent should not silently decide which is correct.

3. Set rules and constraints

Defining what the agent must not do can be as important as defining the task.

For example:

• Do not infer missing clinical information.
• Do not introduce information that cannot be traced to an approved source.
• Do not change predefined terminology.
• Flag conflicting information rather than resolving it independently.

These boundaries help turn an open-ended generation task into a controlled workflow.

4. Define the expected output

Specify the structure, terminology, level of detail and source references expected in the result.

For a CSR section, this might include the required structure and underlying TLFs. For a patient narrative, it could include chronology, terminology and required clinical events.

5. Define verification and human review

An AI agent should also know what it needs to check and when it should stop.

This could include verifying numerical consistency, source alignment, cross-document consistency or unsupported statements.

If critical information is missing or sources conflict, the correct action may be to escalate the issue to the medical writer rather than produce an answer.

What does this look like in regulatory writing?

Consider a patient narrative.

Instead of simply asking an AI agent to “write a patient narrative,” the medical writer can define the patient and event, authoritative clinical data, required structure, terminology, rules for missing information and conditions requiring human review.

For a clinical study report, an agent could retrieve relevant information from TLFs, support drafting and cross-reference statements against source material.

For content validation, the objective changes completely. Rather than generating text, the agent could compare regulatory content with approved sources and identify inconsistencies or unsupported statements.

In each case, effective communication is less about finding a perfect prompt and more about establishing the environment in which the agent operates.

AI agents still require human oversight

This distinction is particularly important in regulated Life Sciences environments.

Regulatory content needs to be accurate, consistent and traceable to its underlying evidence. An AI agent does not automatically make a process compliant or validated.

Appropriate controls still need to be defined according to the system, intended use and associated risk.

Platforms such as Narrativa Navigator apply agentic AI to controlled regulatory workflows, supporting activities such as information retrieval, regulatory content generation, cross-referencing and content validation while maintaining source traceability and human oversight.

The medical writer remains responsible for reviewing outputs and applying scientific and regulatory judgment.

Medical writers do not need to become prompt engineers

Working with AI agents does not mean medical writers need to become AI engineers.

Many of the skills required to supervise an AI agent are already central to medical writing: identifying authoritative evidence, defining requirements, detecting inconsistencies, understanding what can and cannot be concluded from the data and reviewing the final content.

What changes is how those skills are applied.

As AI in medical writing moves from isolated generation toward agentic workflows, the critical skill may be less about writing the perfect prompt and more about communicating intent, context, evidence and boundaries.

And in regulatory writing, knowing when an AI agent should not produce an answer may be just as important as knowing what to ask.

FAQ

How are AI agents used in medical writing?

AI agents can support tasks such as information retrieval, regulatory content generation, summarization, cross-referencing, consistency checking and content validation. Agentic workflows can connect several of these tasks before returning an output for human review.

How should medical writers communicate with AI agents?

Medical writers should clearly define the objective, authoritative sources, context, constraints, expected output and verification requirements. They should also establish when the agent must escalate an issue for human review.

Do medical writers need prompt engineering skills?

Clear instructions remain important, but working with AI agents goes beyond prompt engineering. Medical writers also need to define relevant context, select authoritative sources, establish boundaries and evaluate the resulting content.