My 2026 Workflow for AI Supported Medical Information Standard Response Letters

By Isaac Khan, PharmD · September 22, 2026

What has changed is Google itself. When it provides an AI-generated summary, I use it as a secondary check on the questions and interpretations my audience may encounter. It can suggest leads to investigate, but I verify those leads against the original sources. An AI summary does not validate the completeness or accuracy of a literature search.

Gather and assess your own evidence first. Then use search summaries to spot questions or sources worth a closer look.

You need to get your AI system working within the right context from the start. This is where you set the stage.

Before I draft anything, I also define the audience and the boundaries of the task. In my case, that may mean telling the AI that it is supporting a U.S. oncology pharmaceutical company’s Medical Information department and writing for practicing oncologists, nuclear medicine physicians, and other healthcare professionals.

I then tell it to base the draft only on the materials I provide: the reference letter document, the MI lexicon, the scientific communications platform, and the study, article, poster, abstract, or structured notes in the chat. That source boundary matters. It is intended to reduce unsupported additions and keep the draft anchored to the evidence under review. It still requires source verification.

Include:

• Formatting requirements
• Dos and don’ts
• Templates
• Style guides
• Example letters
• Related guidance

I have been touting the 3 Cs: clear, concise, and compliant. This is where I tell the AI to apply those principles and my organization’s requirements. A prompt can support that goal; qualified review determines whether the response meets it.

For each HCP request, I specify both the request type and the key question. The request type might be primary or secondary efficacy, general safety, a specific adverse event, pharmacokinetics/pharmacodynamics, concomitant use, subgroup data, or a comparison. This keeps the AI focused on the endpoints and findings that actually answer the question instead of merely summarizing the entire paper.

My priority order is simple:

1. Answer the key question in a scientifically balanced way.
2. Support regulatory and company requirements through non-promotional, scientifically balanced language.
3. Keep the response clear and concise for a busy HCP.

Keep the priming message focused, and attach supporting documents such as example letters and style guides as needed and permitted. I prefer a small, relevant set of documents rather than everything available. Attachment limits vary by account and tool. For company work, use an organization-approved environment and follow its rules for confidential information, HCP data, and licensed source material.

Tell the AI to review a single study that you attach or upload and extract the key components you may—or may not—ultimately want to include in your response.

Specifically, ask it to identify:

• One- to two-sentence, answer-first key takeaway
• Study objective as it relates to the question
• Study design, including phase, randomization, control, setting, and prospective or retrospective design
• Population, including tumor type, line of therapy, prior treatment, sample size, and relevant inclusion and exclusion criteria
• Treatment regimen, including dose, schedule, route, and duration
• Endpoints most relevant to the request
• Numeric results that answer the question, including point estimates, confidence intervals, P values, adverse-event rates, or pharmacokinetic parameters as applicable
• Whether findings are primary, secondary, or exploratory
• Other findings only when they meaningfully help interpret the answer
• Limitations that affect confidence in the answer, such as a single-arm design, small sample, post hoc analysis, missing data, short follow-up, or unbalanced arms
• Study conclusions

These are the core building blocks of a response summary. Do not assume the AI will identify them correctly. Tell it what to extract, ask for page or section references, and check the results against the study.

This step is not mandatory, but if you do not already know which graphic elements to include, ask the AI for suggestions.

Where an approved AI tool supports document editing, it may also help assemble and format graphics. Whether the AI creates the visual or only recommends it, ask for a neutral title, the exact rows and columns or axes, the data and units to display, and essential footnotes covering analysis populations, endpoint definitions, follow-up, censoring, study design, and whether an analysis was exploratory. Check every plotted value, label, and footnote against the source.

Also ask for compliance considerations specific to the visual. A graphic should not imply superiority, causation, or a comparison the study was not designed to support. Single-arm, noncomparative, off-label, and exploratory data should be labeled clearly, and abbreviations should remain consistent with the surrounding letter text.

Before moving on: Get these elements into the letter.

For a focused response, I often aim for approximately:

• Three core references; or
• Five pages.

These are drafting targets, not evidence limits. Include additional references when needed to answer the question accurately and represent the evidence fairly. A closing sentence can point readers to additional relevant studies, but it should not replace discussion of findings that materially affect the answer.

Ask the AI to develop summary bullets for the letter and propose a logical flow of headers and subheaders that breaks up the text and makes the response easy to read and navigate.

I ask for one neutral, descriptive header, followed by one or two concise, answer-first overview bullets and then the supporting paragraphs. The header should identify the study setting or population without promotional language, and the bullets should be strong enough that a busy HCP can read only those lines and understand the main answer.

Abbreviation order matters here. I do not define abbreviations in the header. I define them once at first mention, even in the overview bullets, and then use the abbreviation consistently throughout the remaining bullets and paragraphs.

I like to prompt it to make the bullets high-level talking points, with only the key data points (eg numbers) in parentheses as reference points.

This makes it easier for medical science liaisons (MSLs) or fellow MI colleagues working at congress booths or speaking by phone to discuss the data. Give the AI clear instructions and, again, refine the prompt as you go.

At this point, I aim to have a draft that is ready for substantive editing. How close it is to finished depends on the question, the source material, and the quality of the AI output.

• Read it closely.
• Clean it up.
• Make it yours.
• Make sure it makes sense.

Isaac Khan, PharmD, is a Medical Information and Medical Affairs operations leader with more than 20 years of experience across pharmaceutical and biotech organizations. His work includes building global Medical Information capabilities and applying AI-supported tools to scientific content and medical insights. His experience spans oncology, radiopharmaceuticals, and rare disease. He writes about practical AI applications for Medical Information at MedInfoAI.com.

Explore Isaac’s Medical Information and Medical Affairs leadership experience.

Connect with Isaac about AI-supported Medical Information, scientific communication, insights, and building the function that supports them.

MedInfoAI · Isaac Khan, PharmD