Medical Information in the Age of AI: Where It Helps and Where a Human Must Stay in the Loop
AI can make medical information faster, more consistent and more responsive across languages. The teams that benefit most are the ones that decide, stage by stage, where AI does the work and where a medically qualified person makes the call.
Suhina Singh
Founder & CEO, Jonda Health
October 11, 2026
·
7 min read
The enquiry journey at a glance
Seven stages of a medical information enquiry, shaded by where AI can take on most of the work and where a qualified person must decide.
Medical information teams answer questions that matter: a clinician checking dosing in renal impairment, a pharmacist asking about stability outside the fridge, a patient worried about a side effect. The answers must be accurate, balanced, consistent with the local label and delivered quickly, often across several countries and languages. Enquiry volumes grow, content libraries expand, and the expectation of fast, personalised responses keeps rising.
AI can help with much of this. It can read and route enquiries, search large libraries of approved content, draft responses and surface trends. But medical information is also where a confident mistake can reach a clinician or patient directly, and where a missed adverse event has regulatory consequences. The useful question is not whether to use AI, but where it should do the work and where a qualified person must stay in control. The seven stages below set that out.
Stages 1 to 7
Where AI helps, and where people decide
1. Intake
Enquiries arrive by phone, email, web forms, field referrals and chat, in different languages and levels of detail. Capturing them accurately is the foundation for everything that follows.
Where AI helps
Transcribing calls, translating, extracting the product, question, requester type and country into structured fields, and spotting duplicates.
Where people stay in the loop
Confirming the real question when an enquiry is ambiguous, and following up with the requester when key details are missing.
2. Triage
Every enquiry has to be checked for a possible adverse event, a product quality complaint, off-label content and urgency. This is the stage where missing something has the most serious consequences.
Where AI helps
Flagging possible adverse events and complaints, including ones described indirectly or in another language, and prioritising urgent enquiries. Detection should be tuned to over-flag rather than under-flag.
Where people stay in the loop
Every flagged or possible safety case is reviewed by a trained person and handled through pharmacovigilance procedures within the required timelines. AI supports detection; it does not decide what is reportable.
3. Search and retrieval
Answers draw on standard response documents, local prescribing information, published literature and internal data. Finding the right, current source quickly is a large part of the work.
Where AI helps
Searching approved content and literature by meaning rather than keywords, and bringing back the most relevant passages with their sources.
Where people stay in the loop
Making sure the sources are approved, current and correct for the country, including the local label, and judging when the evidence is insufficient to answer.
4. Drafting
This is where generative AI can save the most time, and where the design of the system matters most. A model that drafts only from your approved content behaves very differently from one that draws on anything it has seen.
Where AI helps
Drafting a response from retrieved approved content, tailoring it to the requester and the question, and citing each statement to its source.
Where people stay in the loop
Checking that nothing has been invented, that references exist and say what is claimed, and that the draft answers the question asked without drifting into promotion.
Grounding reduces risk. It does not remove the need for review.
5. Review and approval
Before a response reaches a clinician or patient, someone accountable must be satisfied that it is accurate, balanced, consistent with the local label and appropriate for the requester.
Where AI helps
Highlighting statements without a source, differences from approved wording, and missing required content, so reviewers can focus their attention.
Where people stay in the loop
Approval stays with a medically qualified person. The reviewer is accountable for what is sent, not the tool that drafted it.
6. Delivery
How a response is delivered depends on who is asking. A healthcare professional asking a scientific question, and a patient asking about their own treatment, need different handling.
Where AI helps
Formatting and translating approved responses, and, where carefully designed and approved, answering routine questions with pre-approved content.
Where people stay in the loop
Deciding which questions can ever be answered automatically, making sure patients asking for personal medical advice are directed to their healthcare professional, and offering a clear route to a person.
7. Insight
Medical information enquiries are one of the richest signals a medical affairs function has about what clinicians and patients do not yet understand.
Where AI helps
Grouping enquiries into themes across countries and languages, spotting rising topics, and linking them to field insights and other sources.
Where people stay in the loop
Interpreting what a trend means, deciding what to do about it, and checking that summaries have not lost clinical nuance.
Before you deploy
Eight questions to answer first
These questions apply whether you are building your own tools, adding AI to an existing medical information platform, or evaluating a vendor.
1. What will the AI draw on, and who keeps that content current?
A system is only as reliable as its content library. Decide which approved sources it can use, how local labels and updated responses reach it, and what happens when a document is withdrawn.
2. How are possible adverse events detected, and what happens next?
Test detection on real, varied examples, including indirect descriptions and other languages, and confirm that every flag reaches a trained person and the pharmacovigilance process in time.
3. Who approves each type of response?
Define which responses always need individual medical review, and whether any pre-approved content can be sent without it. Write this down before launch, not after the first incident.
4. How will you measure accuracy?
Agree how you will test the system before launch and monitor it afterwards: accuracy of extraction and triage, correctness of drafts, citation accuracy, and how often reviewers make substantive changes.
5. Where is enquiry data processed, and is it used to train anything?
Enquiries contain personal and health information about requesters and patients. Know where data is stored and processed, whether it leaves the country, and whether any vendor uses it to train models.
6. Can you show how each response was produced?
For audits and inspections, you should be able to show the enquiry, the sources retrieved, the AI draft, the reviewer's changes and the final response.
7. How does it perform across your languages and markets?
Performance in English does not guarantee performance in Thai, Bahasa Indonesia or Vietnamese. Test in the languages and with the local labels your teams actually use.
8. What changes for your team, and how will you support them?
AI changes the medical information role from writing to reviewing and judging. Plan training, workload and accountability so the team trusts the tools without over-relying on them.
Readiness at a glance
Area
Ready
Not yet
Content
Approved, current library with local labels
AI drafts from general knowledge
Safety
Tested detection; every flag reviewed in time
Detection assumed, not tested
Approval
Clear rules on who approves what
Drafts sent without defined review
Accuracy
Tested before launch, monitored after
No measure of accuracy
Data
Processing location and training use known
Vendor terms unreviewed
Audit trail
Sources, drafts and edits recorded
Only final responses kept
Languages
Tested in each language and market
English testing only
People
Roles, training and workload planned
Tools launched without team change
How Jonda Health can help
Using AI safely in healthcare is a question of design, data and clinical judgement together, and that combination is how our team works. You do not need to have every answer before talking to us. There are two ways we can support you.
Jonda Health Services
Design AI into medical information safely
Through our Digital Health Strategy and Automation and Digital Enablement work, we help medical information and medical affairs teams decide where AI fits in the enquiry journey, define where people stay in the loop, prepare content and data, and set up testing and monitoring so tools hold up beyond the pilot.
JondaX
Structure the clinical documents that arrive
Enquiries and safety cases often arrive with lab reports, discharge summaries and other clinical documents in PDFs, scans and images, in more than 10 languages. JondaX turns them into structured, standardised information through a single API, checked by clinically trained reviewers to 99% field-level accuracy. It is ISO 27001 certified and designed to comply with HIPAA, GDPR and PDPA.
Book a free 30-minute conversation
Bring the part of your medical information process you are considering for AI, and we will talk through where it can help, where people need to stay in the loop, and what to prepare first.
