Skip to main content
Not all writing tasks carry the same risk. Summarising a published paper for an internal briefing and drafting a promotional claim for a detail aid are both “writing tasks” — but the consequences of an undetected error are fundamentally different. This playbook uses four risk tiers to determine how much AI contribution is appropriate, what verification is required, and who is accountable for sign-off.

​
The four tiers at a glance

Risk tierAI roleReview processReviewer
LowFirst draft, structuring, summarisationStandard medical writing reviewExperienced medical writer
MediumTransformation, adaptation, extractionEnhanced review with source cross-checkSenior medical writer or SME
HighLimited drafting support onlyExpert review, full verification, formal sign-offMedical advisor, regulatory reviewer, or compliance lead
CriticalSupporting role onlyFull expert review — this is the final quality gateQualified reviewer for the content type

​
Low risk

AI produces a working first draft. Standard medical writing review applies. Low-risk tasks involve structuring, summarising, or reorganising content from clear source materials. The output is a starting point — you review, correct, and refine before it becomes a deliverable. Typical deliverables:
  • Structured summary of a published paper for an internal briefing or literature review
  • First-draft content outline from a briefing document and key messages
  • Reformatted version of existing approved content (e.g., slide deck notes into a written summary)
  • Internal meeting notes or therapeutic area backgrounder from source materials
Review requirement: Standard medical writing review. Check data points against sources, verify the structure matches the deliverable purpose, and refine language and emphasis. What can go wrong: A summary transposes a primary and secondary endpoint result. An outline prioritises the wrong sections for the target audience. A reformatted piece drops a safety finding from the original. These errors are correctable in standard review — but they must be caught.

​
Medium risk

AI is useful, but the transformation introduces drift risk. Enhanced review with source cross-checking applies. Medium-risk tasks involve reframing, adapting, or extracting interpretive content from source materials. The core risk is that the output subtly changes the meaning, emphasis, or scope of the original evidence. Typical deliverables:
  • Key messages extracted from a Phase III paper for a messaging framework
  • HCP slide deck content adapted into a GP-facing leave piece
  • Congress poster summarised into a client-facing highlights report
  • Technical manuscript summary adapted for a payer audience
Review requirement: Enhanced review. Cross-check every claim in the adapted output against the source. Specifically verify that the transformation has not shifted emphasis, dropped qualifiers, merged study populations, or strengthened language beyond what the evidence supports. What can go wrong: A key message frames a non-inferiority result as if it demonstrated superiority. An audience adaptation drops “in patients with moderate-to-severe disease” and the claim reads as applying to all patients. A repurposed summary introduces a comparative framing not present in the original. These errors are subtle, read fluently, and are easy to miss in a surface-level review.

​
High risk

AI provides limited drafting support only. Expert review is mandatory. No AI output enters the deliverable without line-by-line verification. High-risk tasks involve regulatory-sensitive content, patient-facing materials, or claims with compliance, safety, or legal consequences. The cost of an undetected error is not rework — it is regulatory action, patient harm, or reputational damage. Typical deliverables:
  • Promotional claims or value messages for a detail aid, leave piece, or HCP website
  • Patient-facing explanations of treatment benefits and risks
  • Plain language summaries of clinical trial results for public disclosure (e.g., EU CTR lay summaries)
  • Content subject to ABPI, IFPMA, or FDA promotional review
  • Safety narratives or adverse event profile summaries
Review requirement: Expert review by a senior medical writer, medical advisor, or regulatory/compliance reviewer. AI output is raw input only — it does not reduce the scope, rigour, or timeline of the review process. The reviewer is not checking whether the AI did a good job; they are verifying the content as if they had written it themselves. What can go wrong: An AI-drafted promotional claim implies superiority based on a non-comparative study. A plain language summary understates a common adverse event and the document is published as part of clinical trial disclosure. A detail aid includes a claim supported by a subgroup analysis without labelling it as such, and MLR rejects the submission. These are not hypothetical — they are the failure modes this tier is designed to prevent.
At high risk, AI involvement does not shorten the review process. If you are allocating less time for review because AI produced the first draft, you are under-reviewing.

​
How to apply the framework

1

Identify the task's risk tier

Before starting any AI-assisted workflow, determine the tier based on:
  • What type of content is being produced?
  • Who is the intended audience?
  • Is the content regulated or promotional?
  • What are the consequences if the AI output contains an error?
2

Match the review process to the tier

Use the table at the top of this page. When in doubt, default to the higher tier — it is always safer to over-review than under-review.
3

Document the tier in your project files

Record which workflow was used, what risk tier applies, who reviewed the AI-assisted output, and what changes were made during review. This supports transparency, audit trails, and continuous improvement.

​
Handling tier boundaries

Some tasks sit between tiers. Use these rules:
  • Default to the higher tier. Over-reviewing is always safer than under-reviewing.
  • Consider the downstream use. A summary used only internally is lower risk than the same summary adapted for a public-facing document.
  • Consider the regulatory context. Content for regulated markets, promotional use, or patient communication should always be treated as high risk regardless of the task type.

​
Tools aligned to higher-risk workflows

RefCheckr

Supports claim verification at high-risk tier.

MedCheckr

Supports promotional compliance review at high-risk tier.

Patiently AI

Supports patient-facing content creation at medium-to-high risk.

PLS Generator

Supports plain language summary creation at medium-to-high risk.
These tools assist the review process — they do not replace expert judgement. A tool flagging no issues does not mean there are no issues. When a tool flags an issue, a human must assess whether it is a real problem.