Key Takeaways
- Faster AI drafts don't shorten lead times if review stays the same
- The bottleneck isn't drafting—it's cross-department approval flow
- AI-generated documents still need a human record of review and approval
- This week: Measure the approval turnaround for one type of AI-drafted document
59% of employees at domestic finished drug manufacturers use AI in their work, but because AI is used merely as an individual assistance tool within siloed departmental processes, company-wide productivity gains remain minimal. Even if a staff member uses AI to draft the equivalent of 400 minutes' worth of documents in a day, the final completion time (lead time) won't shorten unless the review time taken by team leaders or the Quality Assurance (QA) department also decreases.
On November 17, 2026, ITCEN Cloit hosted the 'Pharma & Bio AX Transformation Seminar' at the 4th floor auditorium of the Korea Pharmaceutical and Bio-Pharma Manufacturers Association, under the theme 'Beyond Demos, Toward Real Results.' Roughly 100 industry professionals attended. The seminar consisted of three sessions.
It's Not About Drafting Speed—What Actually Needs to Change
In the first session, Baek Mi-seon, CEO of Beagle, pointed out that most pharmaceutical company work involves creating and handing off vast amounts of 'documents' and 'records' spanning R&D, commercialization, and post-market management. Baek argued that "speeding up drafting alone can never resolve departmental bottlenecks," and proposed a '4D Loop (Discovery-Decision-Design-Drive)' strategy: rather than running one-off AI pilots, companies should select one or two high-value core tasks (lighthouse projects) and completely redesign the entire workflow from initial drafting to final approval.
A representative example cited was the 'SOP (Standard Operating Procedure) revision' workflow, common to all pharmaceutical companies. This is a multi-agent approach in which the following agents are linked together to carry out tasks in sequence.
- An agent that detects changes in external regulatory notices
- An agent that searches for internally affected SOPs
- An agent that drafts revisions based on approved documents
- An agent that checks for inconsistencies between regulations
However, Baek limited AI's role to finding, comparing, proposing drafts, and flagging omissions. She drew a clear line that final quality-impact judgments and accountable approvals must be carried out directly by an authorized human (Human-in-the-Loop).
AI-Written GMP Documents—Who Is Responsible for Approval?
In the second session, Lee Eui-gyu, a partner attorney at Bae, Kim & Lee LLC, diagnosed that the root cause of hesitation around AI adoption in the field is not a lack of technical capability, but "anxiety stemming from the absence of control mechanisms." In an April 2026 U.S. FDA warning letter case Lee presented, a pharmaceutical company used an AI agent to draft Good Manufacturing Practice (GMP)-related documents such as specifications and procedures, but the Quality Unit (QU) failed to adequately review related matters such as process validation. This resulted in a citation for violating 21 CFR 211.22(c), ultimately forcing the company to halt drug production and distribution.
Lee warned that "during an FDA inspection, the excuse that 'I didn't know the regulation because the AI didn't tell me' simply doesn't hold up," and that AI-generated documents are treated the same as any other company document—meaning companies must keep records showing that a human cross-checked the content against source regulations and personally reviewed and approved it. The six essential records that must be retained to support company-wide AI adoption are as follows.
- The AI tool used and its version
- The input prompts and reference materials used
- The original AI-generated output
- A record of human edits made
- The reviewer/approver and the criteria applied
- The reason for modifying or not using the AI output
Lee also emphasized the need to build a four-stage control system—starting with the responsible business-unit lead, followed by cross-departmental review, executive approval, and board reporting—similar to systems used by global pharmaceutical companies such as Eli Lilly and Sanofi.
Regulatory risks for marketing and sales departments were also addressed. Under Article 68, Paragraph 7 of the Pharmaceutical Affairs Act, effective November 27, 2026, advertisements that use AI-generated images or voices of fictitious doctors, pharmacists, professors, or other experts to recommend or endorse pharmaceutical products will be completely banned.
In the third session, Park So-young, Director at host company ITCEN Cloit, introduced the company's enterprise-wide AI agent integration management platform, 'AgentGo.'
Implications for AX at Korean Companies
If you're already using AI but not seeing results, it's more accurate to track the time it takes for a single document to go from drafting to approval as your performance metric, rather than simply tracking AI usage rates. This is because shortening only the drafting stage while leaving the review and approval stages unchanged will not reduce lead time.
If you're using AI to create regulated documents or documents requiring approval accountability, the right first step is to check which of the six records listed above you are actually keeping today. In the FDA warning letter case, the violation stemmed from the Quality Unit's (QU) failure to properly review relevant matters.
When selecting where to apply AI, rather than running multiple one-off pilots, companies might consider Baek's proposed approach: designating one or two core tasks and designing the entire flow from drafting to approval as a single process. By dividing which steps AI will handle and which steps require human approval at the design stage, the task of maintaining approval records can be built into that same workflow.
There's one thing worth checking this week. Choose one type of document currently being AI-drafted, and measure how many days it takes from draft completion to final approval—and which department holds onto it the longest.
