AI Adoption Should Be Scoped by Workflow, Not Feature — A Pharma AX Approach That Splits Human and AI Roles into Three Layers

Part 5 of the 'Pharma AI Transformation' series, published on August 25, pointed out that the basic unit of AI adoption should not be an individual feature but the entire 'work workflow.' Simply selecting a task to apply AI to does not mean transformation happens just by bolting one AI feature onto an existing process.

Dividing human and AI roles into three layers

The framework for redesign presented in the series rests on separating human and AI roles into three levels.

• AI-led — Document search, data comparison, repetitive processing, draft report generation. Areas where AI takes full ownership to reduce lead time.
• Human+AI (collaboration) — In-depth root cause analysis, review of alternatives, impact assessment following regulatory changes, exception handling. Areas where human intuition and AI reasoning interact.
• Human-led — Quality judgment, regulatory compliance review, accountability verification, final approval procedures. Areas that must remain under human control and leadership.

The article noted that this principle is commonly found across the strategies of McKinsey, Boston Consulting Group (BCG), Microsoft, the U.S. Food and Drug Administration (FDA), and the European Medicines Agency (EMA). These organizations design not only 'what to delegate to AI' but also 'where humans must intervene (human-in-the-loop)' from the outset. Microsoft recommends requiring human approval for actions that are difficult to reverse or that carry regulatory implications, while the FDA and EMA emphasize human-centered design and evaluation of 'Human-AI Interaction' as core principles in their '2026 Good AI Practice' guidelines.

Starting from the business problem, not the technology

Bagel, a company specializing in enterprise AI design, stressed that this process should be approached from a 'work redesign' perspective. A Bagel representative said, "Most companies make the mistake of taking a technology-first approach when adopting AI, starting with questions like 'which cutting-edge technology should we apply to which task.' We instead pursue the reverse approach: first examining the business problem on the ground, redesigning the work, and only then applying the technology." The sequence is to first redefine pain points such as delays, omissions, and errors in actual operations, then divide roles between AI and humans, and finally match the most suitable technology.

The execution platform, and the data prerequisite

As an execution system to run the designed workflow in the field, the article introduced ITCEN Cloit's multi-agent platform 'AgentGo.' The article explained that the platform embeds a human-in-the-loop structure in which AI accesses systems governed by consistent standards to collect data and produce drafts, while practitioners with final approval authority review and approve at key checkpoints. An ITCEN Cloit representative said, "The essence of AI transformation that leading global pharmaceutical companies are focusing on is not the adoption of tools, but systematizing the way frontline practitioners safely collaborate with AI."

However, the article pointed out that certain prerequisites must first be resolved for an agent platform to function properly — namely, the question of which internal data the agent can trust when drawing conclusions. The article noted that pharmaceutical company data is often scattered across multiple systems such as ERP, MES, LIMS, and EDMS, and that even the same key performance indicator (KPI) is frequently defined and calculated differently across departments. On top of this, guidelines from the EMA, Korea's Ministry of Food and Drug Safety (MFDS), and the FDA require that the source of data, its processing, and analytical decisions be transparently traceable and verifiable. This is an industry where it is not enough for the result to be correct — the evidence and process must also be traceable.

As a solution to this data reliability problem, the article cited IDK2's (Heartcount's) 'Decision Agent.' It is a hybrid structure that separates roles: complex numerical computation and statistical validation are handled by an analytics engine, while the conversational AI (LLM) only adds business context to verified results and explains them in natural language. A 'semantic layer' standardizes metric definitions and business rules so that practitioners receive consistent answers even when asking questions in everyday language, and sensitive data is processed within the user's web browser rather than being sent to external servers, making it usable even in closed networks or on-premise environments.

The flow summarized in the article follows this sequence.

1. Identifying the business problem on the ground
2. Work redesign that redraws how humans and AI work together
3. Designing Human+AI roles
4. Operating the agent execution platform
5. Building a trustworthy data pipeline

The series concludes that only once this sequence is in place can frontline staff truly focus on final human decision-making.

Implications for AX at Korean Companies

Although this series covered the pharmaceutical industry, the three-layer classification can be applied as-is even by companies in other sectors considering AI adoption. Once a candidate task for AI application has been selected, the first step should be to divide its stages into 'tasks fully handled by AI,' 'tasks where AI drafts and humans judge,' and 'tasks only humans do.' As Microsoft recommends, marking which stage involves actions that are hard to reverse or that carry regulatory implications reveals where approval checkpoints should be placed.

Bagel's 'reverse-order' approach follows the same logic. If a company has already decided which technology to adopt and is now searching for a task to apply it to, the right sequence is instead to first write down exactly where delays, omissions, and errors actually occur. Choosing technology without first pinpointing the bottleneck leads to what the Bagel representative described as 'forcing the work to fit the technology.'

The data-related prerequisite is even more concrete. If the metrics used in a task that will be handed over to an agent's judgment are calculated with different definitions across departments, unifying those metric definitions must come before adopting the platform. For tasks subject to audits or regulatory response, it is necessary to verify not just the output but whether the data used and the process by which it was handled to reach that conclusion can be traced. If a closed network or on-premise environment is a requirement, whether the structure can process data without sending it externally becomes a selection criterion.

There is one thing to check this week. Pick one task that is a candidate for AI application and write down which of the three layers each of its stages belongs to. If there is a stage in that table where 'who gives final approval' is left blank, the redesign of that task is not yet complete.

Source: [Pharma AI Transformation ⑤] AI Adoption Is 'Work Redesign,' Not a 'Feature' — Dividing Roles Between Humans and AI ...