1 AI-first workflow
AI-first workflows are practical building blocks for moving an organization toward an AI-native operating model.
Clinical trial development is already organized as a chain of governed workflows. The study protocol informs trial operations and the statistical analysis plan (SAP). Data collected in an electronic data capture system moves through SDTM and ADaM into tables, listings, and figures (TLFs). Those results support the clinical study report and regulatory submission. Every step has inputs, rules, handoffs, reviews, and accountable owners.
AI can do more than accelerate one task. An agent can read across artifacts, use tools, execute multiple steps, observe results, and return with evidences or recommendations. Using that capability reliably requires a workflow designed for human-agent collaboration.
In this book, an AI-first workflow is a repeatable way of working in which an AI agent performs a bounded first pass for human-agent collaboration, and people retain responsibility for judgment and approval.
1.1 From AI adoption to AI-native
AI adoption begins when individuals use AI for isolated tasks inside an existing workflow. A person decides when to invoke AI, assembles the context, reviews the response, and moves the work forward. This may improve personal productivity, but the workflow itself remains unchanged.
AI-first is the practical focus of this book. It does not mean AI performs every task or makes final decisions. It means the team considers AI when designing the workflow, then assigns each step to the component best suited to it.
AI-native is the broader operating model: AI is a normal collaborator, documentation is designed for people and agents, and team members increasingly orchestrate work rather than execute every step manually. AI-first workflows help an organization move in that direction, alongside the necessary governance, infrastructure, skills, and culture.
| Stage | How work is organized | Clinical example |
|---|---|---|
| AI adoption | AI is added to an existing task | Ask a chatbot to summarize meeting notes |
| AI-first | The workflow is designed for AI to make a structured first pass | Trigger an agent to compare a changed TLF with its SAP requirements |
| AI-native | Human-agent collaboration shapes the operating model | Reusable agents and skills work across data, analysis, reporting, documentation, and QC |
1.2 How an AI-first workflow works
An AI-first workflow does not begin and end with a prompt. It connects AI to a repeatable process:
Event -> Task -> Tools and agent -> Evidence -> Human decision -> Recorded state
An event starts the workflow, such as a change to an SAP, dataset, program, or output. The task definition states the goal, source artifacts, allowed tools, expected result, conditions for stopping, and responsible reviewer.
The workflow gives the agent current, versioned artifacts instead of relying on whatever happens to be available in a chat. The agent returns a structured finding that states what it checked, the supporting evidence, any uncertainty, and what it could not cover. A human reviews the finding, makes the decision, and records that decision for the next run.
The design separates three kinds of work:
| Work | Best component |
|---|---|
| Apply an exact rule, calculation, or comparison | Deterministic software |
| Interpret written requirements, trace context, or plan variable steps | AI agent |
| Exercise statistical judgment, approve policy, or accept accountability | Human role |
The simplest capable component should do each job. An AI model should not replace a test whose result must be identical every time, and tool access should never be confused with approval authority.
1.3 Where it applies in biometrics
AI-first workflows can support each biometrics pillar without changing who is accountable for the work.
| Pillar | Useful first pass | Human decision |
|---|---|---|
| Data management | Compare edit checks, draft queries, trace unresolved issues | Approve queries and database readiness |
| Statistical science | Find inconsistencies across the protocol and SAP, assemble evidence, draft questions | Approve design, methods, and interpretation |
| Analysis and reporting | Check traceability, execute code, compare TLFs, draft findings | Validate results and approve deliverables |
Many valuable workflows cross these boundaries. Reviewing a TLF may require the SAP, ADaM metadata, executed code, rendered output, and prior decisions. Domain expertise and end-to-end knowledge therefore become more important, not less.
1.4 Human accountability is not optional
In an AI-first team, the human role shifts from manually executing each task to orchestrating agents, tools, and human review. The agent can retrieve, compare, draft, run approved tools, and monitor routine conditions. The professional defines the objective, reviews exceptions, challenges the evidence, improves the workflow, and remains accountable for the decision.
This makes three human capabilities especially valuable: systems thinking, end-to-end knowledge, and exception handling.
An April 2026 FDA warning letter to Purolea Cosmetics Lab described a drug manufacturer that used AI agents to create specifications, procedures, and production records. FDA found that the firm had not adequately reviewed the generated documents for accuracy and CGMP compliance. The firm also reportedly relied on the agent’s failure to mention a requirement as a reason for not knowing that the requirement applied.
The letter concerns pharmaceutical manufacturing and CGMP, not clinical study reporting. Its workflow lesson still transfers: AI-generated content remains a draft until an authorized person verifies and approves it. The workflow must check authoritative requirements, disclose what it did not cover, and preserve the evidence behind the decision. An agent cannot be the source of regulatory truth or its own approver.
1.5 Start with one bounded workflow
An organization does not need to begin with autonomous, end-to-end execution. A useful AI-first workflow can be small:
- Choose one recurring, review-heavy task with visible source artifacts.
- Define one trigger, one expected output, and one accountable owner.
- Automate deterministic checks before adding model judgment.
- Require evidence, uncertainty, coverage, and explicit escalation.
- Run in advisory mode until real use demonstrates acceptable performance.