Applied Clinical Trials · Published 2026-04-10

How Agentic AI Will Realistically Reshape the Clinical Trial Workforce

privateReview: unreviewedConfidence: high
Agentic AIclinical trialsworkforce planningaugmentation levelsthermo fisher scientificagentic-aiHuman-in-the-loopautonomy-levelsverificationGuardrailsclinical-trialsworkforce-planningauditability
Open on YouTube ↗

Summary

Overview

  • Speaker: Krishna Cheriath, VP, head of clinical research digital data and AI, Thermo Fisher Scientific
  • Channel: Applied Clinical Trials
  • Main topic: The integration and workforce impact of agentic AI in clinical trials
  • Purpose: Provide realistic insights into how agentic AI will transform clinical trial execution, workforce structures, and operational roles over the next few years. Krishna Cheriath, VP and head of clinical research digital data and AI at Thermo Fisher Scientific, discusses how agentic AI is poised to realistically reshape clinical trial design and execution. He outlines a four-level framework for AI augmentation (AL0 to AL4) and explains how clinical trial roles will evolve from human-centric to AI-augmented over the next few years.

Topic Map

Introduction to Agentic AI in Clinical Trials

  • Explanation: Introduction of Krishna Cheriath and the perspective that agentic AI is a lasting wave that will transform clinical trial work.
  • Key claims:
    • Agentic AI is here to stay and will fundamentally impact clinical trial work.
  • Examples:
    • Tech waves over 30-35 years in the industry.
  • Terminology:
    • agentic AI
    • clinical trials
  • Why it matters: Establishes that AI integration is a structural industry shift rather than a temporary fad.

The AI Augmentation Framework (AL0 to AL4)

  • Explanation: A structured framework defining four levels of AI augmentation in clinical trials, ranging from human-dominated work to full autonomy.
  • Key claims:
    • Work in clinical trials can be scaled across four levels of AI augmentation.
  • Examples:
    • Chocolate factory sample line illustrating auditing vs. full automation.
  • Terminology:
    • AL0
    • AL1
    • AL2
    • AL3
    • AL4
    • AI augmentation
  • Why it matters: Provides leaders with a mental model to evaluate where clinical trial roles sit on the AI spectrum today and in the future.

Workforce Evolution and Strategy

  • Explanation: Discussion on how clinical research roles (programmers, biostatisticians, clinical data analysts) will shift from human-centric to AI-augmented, mirroring human workforce development strategies.
  • Key claims:
    • All clinical trial roles will be AI-augmented over the next three to four years.
    • A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.
  • Examples:
    • Employee career progression from entry level to mid-level and senior expert.
  • Terminology:
    • workforce planning
    • AI agents
    • entry level AI
  • Why it matters: Organizations must proactively plan workforce development around AI agents with appropriate guardrails.

Key Points

Human plus digital workforce

  • Explanation: The future of clinical research work is a combination of human expertise and digital/AI workforce execution.
  • Evidence: Observed shifts across 30-35 years of industry technology waves.
  • Practical implication: Leaders must critically evaluate current roles and project their evolution over 6 to 24 months.

Need for rigorous guardrails

  • Explanation: While the potential of agentic AI is huge, it must be deployed with appropriate governance and compliance guardrails.
  • Evidence: Clinical trials deal directly with patient-centric data.
  • Practical implication: Must maintain safety, compliance, and patient trust while accelerating scientific progression.

Frameworks, Models & Processes

AI Augmentation Levels (AL0 - AL4)

  • How it works: Classifies the division of labor between humans and AI across four distinct operational tiers.
  • Components:
    • AL0: Current state where work is predominantly done by humans with tools and assistance.
    • AL1: Combination of human and AI assistance where human verifies and validates every output.
    • AL2: Majority of work done by AI, with human validating outputs and acting as verifier/certifier.
    • AL3: AI does the bulk of work while humans perform auditing, sampling, and quality control.
    • AL4: Fully autonomous AI agents executing tasks independently.
  • When to use: Evaluating clinical trial tasks, operational workflows, and automation readiness.

Examples & Case Studies

A chocolate factory assembly line example was used to explain quality control.

  • Illustrates: The difference between inspecting every item versus random sampling and auditing (AL3).
  • Lesson: As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.

Actionable Takeaways

  • Immediate:
    • Critically examine existing roles in clinical research, biostatistics, and data management.
    • Assess tasks against the AL0 to AL4 augmentation framework.
  • Strategic:
    • Develop a structured AI agent workforce strategy analogous to human career progression (entry, mid, expert).
    • Establish strict guardrails ensuring compliance, security, and patient-centric data integrity.
  • Questions to investigate:
    • Where will specific clinical trial roles fall on the AL1-AL3 scale over the next two years?
    • How do we balance autonomous AI execution with regulatory compliance and patient safety?

Claims Worth Verifying

  • All clinical trial roles will be AI-augmented across levels AL1 through AL3 over the next two years. (prediction)

Notable Quotes

"Among the spectrum of AI optimist to AI pessimist, I am an AI realist, and I believe this agentic AI wave is here to stay." "The future of this work is going to be a combination of human plus digital workforce."

Compressed Summary

  • Agentic AI is a permanent wave transforming clinical trial design and execution.
  • Introduces the AL0 to AL4 framework defining levels of human-AI labor division.
  • Predicts clinical roles will shift from human-centric to AI-augmented over 3 to 4 years.
  • Requires structured workforce planning and compliance guardrails for patient-centric data.
  • Keywords: agentic ai, clinical trials, workforce planning, augmentation levels, thermo fisher scientific
  • Core insight: Agentic AI will systematically shift clinical trial roles from human-dominated execution to AI augmentation tiers (AL1-AL3), requiring strategic workforce planning and robust compliance guardrails.

Core insights

5
Mental Modelmedium noveltystrong evidence

The AL0-AL4 framework is organized around one axis: who verifies the output and at what granularity. AL1 = human validates every output, AL2 = human certifies/validates AI output, AL3 = human performs auditing, sampling, and quality control, AL4 = no human. The level is a property of the workflow's verification topology, not of the model's capability.

Why it matters

Gives engineers a capability-independent way to classify and design agent workflows: instead of asking 'is the model good enough', ask 'where in the pipeline does verification live and at what sampling rate'. That reframes autonomy as an oversight-architecture decision, which is directly implementable as different runtime topologies (inline gate vs. out-of-band auditor).

Generalization

Any agentic system operating under audit or safety constraints can be positioned on a verifier-granularity axis (per-output gate -> certification -> statistical sampling -> none), independent of the underlying model's raw capability.

AL1: Combination of human and AI assistance where human verifies and validates every output.
Open source video
AL2: Majority of work done by AI, with human validating outputs and acting as verifier/certifier.
Open source video
AL3: AI does the bulk of work while humans perform auditing, sampling, and quality control.
Open source video
AL4: Fully autonomous AI agents executing tasks independently.
Open source video
Mechanismmedium noveltymoderate evidence

Human oversight degrades predictably in granularity as agent volume rises: 100% verification is replaced by sampling and auditing. The binding constraint is verification bandwidth, not model quality.

Why it matters

Implies that scaling agent throughput forces a separate QA subsystem (sampling policy, statistical acceptance criteria, audit artifacts) rather than simply adding more human reviewers. The pipeline must emit machine-consumable evidence so that sampled outputs can be audited after the fact.

Generalization

Any human-in-the-loop system under volume pressure converges on statistical sampling; this requires the runtime to persist auditable per-step provenance, otherwise sampling is impossible to perform retrospectively.

As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.
Open source video
A chocolate factory assembly line example was used to explain quality control.
Open source video
Architecturemedium noveltymoderate evidence

The AL levels are applied at the task/workflow level, while the workforce claim is made at the role level ('all clinical trial roles will be AI-augmented'). Roles decompose into many tasks, each of which may sit at a different AL level, so role-level autonomy is a derived aggregate rather than a single setting.

Why it matters

Forces explicit task decomposition before any autonomy assignment; a system or org that assigns a single autonomy level per role/agent will misplace verification effort. The correct unit of analysis is the task boundary inside the workflow.

Generalization

Autonomy configuration in agentic systems should be attached to task nodes in a workflow graph, not to agents or job titles; agent-level autonomy is an emergent summary of its task mix.

When to use: Evaluating clinical trial tasks, operational workflows, and automation readiness.
Open source video
All clinical trial roles will be AI-augmented over the next three to four years.
Open source video
Practicemedium noveltyweak evidence

The speaker proposes applying workforce development machinery to AI agents: entry, mid, and expert tiers, with a structured strategy analogous to human career progression.

Why it matters

If taken seriously, this implies agent lifecycle infrastructure that most teams do not build: staged onboarding, tiered supervision levels, explicit promotion criteria between tiers, and guardrails appropriate to each tier. It turns agent deployment into a managed roster with progression rules rather than a fixed prompt/model configuration.

Generalization

Any organization running many agents will need an HR-like control plane: capability tiers, supervision ratios per tier, and promotion/demotion rules tied to measured reliability.

Develop a structured AI agent workforce strategy analogous to human career progression (entry, mid, expert).
Open source video
A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.
Open source video
Predictionhigh noveltymoderate evidence

In a regulated, patient-data domain the stated autonomy target is capped at AL1-AL3, even though AL4 (full autonomy) is defined in the framework. The framework enumerates full autonomy but the deployment prediction excludes it.

Why it matters

Suggests the practical engineering ceiling in high-stakes domains is set by auditability and compliance, not by achievable model capability. Designing for AL4 in such domains is speculative work; designing AL2-AL3 with an audit trail is the near-term tractable target.

Generalization

In any regulated or high-liability domain, the useful autonomy range is bounded by what can be audited and attested, not by what the agent can technically execute.

All clinical trial roles will be AI-augmented across levels AL1 through AL3 over the next two years.
Open source video
Must maintain safety, compliance, and patient trust while accelerating scientific progression.
Open source video

Deep dives

4

Task-level autonomy topology for agentic clinical workflows

Research question

How should clinical trial roles be decomposed into task nodes, and how should each node be assigned an AL level, verifier, and audit artifact schema based on its verification topology?

Why

Role-level autonomy claims obscure the fact that roles decompose into tasks with different verification needs. Assigning one autonomy setting per role or agent misplaces human review effort and makes governance imprecise. A task-node map is the minimum unit for implementable autonomy policy.

AL1: Combination of human and AI assistance where human verifies and validates every output.
Open source video
AL2: Majority of work done by AI, with human validating outputs and acting as verifier/certifier.
Open source video
AL3: AI does the bulk of work while humans perform auditing, sampling, and quality control.
Open source video
Source video

Statistical sampling for AL3 oversight

Research question

What sampling rates and statistical acceptance criteria bound residual defect probability when AI produces the bulk of clinical trial work?

Why

As agent volume rises, 100% verification is replaced by sampling and auditing. That shift turns oversight into a statistical QA problem requiring provenance, defect-rate assumptions, sampling plans, and acceptance thresholds. Without this, AL3 is an aspiration rather than an auditable operating mode.

As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.
Open source video
A chocolate factory assembly line example was used to explain quality control.
Open source video
Source video

Auditability ceiling in regulated clinical domains

Research question

What audit log schema, sampling policy, escalation trigger, and attestation format make AL2-AL3 agents deployable under clinical-trial compliance and patient-trust constraints?

Why

The framework defines AL4 full autonomy, but the stated deployment target caps at AL1-AL3 for clinical trials. That implies the binding constraint is not model capability but auditability and compliance. Engineering effort should target attestable AL2-AL3 workflows rather than speculative AL4 autonomy.

All clinical trial roles will be AI-augmented across levels AL1 through AL3 over the next two years.
Open source video
Must maintain safety, compliance, and patient trust while accelerating scientific progression.
Open source video
Source video

Agent HR control plane: tiered progression and promotion gates

Research question

Do reliability-based tier-transition gates for AI agents improve downstream audit outcomes compared with flat, uniformly supervised deployments?

Why

The workforce analogy proposes entry, mid, and expert tiers for agents. If operationalized, it requires measurable promotion and demotion criteria tied to reliability, plus tier-specific supervision ratios and guardrails. Without measurable gates the analogy is organizational storytelling rather than governance infrastructure.

Develop a structured AI agent workforce strategy analogous to human career progression (entry, mid, expert).
Open source video
A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.
Open source video
Source video

Article ideas

4

Autonomy Is an Oversight Topology, Not a Model Capability

Agent autonomy levels should be classified by who verifies outputs and at what granularity, not by benchmark scores or model capability. AL0-AL4 are workflow architecture choices that determine whether verification is an inline gate, a certification step, an out-of-band sampling function, or absent.

Angle

Reframe autonomy as an oversight-architecture decision and show how different verification topologies map to different runtime designs, latency profiles, and audit requirements.

Source video

The Verification Bandwidth Ceiling: Why Agentic Scale Requires Statistical QA

The limiting factor for scaling agentic work is human verification bandwidth, not model intelligence. When agent volume rises, 100% review becomes impossible, forcing sampling policies, retained provenance, and statistical acceptance criteria as first-class infrastructure.

Angle

Use the manufacturing quality-control analogy to argue that agent platforms need an out-of-band QA subsystem before they can credibly run at higher autonomy levels.

Source video

Stop Assigning Autonomy Per Role: Task-Level Guardrails for Agentic Workflows

Assigning a single autonomy level to an agent or job role misallocates verification because roles contain tasks at different AL levels. Autonomy policy should attach to task nodes in a workflow graph, with per-node verifiers, guardrails, and audit artifacts.

Angle

Show why role-level autonomy is an aggregate fiction and how task-node metadata changes orchestration, evaluation reporting, and guardrail placement.

Source video

Agent HR: When Workforce Metaphors Help and Hurt Agent Governance

Entry, mid, and expert tiers for AI agents can structure supervision and guardrails, but the human-career analogy breaks if it assumes monotonic capability growth or stable tiers. Agent progression needs measured reliability gates and demotion paths, not title-based trust.

Angle

Examine the value and failure modes of treating agent fleets like a workforce roster with promotion criteria, supervision ratios, and tier-specific guardrails.

Source video

Project ideas

4

Task-Level Autonomy Mapper (AL-TAM)

gatehouse

Attaching AL level, verifier type, and audit artifact schema to each task node reduces misplaced human review and defect escapes compared with a single agent-level autonomy setting.

Proof of concept

Build a small orchestration graph for synthetic clinical data review tasks and label each node AL1-AL3 with verifier stubs, provenance logging, and sampling hooks; compare it against a flat agent-level setup on matched task batches.

Measurement

Human review minutes per 100 outputs; defect escape rate; percentage of outputs with complete auditable provenance; reviewer agreement per task node.

Source video

SPC for Agent Outputs

beyond-evals

Sampling plans derived from statistical process control keep residual defect probability below a target while reducing manual review effort versus 100% verification for AL3-style agent workflows.

Proof of concept

Simulate agent outputs with known defect rates, implement a sampling auditor with confidence bounds and acceptance criteria, and compare audit yield and effort against full human review.

Measurement

Residual defect probability; review effort reduction; sample size required to detect a defect shift; false-negative rate at target confidence.

Source video

Agent Promotion Gate

new

Tier-transition gates based on measured reliability over sampled outputs predict downstream audit outcomes better than static agent deployment or tenure-based promotion.

Proof of concept

Run two agent fleets on the same clinical data tasks: one with reliability-based promotion and demotion gates across entry, mid, and expert tiers, and one flat deployment.

Measurement

Audit pass rate; regression incidents after promotion; false promotion rate; demotion detection latency after known regression.

Source video

Clinical Trial Compliance Attestation Harness

gatehouse

Requiring AL2-AL3 agents to emit a standard audit log, sampling attestation, and escalation record is sufficient for simulated clinical-trial compliance review without full output inspection.

Proof of concept

Build a checker that rejects runs lacking required artifacts and test it on synthetic clinical trial workflows that vary AL level, sampling rate, and escalation triggers.

Measurement

Percentage of runs passing compliance checks; manual review time per run; audit findings per 100 runs; attestation completeness.

Source video

Architectural implications

3

Verification responsibility moves out of the inline execution path as autonomy increases.

Before

Human reviewer sits inside the pipeline as a blocking per-output gate (AL1).

After

Human or human-defined policy runs as an out-of-band auditing and sampling function over a stream of AI-produced artifacts (AL3).

Consequence

The runtime must emit structured, retained artifacts and provenance for every unit of work so that sampling can occur after execution; non-blocking QA changes latency and cost profiles and requires a sampling policy with acceptance thresholds.

Source video

Autonomy is assigned per task node rather than per agent or per role.

Before

A single autonomy setting or model configuration governs an entire agent or job function.

After

Each task node in the workflow carries its own AL level, verifier, and guardrail set.

Consequence

Orchestration layers need per-node policy metadata and routing; evaluation must be reported per task node, not per agent.

Source video

Agent fleet management gains an HR-like control plane.

Before

Agents are deployed as static configurations with fixed prompts, tools, and permissions.

After

Agents occupy tiers with defined supervision levels and explicit progression criteria, with tier-appropriate guardrails.

Consequence

Requires measurement of agent reliability per tier as a promotion gate, plus the ability to demote or quarantine an agent after regression — i.e., continuous evaluation wired into deployment rather than one-time validation.

Source video

Tradeoffs and failure modes

3

Sampling-based oversight vs. residual defect risk

Benefit

Human effort scales sublinearly with AI output volume, enabling throughput gains.

Cost or risk

Switching from 100% verification to sampling means defects that fall outside the sample go undetected; the guarantee becomes statistical, not absolute, which is a weaker safety posture for patient-facing data.

As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.
Open source video
Source video

Autonomous execution vs. regulatory compliance and patient trust

Benefit

Higher AL levels accelerate scientific progression and reduce cycle time.

Cost or risk

Clinical trials operate on patient-centric data, so autonomy must be constrained by governance, security, and compliance; the acceleration benefit is capped by the auditability of each automated step.

Must maintain safety, compliance, and patient trust while accelerating scientific progression.
Open source video
Source video

Workforce analogy applied to agents

Benefit

A familiar mental model (entry/mid/expert) gives leaders a concrete way to plan agent rollout and guardrails.

Cost or risk

The analogy imports assumptions from human career development that may not hold for software artifacts (e.g., that capability increases monotonically with tenure, or that tiers are stable over time), risking misplaced trust or over-engineering of agent 'progression'.

A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.
Open source video
Source video

Open questions

4

Where do specific clinical trial roles and tasks fall on AL1-AL3 over the next two years?

Why unresolved

The talk states the framework and a directional prediction but provides no per-task assignments or measurement of current AL levels.

Research direction

Build a task-level inventory of clinical trial workflows and instrument each task against a verification-granularity rubric, producing a baseline map of current AL levels.

Source video

How is autonomy balanced against regulatory compliance and patient safety in a concrete deployment?

Why unresolved

The talk names the tension and asserts guardrails are needed but does not specify guardrail mechanisms, review gates, or evidence standards.

Research direction

Define and pilot a guardrail specification for AL2-AL3 agents: audit log schema, sampling rate policy, escalation triggers, and attestation format acceptable to regulators.

Source video

What sampling rate and statistical acceptance criteria make AL3 oversight sound when AI produces the bulk of the work?

Why unresolved

The chocolate-factory analogy introduces auditing and sampling but no defect-rate assumptions, sample sizes, or confidence levels are given.

Research direction

Adapt statistical process control to agent outputs: model agent error rates, derive sampling plans that bound residual defect probability, and measure whether agent errors are independent enough for sampling to hold.

Source video

What are the promotion criteria that move an agent from AL1 to AL2 to AL3?

Why unresolved

The workforce-tier strategy is proposed by analogy but the talk does not state measurable transition conditions.

Research direction

Define tier-transition gates in terms of measured reliability (e.g., agreement rate with human verifier over N sampled outputs) and test whether they predict downstream audit outcomes.

Source video

Key claims

7
predictionVerification needed

Clinical trial roles will be AI-augmented across levels AL1 through AL3 within roughly two years.

Evidence

All clinical trial roles will be AI-augmented across levels AL1 through AL3 over the next two years.

Question

Is there any measured adoption data showing roles moving to AL2/AL3 within two years, and does the two-year figure conflict with the separately stated three-to-four-year horizon?

Source video
causalVerification needed

As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.

Evidence

As AI takes on more volume, human intervention shifts from 100% verification to sampling and auditing.

Question

Do real deployments show verification bandwidth (not model capability) as the binding constraint that forces the shift from full review to sampling?

Source video
factualVerification not requested

Work can be classified across four distinct tiers of human-AI division of labor (AL0-AL4).

Evidence

Classifies the division of labor between humans and AI across four distinct operational tiers.

Source video
opinionVerification needed

A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.

Evidence

A structured workforce strategy mirroring entry, mid, and expert levels is required for AI agents.

Question

Do tiered agent rosters with progression rules actually improve reliability or governance outcomes versus flat, uniformly supervised deployments?

Source video
predictionVerification not requested

The future of clinical research work is a combination of human plus digital workforce.

Evidence

The future of this work is going to be a combination of human plus digital workforce.

Source video
causalVerification needed

Clinical trials deal directly with patient-centric data, therefore agentic AI must be deployed with governance and compliance guardrails.

Evidence

Must maintain safety, compliance, and patient trust while accelerating scientific progression.

Question

What specific guardrail mechanisms (review gates, audit logs, attestation) are required to satisfy clinical-trial compliance regimes for AL2-AL3 agents?

Source video
opinionVerification not requested

Agentic AI is a lasting structural shift rather than a temporary technology wave.

Evidence

Among the spectrum of AI optimist to AI pessimist, I am an AI realist, and I believe this agentic AI wave is here to stay.

Source video

Connections

4