How Sponsors Can Prepare for Real-Time Clinical Trial Execution

Most clinical teams use AI to assist work. AI agents go further: connected to trial systems, they perform tasks such as sending reminders or updating statuses. Because those actions change trial records directly, a sponsor should be able to explain how the workflow will operate before authorizing it. Which records will the agent use? What can it change? Who reviews consequential actions? What happens when information is incomplete, or a system rejects an update?

That raises the oversight bar because agent’s error can repeat across many records before anyone notices. Answering these questions makes AI risk management in clinical trials a practical preparation process and gives clinical, quality, data, safety, and technology teams a shared basis for deciding where AI can contribute.

This final article in our Real-Time Clinical Trials series explains how to start, control, and expand. For context, read our pillar guide Real-Time Clinical Trials Are Now an FDA Priority. Maxis AI Has Been Building for This Shift.

Why Real-Time Readiness Is More Than a Technology Project

Real-time readiness starts with a change in who does the work: software now carries out some actions under delegated authority. That makes readiness more than a technology project: teams must decide what to delegate, who supervises it, and what confirms completion.

For example, an agent may flag an overdue CTMS task, but it stays open if the sponsor and CRO have not agreed who owns it.

Sponsors must therefore define the operating arrangements around an AI-supported workflow before allowing agents to perform live tasks. This includes identifying the process owner, assigning responsibility for reviewing agent outputs, defining approval authority, establishing escalation routes, and ensuring that qualified personnel are available when human intervention is required.

This cross-functional approach reflects the direction regulators have set. The FDA and EMA's Guiding Principles of Good AI Practice in Drug Development (January 2026) emphasizes a clearly defined context of use, human-centric design, multidisciplinary expertise, data governance, and validation and oversight proportionate to risk. [1]

For the broader team implications, read What FDA's Real-Time Clinical Trial Initiative Means for Clinical Operations Teams.

How Sponsors Can Prepare, Step by Step

Preparation follows a sequence: map the workflow, choose a bottleneck, map systems of record, define AI’s limits, then govern, pilot, and measure before scaling.

Assess the Current Signal-to-Action Workflow

Controls belong where information enters, decisions are made, and work changes hands, so map those first. Trace how recent cases of a recurring issue moved from detection to closure, including email and spreadsheets. Focus on four areas:

Where Signals Originate

Identify the record, report, system alert, or event that initiates the workflow and the authoritative source for it, so the agent acts on the official record, not a copy. Check whether the information reaches the team promptly and provides enough context to act.

Who Reviews Them

Identify who reviews the signal, determines its significance, and authorizes the next step. Confirm that they have the information, decision-making authority, and backup coverage to respond on time.

Where Work Stalls

Separate time spent doing the task from time spent waiting for information, assignment, or approval. Look for repeated handoffs, duplicate checks, and unclear responsibility for closure.

How Escalation Happens

Define what triggers escalation, who receives it, and when a response is expected. Establish a backup route if the owner is unavailable or the issue remains unresolved. Also define when the agent must stop because its authority or information is insufficient.

For the distinction behind this exercise, see Real-Time Trial Monitoring vs. Real-Time Trial Execution.

Identify High-Value Execution Bottlenecks

The best first workflow recurs often, has a clear definition of done, and runs on reliable information. Weigh value against task risk: start where an agent’s error is easy to detect and reverse, not where it could alter clinical data or safety reporting.

These are areas worth assessing, not blanket automation candidates; judgment-heavy decisions stay with people.

Which workflow should sponsors start with? One with stable rules, reliable inputs, narrow permissions, and an outcome someone can independently verify. A big backlog alone is not a reason to start somewhere. If clearing it needs clinical decisions not yet made, or data that disagrees across systems, an agent will only push those problems downstream faster.

Map Systems of Record and Data Flows

Your team must know which system is authoritative for each status, value, and approval. When systems disagree, the agent needs a rule, not a guess.

For each source, document refresh timing, matching identifiers, permitted write locations, and access restrictions. In blinded studies, ensure the agent cannot see or reveal treatment assignments, and limit access to personal information to what the workflow genuinely needs.

ICH E6(R3) recommends defined, validated interfaces between computerized systems, and addresses system validation, data integrity, access controls, and the management of data transfers and system failures. [2] For AI workflows, also test what happens when a transfer fails, arrives incomplete, or is silently rejected.

Define What AI Can and Cannot Do

An agent’s actions take effect directly in trial systems, so its boundaries are the primary control. Specify what it may do alone, what needs approval, and what is off-limits:

Workflow Boundaries

Define the trigger, eligible records, permitted steps, completion condition, and explicit exclusions.

Permissions

Separate permission to read, write, and execute. Match access to the agent’s role and name who can grant or revoke it.

Human Checkpoints

Define what needs human review before proceeding, who can review and approve it, and what information they need. Set response times and designate backup reviewers.

Escalation Rules

The agent should stop on conflicting records, missing authorization, failed updates, or out-of-scope work, and assign an exception owner.

When it stops, the agent should record its progress, hand the case over, and roll back or flag partial updates.

Together, these controls turn AI risk management in clinical trials into operating rules for each task. For the underlying principles, read Why Governed AI Agents Matter in Real-Time Clinical Trials.

Establish Governance Before Scaling

Boundaries set what the agent may do; governance proves it stayed within them. Before broader use, review four areas.

Traceability

Reconstruct a completed case. The AI audit trail should capture the triggering signal, data used, agent and configuration version, actions, approvals (who and when), before-and-after values, and the receiving system’s confirmation, distinguishing an attempted update from an accepted one.

Validation

Validate the end-to-end workflow, not just the model: intended use, acceptance criteria, and tests for routine work, stale or conflicting data, rejected updates, unavailable reviewers, and recovery. Approving individual actions does not replace this.

Accountability

Assign owners for approval, supervision, exceptions, and corrective action, including at every cross-organization handoff.

Reproducibility and Controlled Change

Retain the versions and configurations needed to repeat evaluations and assess changes before release.

ICH E6(R3) recommends recording workflow actions alongside data changes and risk-based validation with change control. [2] It is a basis for evaluation, not agent certification.

For AI compliance in clinical research, what matters is evidence about the deployed workflow, not the model in isolation.

Start With Bounded, Measurable Workflows

An “AI everywhere” rollout overloads reviewers, hides failures, and leaves no baseline.

Instead, give the agent one clearly limited task.

Example: chasing overdue queries in one study. The agent drafts reminders for approval and logs outcomes; it cannot change data or close queries.

Roll out in stages:

1. 1.Evaluate without live actions. Run in test mode with no record changes; compare outputs with reviewed examples.
2. 2.Introduce permitted actions under supervision, with agreed approvals, monitoring, and traceable records.
3. 3.Review before expanding against the acceptance criteria.

Also test how to pause execution in an outage.

Measure Execution Outcomes

Record a baseline first and agree what starts and stops each measurement.

Look at the longest delays, not just averages, and check for hidden costs: more corrections, heavier reviewer load, missed exceptions, or rubber-stamp approvals. A faster workflow that creates rework or weakens oversight is not ready to expand. Measuring this way keeps AI risk management in clinical trials tied to real operational performance.

Build Toward an AI Workforce Operating Model

As more agents go live, sponsors need a shared way to assign work, supervise agents, and manage change.

• Keep an inventory of active agents, owners, tasks, permissions, connected systems, configurations, and versions.
• Use shared processes for incident review and change control.
• Guard the handoffs. An agent receiving work from another should confirm it is complete, current, and authorized.
• Plan reviewer capacity, or automation simply moves the backlog to the approval queue.

The FDA and EMA principles call for scheduled monitoring and periodic reevaluation across the AI lifecycle. [1] Building that rhythm into the operating model keeps AI compliance in clinical research aligned with workflows, data, and system behavior as they change.

Maxis AI’s AI Workforce for Clinical Trials follows this model: governed agents with scoped permissions, human checkpoints, named exception owners, and a reviewable audit trail, starting with one bounded workflow.

Sponsor Readiness Checklist

Before authorizing live execution, confirm that:

• The workflow has an owner and a verifiable completion condition.
• Baseline turnaround, staff effort, and errors have been measured.
• Authoritative records and data flows are documented.
• Access preserves privacy, blinding, and role restrictions.
• Allowed actions and exclusions are explicit.
• Reviewers and exception owners have backup coverage.
• Testing covers routine work, failures, and recovery.
• Action and approval records can be reconstructed.
• Staff can pause execution and manage unfinished work.
• Expansion criteria and ongoing review responsibilities are agreed.

Key Takeaways

1. 1.AI agents require defined boundaries: access, actions, and exclusions set before deployment.
2. 2.Human oversight must be meaningful, with reviewers who can act.
3. 3.Validation must cover the complete workflow, including failures.
4. 4.Accountability and traceability are essential: named owners and a usable AI audit trail.
5. 5.Expansion should be evidence-led, measuring quality alongside speed.

Frequently Asked Questions

Prepare One Workflow for Governed Execution

Real-time execution is built one governed workflow at a time, on five foundations: defined controls, appropriate permissions, meaningful human validation, clear accountability, and traceable actions. Choose a workflow, bring its owners together, and document how each will work.

For sponsors exploring an AI Workforce for Clinical Trials, this exercise also creates a concrete starting point for discussing operational fit.

References