---
title: "AI Agents in Clinical Trial Documents, Humans Approve | Trelice"
description: "AI agents can draft protocol updates, comments, IRB responses, and amendments in Trelice. A human approves every change. Scale with AI without losing governance."
canonical_url: "https://www.trelice.com/ai-agents"
---

AI agents

# AI agents do the legwork. People approve every change.

In Trelice, an AI agent can draft protocol updates, field-level comments, IRB responses, and amendments. Its work lands as drafts inside structured sections with mandatory fields, review workflows, and an audit trail, and a person approves every change. An AI agent never approves. These are agents your team connects. Trelice itself uses AI only for importing existing documents, learning document formats, and drafting consent form sections and their translations, and its protocol authoring, review, and approval do not run on a language model. Trelice allows you to scale and bring AI to the table without losing governance, on both human and machine.

[Request a Tour](/contact) [Read the docs](/docs)

## LLMs and AI agents in plain words, for anyone who has been nodding along in meetings

Clinical research is full of people who are excellent at their jobs and quietly unsure what half the AI vocabulary means. That is nothing to be embarrassed about; the vocabulary is two years old and changes monthly. Here is what the two terms that matter actually mean, in plain words.

A **large language model**, or LLM, is a computer program that has read an enormous amount of text and learned from it how to read and write. Give it a fifty-page document and it can summarize it, answer questions about it, or draft a response to it, in seconds, in fluent professional language. What it does not do is know things the way a database knows things: it produces the most plausible text, which is usually right and sometimes confidently wrong. That one property, fluent but fallible, is the reason everything else on this page insists on human review.

An **AI agent** is an LLM that has been given tools and a goal. A plain LLM answers when you ask it something and then stops. An AI agent works through a task on its own: it can read your inbox, open documents, use software the way a person uses it, and take the next step based on what it found. "Read this IRB letter, find the protocol sections that answer each question, and draft the responses" is a task for an AI agent, not a chatbot.

Why does this matter to clinical trials in particular? Because clinical trials run on documents. A study produces protocols, amendments, consent forms, worksheets, committee letters, and meeting minutes, and most of the work is reading one of them carefully and writing or checking another. Reading carefully and drafting from what was read is precisely the work LLMs are good at, which is why AI agents are arriving in clinical operations faster than in most industries, whether or not anyone plans for them. The question worth planning for is not whether your team will use them. It is whether their work will land somewhere governed.

## The question is not whether AI agents will help. It is where their work lands.

Clinical teams are already using AI agents outside their document systems: an AI agent that summarizes the week’s literature, an AI agent that drafts a reply to a site’s question, an AI agent that turns a meeting recording into notes and action items. The output is useful and the output is homeless. It sits in a chat window or an email draft, and a person retypes the part that matters into the protocol, the response letter, or the tracker.

The riskier alternative is worse: letting the AI agent write directly into a Word document. A Word file has no notion of a mandatory field, a review state, or an approval gate. An AI agent can delete a safety paragraph, change a dose, or quietly rewrite an eligibility criterion, and the document accepts all of it. The only control is a person reading the whole file again, which is the control that was already failing before the AI agent arrived.

Trelice gives AI agent work a governed place to land. An AI agent interacts with the protocol the same way a person does: through named sections and fields, each with rules, each with a review state, each change recorded. What follows is what that looks like across the life of a study.

## The structure is the guardrail

Even when the AI agent inputs the data, the structure is provided by Trelice, and so are the mandatory fields. The AI agent cannot stray from them. A protocol in Trelice is not a page of free text, it is a set of [structured sections](/docs/structured-sections) with typed fields: the planned number of subjects is a number that appears consistently everywhere the protocol states it, an eligibility criterion is an item in a list, a visit in the Schedule of Assessments is linked to the assessments it carries. An AI agent filling these fields is held to the same validation as a person filling them.

That changes what an AI agent error can look like. It cannot leave the objectives section half-written and undetected, because an incomplete mandatory field blocks submission for review. It cannot change the dose in one section and not the other, because facts the protocol states in several places live in one field. It cannot break the chain from objective to endpoint to assessment without [golden threads](/docs/golden-threads) flagging it before approval. The guardrails were built to catch human inconsistency; they catch machine inconsistency the same way.

This is the line we build the whole feature set around: **Trelice allows you to scale and bring AI to the table without losing governance, on both human and machine.** Governance on the human means review workflows, roles, and signatures. Governance on the machine means the AI agent works inside the same structure, the same permissions, and the same audit trail, with one hard boundary described next.

## One thing an AI agent never does in Trelice: approve

An AI agent can draft a section, update a field, write a comment, assemble an amendment, and run a first review that flags inconsistencies for the humans who follow. It can do everything that ends in a recommendation. It cannot do the thing that ends in a decision: in Trelice, every approval is a human approval, for every change, with no exception for small changes and no exception for changes the AI agent itself made.

The mechanics make this real rather than aspirational. Sections move through [review and approval](/docs/reviewing-and-approving) workflows where the approver is a named person with the right role. [Electronic signatures](/docs/electronic-signatures) belong to people. The [approval queue](/docs/approval-queue) is a human’s worklist. A first review by an AI agent can make the human review faster, by flagging what changed, what drifted, and what looks inconsistent, but the human review is the one that counts, and the one the system requires.

The [audit trail](/docs/audit-trail) keeps the two visible. Every change is recorded with who made it, so a reviewer, a QA auditor, or an inspector can always see which content was drafted by an AI agent and which decisions were made by people. Automation that hides behind a person’s name is not governance, it is risk with better typography.

## What AI agent work looks like across a study

The pattern is the same in every case: the AI agent does the reading, the drafting, and the assembling, the work lands in Trelice as drafts and comments inside the normal workflows, and a person reviews and approves before anything becomes part of the study record.

### From research papers and emails to a protocol update

The background of a protocol rests on literature, and the literature does not stop moving when the protocol is signed off. An AI agent can read the new publications, the investigator correspondence, and the internal research summaries, and propose the update where it belongs: the affected sections of the protocol in Trelice, edited as drafts, with the source material cited in a comment on each change.

Because the proposal arrives as a draft inside [protocol authoring](/protocol-authoring), nothing about the process changes for the team. The section shows as changed, the assigned reviewer gets a pending task, and the edit goes through the same review it would get if a medical writer had typed it. The difference is that the medical writer starts from a grounded draft instead of a stack of PDFs.

### Meeting notes become comments on the right fields

Protocol review meetings produce decisions, and the decisions evaporate unless someone spends an evening writing them into the document. An AI agent can take the meeting notes and create the comments on a person’s behalf: the decision about the washout period lands as a [field-level comment](/collaboration) on the washout field, the open question about the second cohort lands on the study design section, each one addressed to the person who owns the answer.

The person the AI agent acted for sees the comments before their colleagues are notified, confirms or edits them, and the discussion continues inside Trelice where it is attached to the exact field it concerns. The audit trail records how each comment was created. Nothing about a meeting is lost, and nobody spends the evening on transcription.

### IRB responses, even when the questions arrived by email

An IRB or ethics committee sends its questions wherever it likes, usually a letter attached to an email. An AI agent can read that letter, match each question to the protocol content that answers it, and draft the responses in Trelice, grounded in the approved protocol rather than in anyone’s memory of it. Where Trelice has already built an [IRB Q&A](/downstream-documents#irb-questionnaire) from the protocol, the AI agent starts from answers that are current by construction.

The draft responses then wait for the people responsible. The study lead reviews each answer against the question, corrects what needs correcting, and approves. Only after that human approval does anything go back to the committee. An AI agent that answers an ethics committee on its own is exactly the failure mode this design exists to prevent.

### During the study, the AI agent starts the amendment

Amendments are where document work hurts most, because they happen under time pressure in a running study. An AI agent can start the [amendment](/protocol-authoring#amendments) in Trelice the moment the need is known: open it, carry in the proposed edits as drafts, draft the reason for each change, and leave the whole package in the team’s pending tasks.

From there the normal machinery takes over. Reviewers work through the changed sections, the approvers approve or send back, and once the amendment is approved the [amendment cascade](/downstream-documents) flags every downstream document that the change touches, consent forms and visit worksheets included. The AI agent compressed the first week of an amendment into the first hour. The decisions stayed where they were.

### From a safety query in the EDC to an open amendment

The sharpest version of this is safety. An AI agent connected to your EDC can watch for the signals the team defines, and when a safety query fires, start the amendment in Trelice at once: the affected protocol sections identified, the trigger documented in the amendment’s reason, the medical monitor and the safety team notified through their pending tasks.

Nothing in the protocol has changed at that point, and nothing will change until the safety team and the approvers decide it should. What the AI agent removed is the gap between the signal and the start of the response, the days in which someone notices the query, finds the protocol file, and schedules the meeting. What the AI agent did not remove is a single decision. The humans arrive at an amendment that is already open, already scoped, and entirely theirs to accept, reshape, or close.

## The loop is enforced by the system, not promised by the vendor

Human in the loop is an easy sentence to say and a hard property to fake in Trelice, because the loop is the product’s normal operation. Work becomes real only by passing through states that require a person: a draft is not reviewed until a reviewer acts, a section is not approved until an approver signs, an amendment is not effective until it completes its workflow. An AI agent can fill the pipeline. It cannot open the valve.

The surrounding machinery keeps the people informed rather than assuming they will check. [Pending tasks](/docs/pending-tasks) and [notifications](/docs/notifications) tell each person what is waiting on them, the dashboard counts what is open per study, and [review workflows](/docs/review-workflows-and-approvals-setup) are configured by the organization, so the number of human eyes on each kind of change is a policy the sponsor sets, not a default the AI agent can negotiate.

## Governance on the human and on the machine

The reason teams hesitate to bring AI into regulated documents is that most tools offer a trade: speed for control. Trelice refuses the trade by making both sides run on the same rails. People and AI agents work through the same [roles and permissions](/docs/roles-and-permissions), so an AI agent holds exactly the access its role grants and nothing more. Both are held to the same mandatory fields and validation. Both leave the same [audit trail](/docs/audit-trail), with every change attributed, so oversight of the machine is the same review discipline the team already applies to each other.

That is what the line means in practice. Trelice allows you to scale and bring AI to the table without losing governance, on both human and machine: more documents, more studies, and more automation, with the same number of uncontrolled changes as before, which is zero.

## Getting started

Everything this page relies on inside Trelice is shipped and documented today: structured sections with mandatory fields, [review and approval](/docs/reviewing-and-approving) workflows, the [approval queue](/docs/approval-queue), [roles and permissions](/docs/roles-and-permissions), and the [audit trail](/docs/audit-trail). If you are planning AI agent workflows around your studies, from research monitoring to EDC-triggered amendments, we would like to hear what you want the AI agent to do, and we will show you where the human approvals sit in each one.

None of this is required. Trelice runs without AI: whether an AI agent works in your studies is your decision, and Trelice itself uses AI only for importing documents, learning formats, and drafting consent form sections and their translations, all of which can be turned off. A team that wants no AI anywhere gets a platform that works exactly as described on every other page of this site.

**Related:** [Protocol authoring](/protocol-authoring)[Collaboration and review](/collaboration)[Scaling clinical programs](/scaling-clinical-programs)[Security](/security)[Developer resources](/developers)

## Bring the AI agent. *Keep the approvals.*

Tell us what you want an AI agent to do across your studies. We will walk through where the work lands in Trelice and show you every point where a human signs off.

[Request a Tour](/contact)
