How bowlyn uses AI. And where it doesn't.
bowlyn builds clinical trial software. Bridge, its first product, uses AI to draft operational guidance that a person who's run trials reviews before it reaches you. This page explains how that works, what AI does, where it doesn't operate, and how your data is handled. It's written for the quality, regulatory, and IT teams who vet vendors.
Bridge is in production rollout. This page describes its security architecture.
Four principles.
A person is always in the loop.
AI never sends, files, or commits anything on its own. Every answer that leaves bowlyn has been reviewed and approved by someone who's run trials.
AI doesn't touch regulated artifacts.
It doesn't generate, modify, or submit content into the Trial Master File, eTMF, regulatory submissions, or IRB packages. A person carries any artifact across the regulated boundary.
AI is used for leverage, not judgment.
It's good at synthesizing protocols, drafting first responses, and reading long documents. It doesn't make operational calls under uncertainty. The workflow is built around that line.
What AI did is transparent.
The person is the author of record. AI is a drafting tool.
What the AI does.
Drafts operational answers for review.
A question comes in through the portal or by email. The AI drafts a first-pass answer from that trial's protocol and the regulations. Someone who's run trials reviews, edits, and approves it before it's sent.
Reads and summarizes documents.
It helps surface key information from protocols, vendor agreements, monitoring reports, and change orders, flagging inconsistencies. Answers are checked against the source before anyone relies on them.
Where the AI does not operate.
It does not submit or file anything to a regulated system. It does not write your regulatory documents. It does not communicate directly with sites, CROs, vendors, regulators, or patients. It does not make decisions for the sponsor. And it is designed to decline protected health information.
Your data.
Where it runs.
Model inference runs on AWS via Amazon Bedrock, under a business associate agreement with AWS.
What's sent.
Operational context: program details, the question, document excerpts. It's minimized and structured. Protected health information is not sent.
Training.
Data sent through Bedrock is not used to train foundation models.
Tenant isolation.
Every customer's data is isolated at the database level. An automated test runs on every code change and blocks any merge that could let one customer's data reach another.
Encryption.
Encrypted in transit and at rest.
Sign-in.
People who work at bowlyn sign in with multi-factor authentication. Customer sign-in uses single-use, expiring links.
Audit trail.
Every question, answer, and review is logged. Records are immutable and built for audit and inspection.
Deletion.
On offboarding, customer data is exported and then hard-deleted, with a verifiable check that nothing remains. Immutable audit stubs of the event are kept for compliance.
Failure modes bowlyn designs against.
Hallucination.
Mitigated by review on every answer. The system isn't relied on to be right; a person who's run the workflow is the final filter.
Out-of-scope confidence.
The AI's role is constrained by design, and review catches answers outside its competence.
Drift.
Output quality is monitored through ongoing review and periodic testing against known-good examples.
Operational support. Nothing more.
bowlyn's software is operational support. It is not a medical device, not clinical decision support, and not a system of record. It doesn't make safety, dosing, or eligibility calls. It tracks the FDA's 2025 draft guidance on AI in drug development and designs its workflow consistent with the principles there: human oversight, traceability, and a defined role for AI inside a person-driven process. Sponsors remain responsible for their submissions, GCP compliance, and oversight of their programs. bowlyn supports those obligations. It doesn't assume them.
Quality and regulatory teams are welcome to review this in more depth.
Last updated June 2026.