AICA's Security Architecture: What Your IT Team Needs to Know
The technical details behind how AICA handles your data (written for the people who have to approve it).
If you’ve been following our AICA series so far, you already know what the platform does, you’ve read through our FAQ, and you’ve seen how quality teams can put it to work across internal audits, supplier audits, and inspection readiness.
Today’s piece is for a different audience (or maybe more accurately, the same audience’s next conversation).
Because here’s what often happens in practice when a quality leader decides to evaluate a tool like AICA: they bring it to IT. And IT often has questions about where data lives, how it moves, who can access it, how it interacts with AI models, and what deployment options look like.
We’ve been fielding these questions since launching earlier this month, and we want to answer them publicly so you have something concrete to share internally. No hand-waving or vague assurances here. Just actual answers.
Let’s walk through each one. As always, if you or others have questions we don’t cover here, head over to aica.thefdagroup.com to ask.
Where your data lives in AICA
AICA is hosted on Microsoft Azure servers located in the United States. All application databases and storage services are hosted in the Azure region selected for the deployment, so your information remains physically within U.S. data centers. Azure does not move data outside the country unless explicitly configured for cross-region replication, and AICA’s default configuration does not enable that.
This is a straightforward point, but it matters a lot for many teams, especially those with stringent data protection concerns. For organizations subject to data residency requirements or internal policies about where regulated data can be stored, the answer is clear: your data stays in the U.S., in Azure, under defined access controls.
How your data is protected in transit and at rest
AICA uses a multi-layered security architecture. Here’s what that means in specific terms.
In transit: All exposed connections use TLS encryption. Every data transmission between your browser and the AICA platform, and between AICA’s internal components, is encrypted in transit.
At rest: Sensitive information stored within the platform (your uploaded documents, analysis reports, and associated data) is protected by the built-in encryption mechanisms of Azure’s cloud storage services. Documents are stored in isolated, access-controlled environments.
Network architecture: The platform minimizes its attack surface through reverse proxy implementation and API management. Cloud resources are network-isolated via firewalls and private connectivity, meaning the infrastructure isn’t broadly exposed to the public internet.
Authentication and access controls
AICA uses mandatory token-based authentication built on JWT (JSON Web Token) standards, integrated with OpenID Connect-compliant identity providers. In practical terms, this means the platform plugs into standard enterprise identity infrastructure.
Access within the platform is managed through centralized role-based access control (RBAC) with granular permission management. Your organization’s admin defines roles and permissions during onboarding (like who can upload documents, who can review observations, and who can export reports), and group-based access controls provide precise resource authorization.
All access is encrypted, logged, and fully auditable within the secure Azure environment.
Who can see your documents
This is probably the single most common question we hear from IT and quality teams. The short answer: essentially no one, unless you have a problem and ask for help.
Your audits and uploaded documents are not viewed by humans. No one at The FDA Group or EPAM, our development partner, personally reviews your documents unless you explicitly opt in to share them for model improvement through an optional setting.
A very small number of authorized FDA Group employees may access data only when needed for two specific scenarios: addressing platform security issues, or resolving technical problems you report. This access is tightly controlled, audited, and restricted under strict privacy and security policies.
How AICA interacts with AI models, and what happens to your data
This is where IT teams tend to have the most nuanced questions, and rightly so.
Here’s how it works:
AICA functions as an orchestration layer. It takes your uploaded QMS documents and the regulatory requirements you’ve selected, structures the analysis, and routes the work through cloud-based large language models (specifically Azure OpenAI) using The FDA Group’s designated cloud AI subscriptions.
The key question everyone asks: does my data get used to train the AI model? No, not directly. AICA retains conversational data and prompts in system logs exclusively for governance and compliance purposes. Your documents are not incorporated into model training datasets. When interfacing with the third-party LLM, Azure OpenAI's processing, retention, and use of data are governed by the service agreements between The FDA Group and Microsoft, which, under Microsoft’s enterprise AI terms, means your data is not used to improve or train foundation models.
All log data is stored within Azure Blob Storage provisioned within The FDA Group’s cloud infrastructure. Access controls and security configurations for that storage remain under the exclusive administration of The FDA Group’s IT department.
A little more context on how we “train” AICA:
AICA is not a general-purpose AI that learns from everything it processes (like ChatGPT or any other "off the shelf” LLM). The platform was built on a foundation of curated biopharma QMS documents provided specifically for that purpose by regulatory subject matter experts.
When you upload documents to AICA, they are analyzed and a report is generated. The system does not automatically absorb, retain, or learn from your documents as part of that process. Your data is stored in isolated, encrypted environments on Microsoft Azure’s U.S.-based infrastructure, and it is never exposed to other customers or made available in any public domain.
Our legal agreements do include provisions that allow The FDA Group to use customer data to improve AICA’s capabilities over time. However, this is not something that happens passively or automatically. AICA’s architecture requires direct human involvement from qualified subject matter experts to review outputs and provide structured feedback before any model improvements can take place. In practice, this means any use of customer data for training would involve a deliberate, resourced engagement, not a behind-the-scenes process.
If your organization prefers that your data not be used for model improvement under any circumstances, you can opt out, and that preference will be honored. We recommend discussing your specific requirements with us during the agreement phase if this is a concern.
Data privacy controls
AICA includes several configurable features designed to support your organization’s data privacy requirements:
Document retention is entirely in your control. Documents can be deleted immediately after AICA completes its analysis if you prefer not to store them in the platform. You can also configure automated deletion policies based on your organization’s data retention requirements (for example, automatically purging conversation logs and sensitive files after a set interval).
Analysis reports don’t contain your source documents. The reports AICA generates include references to your documents (document names, specific sections), along with the gap analysis and recommendations. But the actual QMS documents themselves are not embedded in the reports. You can export reports at any time and manage them according to your internal document control procedures.
An interceptor framework examines communications bidirectionally — both user prompts going to the LLM and responses coming back. This lets us implement custom validation logic for detecting and handling sensitive information or personally identifiable information (PII), and to modify or block requests containing prohibited content. These capabilities help minimize and control the lifetime of PII within the platform.
The collection of conversation logs is configurable. If your organization’s policies require it, you can disable conversation logging entirely.
What about hallucinations and content safety?
These are fair questions for any platform built on large language models. Let’s walk through it:
AICA functions as an orchestration layer — it structures and directs the analysis, but does not incorporate its own generative AI logic. What that means in practice: AICA’s value comes from how it structures the audit task, how it maps your documents against specific regulatory requirements, and how it presents findings for human review. The platform is not generating free-form creative output.
That said, the limitations inherent to LLMs — including the potential for inaccuracies — are why AICA is designed with human-in-the-loop as a non-negotiable requirement. Every observation the platform generates must be reviewed and approved by a qualified quality professional before the report is finalized. The platform is explicit about this: AI-generated findings are preliminary until a human validates them.
Responsible AI measures, content filtering, and safety controls operate at the LLM layer through Azure OpenAI’s built-in capabilities, which are governed by Microsoft’s responsible AI policies.
Deployment options
AICA’s default deployment is cloud-based SaaS, meaning you access it through your web browser, no local installation required. This works for most organizations and gets you from contract to first audit in one to two weeks.
For organizations with more stringent security requirements around the cloud, AICA can be deployed on your own private cloud infrastructure. In this configuration, the platform runs within your cloud perimeter. Your data never leaves your controlled environment, and your IT team manages access at the infrastructure level.
Private cloud deployment is a more involved implementation that we handle on a case-by-case basis, and several enterprise firms have already expressed interest. If this is a requirement for your organization, it’s worth raising early in the conversation so we can plan the deployment accordingly.
Technical requirements at a glance
For the standard cloud deployment, the technical requirements on your end are minimal:
Supported browsers: Google Chrome (latest version), Microsoft Edge (latest version), and Safari (latest version) are fully supported. Firefox and Internet Explorer are not currently supported.
Supported devices: Desktop and laptop computers are fully optimized. Tablets have limited support in the current release phase. Mobile devices are not supported in the current release.
File formats accepted: PDF files (including scanned documents via OCR) and Microsoft Word documents (.doc and .docx). The maximum file size per document is 512 MB.
Concurrent users: Multiple users can access AICA simultaneously, with role-based permissions defined by your organization’s admin during onboarding.
No specialized technical expertise, local software installation, or IT configuration is required on the end-user side.
Intellectual property and liability
Two more questions that come up during IT and legal review:
Who owns the content AICA generates? Intellectual property rights and copyright ownership for any content generated through the platform are governed by the terms and conditions in the service agreement between your organization and The FDA Group. AICA functions as an intermediary tool; it doesn’t independently claim ownership of outputs.
Who is responsible for AI-generated content accuracy? Liability considerations regarding content accuracy and appropriateness are determined by the contractual arrangements between your organization and The FDA Group as specified in the applicable service agreements. This is also why human review of all findings is built into the platform’s workflow: it ensures that your qualified quality professionals, not the AI alone, are responsible for the conclusions in your final audit reports.
A note on what’s still evolving
We want to be transparent: AICA is a newly launched platform, and some of the technical details covered here will continue to develop. Enhanced concurrent user scaling, additional eQMS integration capabilities, and expanded deployment configurations are all on the roadmap. As these evolve, we’ll update our technical documentation and let customers know what’s changed.
If your IT team has questions that aren’t addressed here, we want to hear them. Every technical question we field helps us build better documentation and a better product. Reach out through our contact form.
What’s next in this series
We've now covered the platform overview, the FAQ, practical use cases, and the security and IT deep dive. Coming up next:
The benchmarking data behind AICA: A closer look at our internal auditor consistency study — what the overlap numbers mean and what they reveal about the state of traditional auditing.
The regulatory landscape for AI in compliance: Where AI tools fit in a GxP environment, how to think about validation, and what the industry conversation around AI-assisted auditing looks like right now.
Have questions or want to get a demo?
If you've been following this series and want to see the platform for yourself, request a demo at aica.thefdagroup.com. We’re currently in conversation with a number of firms with onboarding starting soon. Talk to us!
What is AICA?
AICA (the Audit Intelligence Compliance Assistant) is an AI-powered compliance auditing platform purpose-built for pharmaceutical and biologic companies.
Developed by The FDA Group in partnership with EPAM Systems, AICA is trained by 15+ FDA and ISO regulatory experts on real-world pharma QMS documentation. Upload your documents, select which regulations to audit against, and get a comprehensive gap analysis of your entire QMS in hours, not days or weeks. Every finding goes through human review before the final report.
See it in action:
AICA currently supports 21 CFR Parts 11, 211, and 600, with additional regulations in active development.
Learn more, read our FAQs, or request a demo at aica.thefdagroup.com.