Secure AI document analysis in regulated industries

Document analysis can reduce the time spent finding, comparing, and compiling information. In finance, research, life sciences, and other regulated industries, it is not enough for the result to seem reasonable. The information must be processed in the right environment, access must be controlled, and every key statement must be verifiable.

The same basic technology can be used in several industries, but the requirements differ. Therefore, the model, deployment, and workflow should be selected based on the classification of the documents to be analyzed, and the severity of the consequences of an error.

What is meant by document analysis?

Document analysis can include:

• searching within and querying a document archive

• extraction of amounts, dates, parties, or other fields

• summarization with source references

• comparison between versions or documents

• classification and prioritization

• drafting reports and decision support materials

Structured extraction and free-form summarization carry different risks. If an amount is to be forwarded to a calculation, strict validation is required. A report draft can allow for more flexibility but must be reviewed before use.

How to connect a secure chat to your organization's own data

Why regulated industries require more control

Documents may contain personal data, trade secrets, security-classified information, or data subject to sector-specific requirements. Incorrect output can affect reporting, decisions, or regulatory compliance.

The organization therefore needs to be able to demonstrate purpose, data flow, access, model version, testing, and human control. In finance, DORA's requirements on ICT risk and third-party dependencies may be relevant.[1] When personal data is processed, GDPR and the Swedish Authority for Privacy Protection's (IMY) guidance apply.[2]

AI Act, NIS2, DORA & GDPR – which requirements affect your company's AI solution?

Information classification and model selection

Classify documents before choosing a model. Ask which environment the information is allowed to leave, which countries and subcontractors are permitted, and what data retention is accepted.

Highly sensitive documents may require a local model hosted in Sweden or in other controlled infrastructure. Less sensitive material can be processed in an external enterprise service with agreed terms. A hybrid architecture can choose the model based on classification.

On-premise deployment provides control but requires self-responsibility for patching, capacity, and incidents. Assess the entire chain, including document indexes, logs, backup, and administrative access.

Local language models or cloud models – how do you choose?

RAG, sources and traceability

RAG can retrieve relevant document sections and provide the model with controlled source material. The user should be able to see exactly which source and version supports the answer.

A source reference is not the same as proof. Test that the source actually supports the statement. For extraction, the result should link to the page or section, and ideally display the original text next to the value.

Log the model and index version, as well as which sources were used. Limit log content and retention according to classification.

What is RAG? How companies use their own data in a language model

Document-level permissions

Users should only be able to search within documents that their role is authorized to read. Permissions must follow the document during indexing and be updated when the document is moved or access changes.

The platform administrator role must be separated from the right to read all documents. Use least privilege for integration accounts and log privileged access.

Test the same query with multiple user profiles. Unauthorized material must not be retrieved into the model's context.

Permission-controlled AI with Microsoft Entra ID – how access works

Testing and human validation

Build test cases with representative documents, tables, scans, attachments, and conflicting versions. Measure extraction per field and source grounding for summaries.

Test for correct rejection when the text is illegible, source material is missing, or the user lacks permission. For critical data, a domain expert should approve the result before it is recorded or published.

NIST's GenAI profile emphasizes documented risk management, measurement, and monitoring.[3] The OWASP risk catalog provides support for security testing of, among other things, manipulated documents and insecure output.[4]

Two different sets of requirements

Example 1: Research and technical reports

An organization may need to search, summarize, and draft reports without sensitive data leaving the controlled environment. With a local model and RAG, documents and analyses can be kept within a closed infrastructure, while the researcher retains control over both sources and conclusions.

Example 2: Regulated financial operations

A securities company may want to analyze documents, generate report drafts, and support cash flow analysis. Requirements for permissions, logging, document versions, and human approval become central. DORA and other sector-specific requirements need to be assessed and managed within the organization's own governance.

The choice between a platform-based and a custom-developed solution should be based on actual requirements, not on a predetermined architecture.

Read more about our customers

Requirements before a pilot

• Document type, owner, and information classification are defined.

• Permitted model and deployment environment are decided.

• Access follows users and documents.

• Extraction and generation have separate quality metrics.

• Sources are displayed at a relevant level of detail.

• The test library contains difficult and prohibited cases.

• Human approval is placed before critical use.

• Logging and retention are proportionate.

• Incident response, backup solution, and exit process are documented.

• Customer-specific results are reported with method, time period, and clear source.

From idea to pilot: what is a PoC, prototype, and pilot?

Sources

1. European Commission, DORA – cyber resilience:
   https://finance.ec.europa.eu/digital-finance/cyber-resilience_en

2. IMY, Guidance on GDPR and AI:
   https://www.imy.se/verksamhet/dataskydd/innovationsportalen/vagledning-om-gdpr-och-ai/

3. NIST, Generative AI Profile:
   https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

4. OWASP, Top 10 for LLM and GenAI Applications:
   https://genai.owasp.org/