Custom AI data masking, deployed in your VPC.
We adapt sensitive-data detection and masking to your business data, languages and AI workflow, then install and validate the solution in your environment.
Original business data
Detect · mask · apply policy. Original values, mappings and credentials stay here.
You select the provider, model and region.
Runs in your environment
Protection is installed inside your VPC.
You keep the keys
Model credentials remain under your control.
No production-data upload
Herkus does not receive your production data.
Built for the data your business actually uses.
Standard PII is only the starting point. We configure the protection around the language, terminology, identifiers and exceptions in one real application.
Protect your own entities
Customer references, contracts, suppliers, prices, projects and other identifiers specific to your workflow.
Adapt to your terminology
Languages, domain terms and data formats are customized around what your application really receives.
Choose the protection behavior
Define what should be masked, replaced, allowed or, where the workflow requires it, restored.
Protect before AI. Check before restoration.
The model receives policy-transformed content, not the original sensitive values. If restoration is needed, the response is checked inside your controlled path first.
Protect
Detect and mask sensitive values inside your VPC.
Process
Send only the protected content to your chosen AI.
Check
Evaluate the returned result against the active policy.
Restore
Restore approved values only when the workflow requires it.
Protect the data without losing the task.
Before
Customer Aurelia GmbH called from +370 612 34 567 about order 48213. Summarize the ticket and suggest next steps.
Sent to your chosen AI
Customer [CUSTOMER_1] called from [PHONE_1] about order 48213. Summarize the ticket and suggest next steps.
Before
Met Jonas Petraitis (jonas.p@example.lt) about the Kaunas warehouse contract. Draft a follow-up email.
Sent to your chosen AI
Met [PERSON_1] ([EMAIL_1]) about the [PROJECT_1] contract. Draft a follow-up email.
Before
Project Nemunas margin is 18% with supplier Baltic Steel UAB. Translate this paragraph into English.
Sent to your chosen AI
Project [PROJECT_1] margin is [AMOUNT_1] with supplier [COMPANY_1]. Translate this paragraph into English.
Illustrative examples. Which entities are protected is defined with you for your application.
A working protection path, not another platform to configure.
Start with one representative application. Herkus handles the detailed configuration, VPC installation and validation with your team.
01 · Assess
Review the workflow, its data boundary and selected AI provider.
02 · Configure
Adapt the languages, business entities and protection rules.
03 · Deploy and validate
Install and test the complete path inside your VPC.
04 · Maintain
Refine the protection as your data and workflow evolve.
- Your sensitive valuesOriginal production values are not sent to Herkus.
- Your credentialsModel and application keys remain customer-held.
- Your mappings and evidenceProtection state and audit records stay in your environment.
- Your AI providerYou choose the provider, model and processing region.
What is AI data masking?
AI data masking identifies and replaces sensitive values before content is sent to an external AI model. It preserves the context required for the task, keeps the original values inside the customer environment and restores only approved values after the result is validated.
What AI data masking protects
The protected data is not limited to standard personally identifiable information. Depending on the application, it can include names, email addresses, phone numbers, account references, customer identifiers, contract values, project names, internal product information and combinations of fields that reveal confidential business context.
A useful masking policy therefore starts with the application and its data. A generic list of patterns is only the first layer.
How the protected path works
- Discover: identify sensitive values and business-specific data in the application request.
- Mask: replace those values with controlled substitutes before the external LLM call.
- Process: send only the masked content to the selected model or provider.
- Validate: check that the returned result belongs to the active request and satisfies the restoration policy.
- Restore: return eligible original values inside the customer-controlled application path.
If validation fails, the protected values should not be restored.
Masking, redaction and pseudonymisation
Redaction removes or hides information. Masking replaces sensitive values with representations that can preserve the structure of the task. Pseudonymisation uses substitutes that may remain consistent across an approved scope. A real implementation can combine these approaches according to the data type and application requirement.
Why the implementation is customer-specific
Over-masking can make an AI result unusable. Under-masking can expose information that should never leave the application boundary. Consistent references, exceptions, validation, restoration and logging all depend on the workflow. Herkus handles that detailed work: rule design, integration, testing and refinement around one real application.
Frequently asked questions
Does AI data masking delete the original data?
Not necessarily. Masking replaces sensitive values in the content sent to the model. The original values can remain in the customer environment and may be restored only when the returned result passes the configured validation and policy checks.
Can an LLM still complete a task after data is masked?
Often yes, when substitutes preserve the references and context required for the task. This is why masking should be tested against the real application rather than treated as a generic filter.
Is AI data masking tied to one model provider?
The protection boundary can be designed independently of the selected model. The actual integration and validation rules still need to match the customer application and provider interface.
Related guides: PII masking before an LLM call · customer-side AI data masking
See how custom masking fits your application.
Tell us what the application does and which types of data it handles. We will define the customization scope and propose a VPC PoC.
Please do not send production data, credentials or sensitive samples by email.