Data Privacy & Handling

Manage PII, retention, and where your data flows

Conversation data moves through a clear lifecycle: collection, anonymization, storage, and deletion. You control the two stages that vary most between Workspaces: how personal information (PII) is identified and replaced, and how long conversation data is kept before deletion.

For certifications, the security FAQ, and how Helvia.ai protects your data end-to-end, see the Security Overview page.

Anonymization and PII Handling

Personal information in conversations is automatically detected and replaced before that data is stored or sent to third parties. Anonymization is activated and configured per agent under Designer > Privacy & Security > Anonymization, so each agent can carry rules appropriate to its domain.

These mechanisms work together, each addressing a different sensitivity level:

Entity-Based Detection

Recognize named entities like people, locations, dates, and money, or match custom regex patterns

Custom Anonymization Service

Plug in your own detection endpoint when the built-in entities are not enough

Full Message Obfuscation

Replace the entire user message before storage for the highest-sensitivity scenarios

Entity-Based Anonymization

Entity-based anonymization uses Named Entity Recognition to detect Personally Identifiable Information (PII) in free text and substitute it with placeholder values. The detector recognizes a fixed set of categories, such as names, locations, dates, and monetary values.

Detection runs on every incoming user message when configured, before it enters the agent's processing pipeline, so the data is replaced before it ever reaches the language model or persistent storage. It can also be applied to the full conversation transcript before export to external systems. Contact support for special configurations beyond the defaults.

To add a detection and replacement rule:

1

Open the Agent's Anonymization Settings

Go to Designer > Privacy & Security > Anonymization.

2

Add a Detection Rule

Under Anonymization Settings, select Add Data Type to insert a new row. You can add as many rows as you need, one per entity type or regex pattern you want to detect.

3

Choose What to Detect

Pick the Data Type category. The full set of supported categories is listed below.

4

Set the Replacement

Enter the text that will replace matches in Replacement Text. Leave it blank to fall back to the Default Replacement value set for the section.

5

Provide a Regex Pattern

Only required when Custom Regex is selected as a data type. Enter the Regex pattern and the Replacement Text. Use this for identifiers specific to your domain, such as account numbers or internal case IDs.

6

Save the Rule

Select Save Changes to apply.

The full list of supported entities is:

Entity	What It Covers
PERSON	People, including fictional
GPE	Countries, cities, states
NORP	Nationalities, religious or political groups
FAC	Buildings, airports, highways, bridges
ORG	Companies, agencies, institutions
LOC	Non-GPE locations, mountain ranges, bodies of water
PRODUCT	Objects, vehicles, foods (not services)
EVENT	Named hurricanes, battles, wars, sports events
WORK_OF_ART	Titles of books, songs, and other works
LAW	Named documents made into laws
LANGUAGE	Any named language
DATE	Absolute or relative dates and periods
TIME	Times smaller than a day
PERCENT	Percentages
MONEY	Monetary values, including currency
QUANTITY	Measurements such as weight or distance
ORDINAL	First, second, third, and so on
CARDINAL	Numerals that do not fit another type

Tune detection to your data: Entity recognition is model-based, so unusual names or domain-specific identifiers can slip past the default categories. Add Custom Regex rules for the most important patterns.

Custom Anonymization Service

If the built-in detector does not cover a category specific to your domain, point the agent at your own service instead. In the agent's anonymization settings, enable Service Configuration, supply the endpoint URL, and add any HTTP headers your service requires for authentication.

When to use a custom service: industry-specific identifiers (medical record numbers, account numbers, internal case IDs) or jurisdictions where you need detection beyond the standard entity set.

Full Message Obfuscation

For the highest-sensitivity scenarios, enable Obfuscate User Input to replace the entire user message with an obfuscated string before it is stored or sent downstream. Use this when entity-level redaction is not enough and no portion of the original message should be preserved.

With full obfuscation on, the language model only sees _censored_ in place of the user message and cannot respond to the original content.

Data Retention

The platform retains conversation transcripts for a configurable window and deletes them automatically when that window expires. Data retention has two levels:

Workspace Default

Set under Workspace > Settings > Configuration in the Data Retention field. Required at Workspace creation.

• Range: 1 to 24 months

• New Workspaces default to 3 months.

• Applies to every agent in the Workspace unless overridden

Per-Agent Override

Set under Designer > Privacy & Security > Anonymization in the Data Retention field for an agent, and selectable when creating a new agent.

• Range: 1 to 24 months

• Overrides the Workspace default for this agent

• Select Inherit from Workspace (N months) to drop the per-agent override

When the retention period expires, the data is completely removed from production databases. Archived copies remain in backup storage for up to two years for disaster recovery, unless a shorter window is agreed in your contract.

Retention covers stored conversation transcripts (chat sessions). Audit logs, Knowledge Base content, and Workspace configuration follow separate retention rules.

Data Loss Risk: Setting or shortening retention permanently deletes data older than the new window on the next cleanup run.

Where Your Data Lives

All Helvia.ai platform services and databases operate within the European Union, on managed cloud infrastructure with geographic redundancy for disaster recovery.

EU-only Hosting

All processing and storage happens inside European Union data centers

Cloud-Native

Hosted on AWS under their ISO 27001 and SOC 2 certified programs

Geographic Redundancy

Backups and replicas distributed across availability zones for continuity

Encryption

All customer data is encrypted in transit and at rest. The same encryption applies to backups and to data flowing between internal services.

State of Data	Protection
In transit	TLS/SSL across all internet communications, including traffic to LLM providers
At rest	AES-256 encryption applied at the storage layer
Database-level	Sensitive fields encrypted inside the database so data stays protected even on direct access
Passwords	Hashed with modern algorithms; never visible to administrators
Backups	Encrypted with the same standards as primary storage and held in secure, segregated locations

Data Minimization

The platform collects only what each processing purpose requires and removes data when that purpose ends. These principles apply throughout, from the data your agents receive to the records kept in Observatory.

Purpose-Bound Collection

Each data point is tied to a defined processing purpose and not collected for hypothetical uses

Least-Privilege Access

Role-based access ensures users see only the data their role requires

Routine Review

Stored data is reviewed regularly and removed if no longer necessary

Automatic Deletion

Retention windows enforce removal without relying on manual cleanup

Data Flows to Third Parties

Conversations sometimes need data to leave the platform, whether to generate a response with a language model or to update an external system. Every transfer is protected by the same controls used everywhere else.

Encrypted in transit

All outbound data travels over TLS-encrypted channels, so it stays protected end-to-end between Helvia.ai and the destination.

Anonymization available before export

Personal data can be detected and replaced before any of it is sent to a language model or downstream system, using the same anonymization rules configured per agent.

Vetted providers only

Every third-party provider Helvia.ai integrates with is reviewed against recognized security standards such as ISO 27001 and SOC 2 before being added.

All third-party integrations, LLM providers included, run through your own provider account and credentials. This means two things:

• The relationship is governed directly by your contract with that provider, including any data-use and training terms

• The processing region follows the credentials you supply, which can be configured to be inside or outside the EU

Best Practices

• Configure anonymization per agent: Match the rules to the data each agent actually handles, rather than applying one set across every agent

• Use a custom service for domain identifiers: The built-in entities are broad; plug in your own service when you need medical IDs, account numbers, or other domain-specific patterns

• Tune data retention per agent: Adjust the per-agent retention for agents handling more sensitive conversations according to your contractual obligations

• Audit which provider your LLM calls use: Customer-owned accounts give you direct control over training opt-outs and data-use terms

You now know where your data lives, how it is anonymized and retained, and what reaches third parties at each step of the lifecycle.