AI Privacy

Avut PIM uses AI only for selected product, import, translation, media, tariff, assistant, and MCP-enabled workflows. The service is designed to send only the context needed for the requested task.

This page supplements Avut’s Privacy Policy, Terms, Data Processing Addendum, and Subprocessors and Third-Party Providers register. It explains how AI-assisted features in Avut PIM may process customer-controlled data, AI-generated content, MCP-related data, and related operational records. This page does not replace the role allocation, transfer terms, retention commitments, or customer-specific terms described in those documents.

How to read this page

For this page:

• Customer-controlled data means data that the customer or its users add to the workspace, including catalog content, product text, images, import samples, prompts, instructions, and other tenant content.
• AI-generated content means model outputs such as translations, descriptions, classifications, alt text, mapping suggestions, assistant replies, and generated visuals.
• System-generated data means job status, approvals, usage counters, timestamps, queue events, and similar service data generated while Avut operates the feature.
• Operational logs means security, access, error, and infrastructure logs used to protect, troubleshoot, and operate the service.

Roles and scope

For customer-controlled workspace, catalog, media, import, export, integration, MCP, and AI workflow data, Avut generally acts as a processor on behalf of the customer under the applicable agreement and DPA. For account administration, billing, service security, abuse prevention, operational monitoring, and certain support or audit records, Avut may act as a controller or independent controller as described in the Privacy Policy.

Customers remain responsible for the accuracy, legality, and rights to the data they enter, import, synchronize, or publish through the service.

Avut-operated AI providers and transfers

When an Avut-operated AI feature requires model processing, Avut may use the AI provider configured by Avut for that feature as part of the Avut service. Avut’s built-in AI features currently use OpenAI through the API for supported translation, inline product content generation, assistant workflows, alt text, image generation, tariff suggestions, and import mapping suggestions.

Where personal data is transferred outside the EU/EEA as part of an Avut-operated AI feature, Avut relies on the transfer safeguards described in the Privacy Policy, DPA, and public provider register, including applicable data processing terms and lawful Chapter V transfer mechanisms where required.

Customer-directed MCP clients

Avut may provide a standards-based remote MCP server through which a customer-authorized external AI or MCP client can use a deliberately limited catalog of Avut tools. Examples of compatible client types may include ChatGPT, Claude, Codex, MCP Inspector, and other standards-compatible clients, but compatibility does not mean Avut selects, controls, endorses, or contracts with every such provider.

When you connect an external MCP client to Avut, the external client is selected and authorized by the customer or user. In those cases, the external provider receives and processes prompts, conversation context, tool requests, and data returned by authorized tool usage under the customer’s own account and provider relationship. External-client output, behavior, availability, security, retention, residency, and training settings remain subject to that provider.

Customers should connect an external client only if their organization permits that provider to process the selected workspace data.

Model training and service improvement

Avut does not intend to use customer-controlled data submitted through Avut-operated AI features to train Avut’s own general AI models. Avut also does not intentionally opt customer-controlled data into a third-party provider’s voluntary general-model training or data-sharing program where such optional settings exist.

OpenAI states that API inputs and outputs are not used to train or improve its models by default unless the API customer explicitly opts into data sharing. Under OpenAI’s default API controls, abuse-monitoring logs for the Chat Completions and image endpoints used by Avut may be retained for up to 30 days.

Provider-side data handling, retention, abuse monitoring, and model-improvement settings nevertheless depend on the provider configuration and contractual terms then in effect, including any approved modified-abuse-monitoring or zero-data-retention controls. Customers should review the provider relationship that applies to their use case, especially where they independently connect an external MCP client.

Storage, outputs, and logs

Depending on the workflow, Avut may store prompts, selected inputs, model outputs, user approvals, rejected suggestions, assistant messages, and limited job metadata in a tenant-scoped manner where needed to display results, preserve workspace history, support troubleshooting, maintain auditability, or enforce security.

Avut does not intentionally keep full prompts or full model responses in separate long-lived operational logs where this is not needed for the service. Operational and security logs may still record limited metadata such as workspace identifiers, user identifiers, job identifiers, provider, timestamps, status, token counts, and error summaries.

Generated images or other generated assets that the customer chooses to keep may be stored as tenant media.

For MCP, Avut stores MCP client and grant metadata, operation and audit records, and opaque token or authorization-code hashes. Avut does not store raw OAuth access tokens or refresh tokens for MCP connections.

Retention

AI-related and MCP-related data is retained for as long as needed to provide the requested feature, preserve workspace history, support customer review and approvals, maintain security and auditability, troubleshoot incidents, meet legal obligations, and complete customer offboarding, deletion, and backup processes.

Where specific operational schedules apply, Avut currently targets or uses the following retention periods:

• finalized AI, translation, import-mapping, and related workflow jobs or events: normally 60 days;
• security and audit records, including MCP-relevant audit records: normally 365 days;
• MCP authorization-code hashes: 1 day;
• expired or revoked MCP token hashes and completed write-operation hashes: 31 days;
• short-lived MCP write previews: 10 minutes;
• temporary privacy or operator export packages: normally 14 days.

Other workspace-level AI content, such as assistant messages, AI approvals, or generated content retained as part of workspace history, may remain available until the customer deletes the relevant data or the workspace is offboarded, subject to backup cycles, legal obligations, and security needs.

Human review and limitations

AI-generated content is provided as draft or decision-support output. It may be inaccurate, incomplete, outdated, biased, or unsuitable for a specific market, tariff classification, channel, language, or legal requirement.

Customers and authorized users must review AI-generated suggestions before applying, publishing, or relying on them. Avut’s AI features are intended to support human workflows and are not designed to make decisions about natural persons that produce legal effects or similarly significant effects without meaningful human review.

Tariff suggestions are decision-support only and must be reviewed before use in customs, compliance, or external filings.

Security and confidentiality

Avut applies technical and organizational measures designed for a business SaaS product handling customer workspace and operational data. Depending on the workflow, such measures may include tenant isolation or equivalent segregation controls, role-based access, permission checks, transport encryption, secure secret handling, audit logging, bounded responses, output minimization, revocation controls, and operational monitoring.

For MCP connections, safeguards include OAuth authorization-code flow with PKCE, workspace-bound authorization, current-role and permission checks, strict schemas, bounded results, output minimization and redaction, revocation, rate limiting, optimistic concurrency, idempotency, and audit logging.

No system can guarantee absolute security, and any connected external client remains subject to its own provider’s security and privacy practices.

Feature categories

Avut’s built-in AI features currently include the following categories:

• Catalog translation: translates tenant product text and may suggest glossary terminology using the context needed for the translation task.
• Inline product content AI: generates, improves, structures, summarizes, or normalizes product catalog content using relevant product fields, field metadata, and user instructions.
• Media AI: generates product alt text and tenant-requested packaging or other image-related visuals using relevant image references, product context, and user prompts.
• Tariff suggestions: suggests tariff classifications from product context, candidate codes, and optional image context.
• API import mapping suggestions: suggests product import mappings and dictionary extraction from sampled external payloads.
• Assistant runtime: answers workspace-context questions, explains workspace data, and suggests product-data fixes or workflow actions.

For Avut-operated AI features, Avut is designed not to intentionally send connector secrets, raw API keys, or unrelated tenant records to the AI provider for the task.

Guarded MCP writes

Read access through MCP is distinct from write access. The standard OAuth scope is mcp:tools. A separate mcp:tools.write scope is required for reviewed write functionality.

When enabled, the current public write catalog is deliberately limited. It currently supports a two-step model for one existing product: a preview step that validates a proposed status change and creates a short-lived preview without changing product data, and a confirm step that applies only the exact previously previewed operation if all then-current controls still allow it.

Confirmation is bound to the exact grant, user, client, workspace, operation, and current product version. If the product changes after preview, confirmation fails without applying the stale operation. Idempotency and replay protections are used to prevent an already completed confirmation from applying the same change twice.

The current public MCP catalog does not provide arbitrary field editing, product creation, product deletion, bulk updates, raw SQL execution, arbitrary URL fetching, administrative debugging functions, or unbounded exports.

Customer responsibilities

Customers should configure AI and MCP features only for authorized users, avoid including personal data that is unnecessary for product information workflows, and review whether any customer-specific legal, contractual, or sector requirements apply before relying on AI-generated output or connecting an external MCP client.

Customers remain responsible for their own legal basis, notices to their end users or staff, and any DPIA or internal approval process required for their particular use of personal data in Avut.

Workspace owners and administrators are responsible for managing connected-client access and revoking connections when no longer needed.

Contact and updates

Questions about AI-assisted processing, MCP-enabled processing, privacy, retention, or deletion can be sent to the contact method listed in Avut’s Privacy Policy.

This page may be updated when Avut changes AI providers, AI features, MCP functionality, retention commitments, transfer safeguards, or other material processing practices.