AI Transparency Statement

Effective date: August 5, 2026 · Last revised: August 5, 2026 · Version 2026.08.05

This Artificial Intelligence Transparency Statement (the “Statement”) is provided by the operator of Tavi (“Tavi,” “we,” “us,” or “our”) to describe, in a reasonably detailed manner, the categories of machine-learning and related automated systems used to operate the Tavi WhatsApp memory assistance service and associated web surfaces (collectively, the “Service”), the purposes for which customer Content (as defined in our Privacy Policy) is submitted to such systems, and our posture with respect to model training. This Statement is informational and descriptive; it does not modify the Terms of Use or Privacy Policy except where expressly incorporated by reference therein. The specific vendors, model identifiers, and routing paths we use are confidential operational details and may change without prior individual notice as we optimize latency, quality, cost, and availability; material changes to our training posture will be reflected in an updated revision of this Statement and, where required, in the Privacy Policy.

1. Executive summary

Tavi uses third-party artificial intelligence and related automated systems—including large language models, embedding models, multimodal / vision models, and speech-to-text models—to understand messages you send, organize and retrieve what you have saved, read text in images you upload, transcribe voice notes, and compose answers grounded in your own Content. Customer data submitted in connection with these processes is used to deliver the functionality of the Service to you. Tavi does not use your Content to train, fine-tune, or otherwise improve general-purpose foundation models for Tavi or for unrelated third parties. Inference may still be subject to the data-handling terms of the upstream AI subprocessors through which requests are routed; those subprocessors’ policies govern their independent Processing. For competitive and security reasons, we do not publicly disclose the specific model names or AI vendor identities in use at any given time.

2. Scope of automated processing

Automated systems may be invoked, without a separate per-request confirmation dialog in WhatsApp, when reasonably required to:

• Classify message intent and route requests to the appropriate Service workflow (save, search, remind, manage, FAQ, and related flows);
• Extract structured fields from free-form notes (for example, dates, entities, or reminder parameters);
• Generate vector embeddings for hybrid retrieval over your stored memories;
• Plan and execute search over your library and compose natural-language answers citing or summarizing retrieved items;
• Perform optical character recognition and image understanding on photos and screenshots you send;
• Transcribe voice notes into text for storage and later retrieval; and
• Support limited product-help drafting grounded in Tavi’s own knowledge materials (not open-web general chat).

The Service is intentionally scoped as a memory assistant for Content you elect to store. It is not offered as a general-purpose chatbot for arbitrary world-knowledge queries unrelated to your library.

3. Categories of AI capabilities

Without disclosing proprietary stack details, the Service may employ one or more of the following capability categories, each selected and substituted by Tavi as operationally appropriate:

You do not select models per message. Tavi selects and may change the underlying systems in order to maintain Service quality and operational constraints. Where an automated call fails, heuristic or fallback logic may apply for certain flows.

4. Third-party inference subprocessors

To obtain inference outputs, Content (or derivatives thereof) may be transmitted to one or more unaffiliated AI inference subprocessors, including model hosts, inference gateways, and speech-transcription processors, under contractual arrangements with Tavi. Hosting of stored Content, account records, and embeddings remains on our primary cloud infrastructure, separate from ephemeral inference requests except to the extent Content must be transmitted to obtain an inference result.

Message transport continues to involve Meta / WhatsApp and our messaging connectivity provider. Payment and export-email subprocessors are described in the Privacy Policy and are not AI model hosts, except insofar as they Process account or export artifacts you request. The identities of AI model providers and gateways are omitted from this public Statement; a confidential subprocessors list may be made available to regulators or enterprise counterparties upon lawful request where required.

5. Training, fine-tuning, and data use

Customer Content processed through the AI features of the Service is processed for the purpose of delivering that functionality to you (inference, indexing, retrieval, transcription, and related Service operations). Tavi does not use customer Content to train, fine-tune, distill, or otherwise improve general-purpose foundation models. Tavi does not sell customer Content to model providers for advertising or for unrestricted training corpora.

Notwithstanding the foregoing: (a) Content must be transmitted to subprocessors to obtain inference outputs, and those subprocessors may retain limited logs or apply safety filters in accordance with their own terms, documentation, and enterprise / API data-use commitments in effect at the time of the request; (b) Tavi may Process aggregated or de-identified operational metrics (for example, error rates, latency, or usage counters) that do not include your note bodies, for reliability and cost management; and (c) where a subprocessor offers API or business pathways whose documented posture excludes provider training on customer inputs, Tavi’s configuration objective is to use such pathways—subject always to the subprocessor’s then-current contractual terms. You should review the privacy and data-use policies of any AI vendor whose identity becomes known to you through other means for their independent practices.

Derived artifacts that remain within your account—such as embeddings, transcripts, OCR text, and extracted fields—are stored as part of your library to operate the Service for you, and are subject to export and deletion controls described in the Privacy Policy. Those artifacts are not contributed by Tavi to public model-training datasets.

6. Human oversight and limitations

Authorized Tavi personnel may access operational tooling to debug delivery failures, investigate abuse, or fulfill support and deletion requests, as described in the Privacy Policy. Routine message handling is automated. Model outputs may hallucinate, omit relevant memories, or mis-parse dates and entities; critical decisions should not rely solely on automated output. The Service does not provide professional advice.

7. User controls and alternatives

You control what Content enters the Service by choosing what to send on WhatsApp. You may delete individual items or your entire account, and you may export a copy of your library, via Account and in-chat management flows. Refraining from sending a category of Content (for example, voice notes or images) avoids the corresponding transcription or vision Processing. Disabling or abandoning the Service, and deleting your account, ceases ongoing inference on new Content, subject to residual retention windows in backups and subprocessor logs as disclosed in the Privacy Policy.

8. Updates to this Statement

We may revise this Statement to reflect changes in capability categories, subprocessors, or regulatory expectations. The revision date above will be updated when changes are published. For Processing practices of broader applicability (retention, rights, international transfers), the Privacy Policy controls in the event of a conflict with descriptive language in this Statement.