AI and Data Processing Disclosure

Last Updated: 2026-08-10

Effective Date: 2026-08-10 for users first accepting the MainBook Terms on or after that date; 2026-09-10 for users who accepted an earlier version.

This AI and Data Processing Disclosure ("Disclosure") explains how the MainBook service ("Service") provided by Human Beyond LLC ("MainBook", "we", "us", "our") uses artificial intelligence ("AI") and machine-learning ("ML") technology to process documents you upload, and the specific commitments and limitations that apply.

This Disclosure is incorporated by reference into our Terms of Service, Privacy Policy, and Data Processing Agreement. It supplements but does not replace those documents. Capitalized terms not defined here have the meanings given in the Terms of Service.

1. Summary

MainBook uses optical character recognition ("OCR") and large-language-model ("LLM") technology, provided by third-party AI vendors, to convert your bank or credit card statement PDFs into structured data formats.
The current AI sub-processors are listed in our Sub-Processors page, which is updated as our sub-processor mix changes.
We do not use your Content, your Output, or extracted financial data to train or improve our own AI models.
We engage AI sub-processors that, according to their published terms applicable to commercial or API customers, are configured for no-training defaults or are contractually committed not to use Customer data to train their models, in each case where such configuration or commitment is available.
AI is probabilistic. Output may contain errors, omissions, or hallucinated data. You must independently verify Output against the source documents before relying on it. See our Disclaimer for the full warning.

2. How the Service Uses AI

When you upload a document, the Service performs the following steps:

2.1 Receipt and storage. Your browser uploads the document over an encrypted TLS connection directly to our object storage (DigitalOcean Spaces) using a time-limited upload authorization issued by MainBook. MainBook then retrieves the document from that storage for processing. The document file is encrypted at rest.

2.2 Text extraction and Optical Character Recognition (OCR). Where a PDF contains a usable digital text layer, MainBook may parse that text within our own infrastructure. For scanned PDFs, images, and certain quality-control re-reads, the relevant source PDF, image, or rendered page image is transmitted to our OCR sub-processor (Mistral AI, as of the Last Updated date above), which returns text and a layout-aware structured representation.

2.3 Structured extraction (LLM). Extracted text and related structured context are transmitted directly to our large-language-model ("LLM") sub-processor (Google, via the Gemini API, as of the Last Updated date above). In a limited final quality-control fallback, MainBook may also send Google a rendered PNG image of one source-document page. The LLM identifies transaction rows, extracts dates, descriptions, amounts, balances, and other relevant fields, and returns structured data. We do not transmit the original PDF file itself to Google. The specific LLM model used is selected based on runtime configuration and may change over time without notice (we always reserve the right to choose the model best suited to a given document). We access the LLM provider on a paid (billed) basis.

2.4 Validation and quality control. The Service applies automated mathematical and logical checks (for example, comparing the sum of credits and debits against opening and closing balances) and may re-run extraction on documents that fail these checks. Any validation status, flag, score, warning, or similar indicator is only an automated signal. It is not an audit, certification, or guarantee that the Output is accurate, complete, reconciled, or compliant with accounting, tax, or legal requirements.

2.5 Storage of Output. The structured Output is stored in our application database and made available to you for review, editing, and export.

2.6 Auto-deletion and deletion requests. Source documents and Output uploaded through an authenticated Account or the Developer API are scheduled for automatic deletion ninety (90) days after upload; guest uploads are scheduled for deletion after twenty-four (24) hours. Our active-system deletion sweep runs daily and also covers abandoned uploads that never complete submission. Where the Developer API permits a deletion request for a completed job, the job becomes unavailable through the Service immediately and its active-system physical purge is scheduled for the next sweep, normally within twenty-four (24) hours; a failed storage deletion may be retried on a later sweep. Encrypted database backups may retain a recoverable copy for up to seven (7) additional days, isolated from normal processing, and third-party providers may retain limited abuse, security, or legal-compliance logs under their published terms. See our Privacy Policy.

3. AI Sub-Processors

The current list of AI sub-processors is set out in our Sub-Processors page, which is the single source of truth for vendor identity and is updated as our vendor mix changes. As of the Last Updated date above, the primary AI sub-processors are:

Mistral AI SAS (France). OCR provider. For scanned PDFs, images, and certain quality-control re-reads, processes the relevant source PDF, image, or rendered page image and returns a text and layout representation. Digital PDFs with a usable text layer may instead be parsed within MainBook's infrastructure. Subject to Mistral's then-current published terms applicable to its API customers.

Google, via the Gemini API. LLM provider. Receives extracted text and related structured context and returns the structured extraction (transaction rows, dates, descriptions, amounts, balances). In a limited final quality-control fallback, Google may also receive a rendered PNG image of one source-document page. We access the Gemini API directly, on a paid (billed) basis, and are therefore subject to the terms Google applies to paid Gemini API access. We do not transmit the original PDF file itself to Google. Google may process data in countries where Google or its agents maintain facilities, subject to the transfer safeguards described in our DPA.

If the AI sub-processors listed here change (for example, if we replace Mistral with a different OCR provider, or switch to a different LLM provider), this Disclosure will be updated and the change will be reflected in the Sub-Processors page with at least thirty (30) days' advance notice. Material changes (for example, addition of a new AI sub-processor or a change in the category of underlying providers) trigger our standard sub-processor notice mechanism described in Section 5 of our Data Processing Agreement.

4. Our "No Training" Position

4.1 We do not train our own models. MainBook does not maintain or train its own AI or ML models. We do not use your Content, your Output, or any data derived from your use of the Service to train, fine-tune, or improve any AI or ML model owned, developed, or controlled by us or by our affiliates.

4.2 Sub-processor training controls. We use commercial API configurations intended to prevent Customer data from being used for model training, subject to each provider's then-current terms and the qualifications below:

Mistral AI (OCR). We configure the applicable Mistral organization or account not to permit training on Customer data where that control is available, and we do not intentionally route Customer data through Mistral Labs or Preview offerings. Mistral's terms may permit training where a customer opts in, provides feedback, agrees otherwise in an order form, or uses a Labs or Preview offering.
Google (Gemini API, LLM). Google's terms for paid Gemini API access provide that prompts, files, and responses are not used to improve Google's products. This protection depends on the production API key accessing a Google Cloud project with active billing.

These statements reflect the vendors' then-current published terms and the configuration we maintain for the Service; we link to authoritative pages from our Sub-Processors list and through the vendors' own legal pages. We periodically review the relevant billing and data-use settings, but we cannot independently audit a provider's internal systems.

4.3 Limits of our position. These no-training protections are the vendors' contractual and published representations, on which we rely, and a vendor may change its terms or controls over time. Separately, a vendor may retain inputs and outputs for abuse monitoring, security, and legal compliance. Under Mistral's standard API policy, inputs and outputs may be retained for a rolling period of up to thirty (30) days unless an approved zero-data-retention configuration applies. Google's paid-service terms permit prompts and responses to be logged for a limited period for abuse detection and legal compliance unless an approved zero-data-retention configuration applies. Such limited retention is distinct from model training. If a vendor materially changes its data-use terms, or if we change AI sub-processors, we will update this Disclosure as described in Sections 3 and 9. If you require a stricter or independently contracted guarantee — for example, a zero-data-retention configuration or a direct contractual no-training commitment from a specific provider — please contact us at hello@human-beyond.ai and we will discuss whether your requirements can be accommodated under your specific configuration.

4.4 Format observations and de-identified telemetry. During the retention period for a source document and its Output, we may keep pseudonymous format observations linked to that job and an opaque Account or guest identifier. Those observations are deleted with the underlying job. We may retain a de-identified format profile containing normalized institution name, document type, language, currency, column-role and locale signatures, and aggregate outcome counts. A retained format profile does not contain names, account numbers, transaction descriptions, transaction amounts, balances, source files, or row-level Output. We also collect aggregated or de-identified operational telemetry such as processing duration, error rates, confidence scores, and file-size distributions. We use this data to operate, secure, and improve the Service, not to train AI models.

5. Probabilistic Nature of AI — Risk of Errors

5.1 No guarantee of accuracy. Output from the Service is generated by probabilistic AI models. Output may not be accurate, complete, reliable, or fit for any particular purpose. Errors that AI-based extraction may produce include but are not limited to:

(a) misreading numerical values (mistaking "1" for "7", transposing digits, dropping or duplicating digits);

(b) misreading or inverting dates;

(d) omitting transactions from the Output;

(e) duplicating transactions in the Output;

(f) misattributing transactions across accounts;

(g) producing fabricated or "hallucinated" entries that do not appear in the source document;

(h) misinterpreting the layout, structure, or context of the source document;

(i) producing different Output for the same input on repeated runs.

5.2 Your obligation to verify. You must independently verify the Output against the original source documents before using, sharing, or relying on it. The Service is intended to assist you; it is not a replacement for human review. See Section 8 of the Terms of Service and Section 3 of the Disclaimer for the full obligation.

5.3 High-stakes decisions. You agree not to use the Output, without independent human review by a licensed or qualified professional, for any decision in the following sensitive areas: financial activities and credit; insurance; legal; medical; employment, housing, or education; essential government services; product safety; national security; migration; or law enforcement. See Section 8.3 of the Terms of Service and Section 4 of the Disclaimer.

6. Data Flow and Data Residency

6.1 Data flow. When you upload a document, Customer Personal Data (as defined in the DPA) flows from your browser directly to DigitalOcean Spaces object storage. MainBook retrieves the document and either parses a usable digital text layer within our infrastructure or conditionally sends the relevant PDF, image, or rendered page image to Mistral for OCR. Extracted text and structured context, and in a limited fallback one rendered source-page image, are then sent directly to Google through the Gemini API. The source document and Output are stored in our object storage and application database as described above.

6.2 Cross-border processing. Some sub-processors are located in countries other than the United States (for example, our OCR sub-processor Mistral AI is based in France). Some sub-processors may use sub-processors of their own in additional locations. The locations applicable as of the Last Updated date above are listed in our Sub-Processors page.

6.3 Transfer mechanisms. Cross-border transfers of personal data from the EEA, the UK, or Switzerland to other jurisdictions are made on the basis of the transfer mechanisms set out in Section 6 of our DPA (Standard Contractual Clauses, UK Addendum, and equivalent mechanisms).

7. Output Ownership and License

7.1 You own the Output. As between you and us, you own the Output generated from your Content, as set out in Section 9 of the Terms of Service.

7.2 Use of Output. Output may not be used in any manner that violates the Acceptable Use Policy, including but not limited to use for high-stakes automated decisions without human review, or use to train, develop, or improve any competing AI product, model, or service.

7.3 Similarity of Output. Due to the nature of AI, Output may not be unique. Other users may receive similar Output from similar inputs. Our assignment of Output rights does not extend to other users' Output.

8. Human Review

We may, on a limited basis and only as reasonably necessary to operate, secure, and improve the Service, have human personnel review specific items of Output or specific Customer documents — for example, to investigate a reported error, to triage a security or abuse incident, or to debug a processing failure. Such review is performed by personnel bound by confidentiality obligations, is logged, and is conducted under the principle of least privilege. We do not engage in routine human review of Customer documents or Output as part of the conversion pipeline; the conversion pipeline is automated.

9. Updates to This Disclosure

This Disclosure may be updated to reflect changes in our AI sub-processors, in their published terms, or in our own practices. Material changes (for example, addition of a new AI sub-processor or a change in the category of underlying LLM providers) are subject to the thirty (30) day notice mechanism described in Section 5 of our DPA. Non-material changes (for example, formatting, clarifications) take effect upon posting.

10. Contact

For questions about AI processing in the Service, or to request additional information about a specific sub-processor or underlying provider:

Human Beyond LLC Email: hello@human-beyond.ai