AI Processing Notice

This notice explains how FinPera uses AI on the data in your workspace - for example when it reads, indexes and classifies what you upload, when you ask FinPera AI a question, or when you generate a draft. The terms that govern that processing - that it is part of the Service, what the controls do, the commitment on training, the limits on relying on AI output, and your confirmation that you hold the permissions needed for information about other people - are in the AI Processing section of the Terms of Service, and are not repeated here. This notice also complements our Privacy Policy and the Data Processing Agreement.

FinPera AI reads documents to extract key fields, classify the document type, surface potential risks and tasks, and suggest where to file them. It is designed to ground its output in the supplied material and flag gaps; it can still be wrong.

We select commercial or enterprise AI configurations where available. What a provider then does on its own systems is governed by that provider's terms. Retention periods, logging, abuse monitoring, safety review and training exclusions are set by each provider, vary by model, endpoint, region and configuration, and change from time to time. A provider may retain prompts, outputs, logs or metadata for service operation, security, abuse monitoring or legal compliance. The provider terms linked from our sub-processor page give the current position.

Financial calculations are designed to be performed in deterministic code rather than by a language model: the AI explains and drafts, and the numbers are computed deterministically. You should still review any figure before relying on it.

Where a value sits in a field we have classified as sensitive, FinPera applies redaction or masking on a best-efforts basis before text is sent to a model, depending on the feature, the data type and your organisation's configuration. These controls may cover identifiers such as PAN, Aadhaar, GSTIN, card and account numbers. This is why the Terms ask you to keep sensitive identifiers in the fields designated for them.

Some organisations can additionally turn on reversible anonymisation. Where it is enabled, FinPera attempts to substitute certain names - people and organisations - with stand-ins before the text leaves us, and to put the real names back in the answer returned to you, so the substitution is invisible in normal use.

It is a risk-reduction measure, not a guarantee, and its coverage is limited by design. It works by recognising entities that already exist in your own records - for example a team member, investor, client or contact you have stored with us - plus pattern-matching for some email addresses and phone numbers. It follows that:

• it is off unless your organisation has enabled it, and it does not apply to every feature or every request;
• a name, entity or identifier that is not already in your records will generally not be recognised, and will be sent as it is;
• it does not attempt to find every name in a body of text, so text that mentions people or organisations you have not stored with us will pass through unchanged;
• it does not detect every category of sensitive information, and detection can be incomplete, inconsistent between requests, or wrong.

We are continuing to improve it. We give no assurance that any particular name, identifier or item will be detected or substituted on any given request, and you should not rely on it as a substitute for deciding what to submit.

Before text is sent to an AI provider, FinPera looks for certain US tax and payment identifiers - Social Security numbers, individual taxpayer identification numbers, employer identification numbers, and labelled bank account, routing and card numbers - and replaces any it recognises with a fixed placeholder that cannot be turned back into the value. This applies to text, including text read from a document with a text layer. Detection is pattern-based: an identifier written in an unusual form, or without the label some patterns rely on, can be missed, and an ordinary number can occasionally be redacted by mistake.

These controls are aimed at structured fields, known entities and recognisable patterns in text - not at everything inside a document. When you send a file for analysis, its content has to reach the model substantially intact - that is what makes extraction possible at all - so anything inside it travels with it, including personal, confidential or sensitive information you did not intend to send. An image, a scanned page, or a file sent to a model as a file is not redacted by the controls above. We give no assurance that any particular item will be detected, masked or removed. If a document should not leave your control, redact it before you upload it, or, in a professional practice vault, switch it to "Do not send this document to AI" (section 9). If in doubt, do not send the file.

Analysis runs on our infrastructure providers and through the sub-processors listed at finpera.ai/legal/sub-processors. FinPera requires each to process personal data only for the contracted service and subject to confidentiality, security and privacy terms appropriate to the service.

Uploads are subject to file-type and magic-byte checks, malware scanning, injection detection, or other controls where applicable to the upload route and file type. Data is encrypted in transit and at rest, access is restricted by role, and AI activity is logged for audit, security, cost tracking, debugging and compliance purposes. The security measures described here are for information, reflect our practices at the time of writing, and may change. Our commitments are those in the Terms of Service and the Data Processing Agreement.

Extracted text and AI index entries are kept while the document exists and retention is needed to provide the Service. On deletion they are removed from active systems, subject to backup cycles, legal holds, audit logs, and retention required by law (for example statutory financial records).

Each document (professional practice vaults): a document in a professional practice vault can be switched to "Do not send this document to AI" in its settings. While the switch is on, its contents are not sent to AI providers, and AI features that draw on your documents - such as questions, reviews and drafting from vault references - leave it out. The document is still stored, security-scanned and available to you as normal. Results generated before the switch was turned on - such as a completed review, a draft, or a report saved to the Vault - are not withdrawn and may still be used. This per-document control is not yet available in company or investor workspaces.

Company and investor workspaces, where available: an administrator can stop workspace content being sent to AI providers. This applies to new processing; results already generated, such as search indexes, may continue to be used.

Professional practices: in DPDP assessments you can record each engagement's AI permission. Obtaining that permission from your client is covered by your confirmation in the Terms, not by this record.

Always running: security scanning, text extraction on your device, keyword search, masking and logging continue whatever these controls are set to.