Trust & Control
Security, Privacy & AI
For a CPA firm the first question about AI is not what it can do, it is who stays responsible. This page describes how WorkpaperOS is actually built - where documents live, what an AI agent can reach, what gets recorded, and how you switch it off. We would rather be specific than reassuring.
Your data and AI models
We do not train on your work
WorkpaperOS does not train its own AI models on your documents or engagement data. Depending on the feature and your settings, document extraction and AI assistance may use Google Cloud, OpenAI, Anthropic, or a connected AI provider. Each provider's applicable terms govern its processing; using its service is not a claim that WorkpaperOS holds that provider's certifications.
With MCP, the account is yours
When you connect Gemini Spark, Claude, ChatGPT, or another MCP client, the AI runs under your subscription, not ours. Whatever data terms you hold with that provider are the terms that apply - including an enterprise agreement your firm has already negotiated. We are the system it connects to, not the party in the middle of your AI contract.
Connect your AI - the MCP server URLs
https://attestwork.workpaperos.com/api/mcphttps://book.workpaperos.com/api/book/mcpWhat an AI agent can actually reach
The same permissions as the person
You invite someone and choose exactly what they may open - which areas of the books, read-only or read-and-write, whether they can download files or run reports. Those choices are enforced on every MCP tool call, not just in the browser. An agent acting for a read-only user cannot post a journal entry, and a tool outside that user's areas is refused before it touches any data.
Refused by default
Every tool is mapped to a permission, and anything unmapped is refused rather than allowed. If we cannot establish who is asking and what they are entitled to, the call fails closed. New capability has to be granted deliberately; it never arrives switched on.
The built-in assistants
Same rails as the connector
AttestWork and Books ship with built-in assistants. They are not a separate, softer path into your data: every tool call they make goes through the same firm-scoped, audited tool layer as a connected MCP client, is checked against the same permissions, and lands in the same log. Drafts stay drafts until a person signs off.
Approval cards, and documents as data
Anything sensitive - a sign-off, a deletion, a file move, posting an entry, creating an engagement - stops at an in-chat approval card and waits for a click from a person. The assistants are built to treat document text as data, not instructions - and no matter what a document says, anything sensitive still stops at that approval card.
Where documents live
Private storage, row-level isolation
Engagement files are held in private storage and delivered through authorized, short-lived transfers. Database row-level security policies and application authorization both help isolate firm data. The controls depend on the data and access path; we do not claim that every database table relies on the same policy.
Scoped file exchange, not a folder mirror
AttestWork Files and Books Files use the product's existing scoped MCP authorization. Original files travel over short-lived HTTPS transfers, outside MCP message payloads. The assistant receives only the access you authorize; your local folders are not continuously mirrored into the app.
Books bank feeds and payments
WorkpaperOS supplied the business, application, and verification information requested during provider onboarding before enabling these production integrations. Approval covers the specific services made available to Books; it is not an independent security certification of WorkpaperOS.
Bank feeds through Plaid
Books uses Plaid Link and Transactions to connect authorized bank accounts and import activity. Plaid's full production registration includes company and application information and, for access to certain US bank APIs, a security questionnaire. Bank-feed access does not authorize WorkpaperOS to initiate transfers. Plaid Transfer would require a separate approval. Plaid reports a SOC 2 Type II examination and ISO 27001 certification for its own services.
Invoice payments through Stripe
Books uses Stripe Connect for a firm's payment account and Stripe-hosted Checkout for customer invoice payments. Stripe reviews business, identity, website, and payment-risk information and may request supporting documents before or after live activation. Connected firms complete their own applicable Stripe verification. Stripe is a PCI Level 1 service provider; that status belongs to Stripe, not WorkpaperOS.
Infrastructure and provider assurance
WorkpaperOS uses established providers for particular services. Their independent assessments cover the providers' stated services and scope, not the WorkpaperOS application or firm. Our own configuration, access controls, and operations remain our responsibility.
- Supabase provides database and authentication services and reports SOC 2 Type II compliance and ISO 27001 certification.
- Vercel hosts application delivery and reports a SOC 2 Type II attestation and ISO 27001:2022 certification.
- Google Cloud Document AI supports document extraction; Google describes its applicable third-party security audits and service controls.
- Cloudflare R2 stores application files and documents its encryption at rest and in transit.
- Render hosts supporting document services, including ONLYOFFICE and OCR, and reports SOC 2 Type II and ISO 27001 assurance.
You stay responsible
AI proposes, you post
AI tools act within the permissions and workflows available to the connected user. Some actions create drafts for human review; others can change records when write access is granted. Professional judgement and required sign-offs stay with your firm.
MCP activity and audit trail
MCP activity is logged with the firm, user, connected client, tool, and outcome. Business changes also retain their application audit records. Operational logging can fail independently, so we do not describe every attempted call as guaranteed to appear in the MCP activity log.
Revoke in one click
An administrator can revoke a connected AI client from settings. Subsequent authorization checks evaluate that revocation. For an urgent access concern, contact us so we can verify the connection is disabled.
If there is a specific control your risk assessment turns on, ask us and we will tell you plainly whether it exists.