AI & Data Privacy

Last updated: October 9, 2026

LedgerBeaver LLC, a Texas limited liability company, provides a hosted finance workspace where you choose which external AI agents to connect. Separate in-app features can also use providers engaged by LedgerBeaver. This page explains who sees what.

1. Who sees your data

Your data lives in your LedgerBeaver workspace. Invoices, bills, journal entries, and contacts are stored in your company's LedgerBeaver workspace, isolated from other companies. A connected AI provider may receive the data needed for a request; its retention and use follow your provider account, plan, and settings.
A connected agent can disclose data to its providers. If your agent uses a cloud model or another remote service, prompts, files, tool results, financial records, and related personal data may reach that service within the access you authorize. Review both the agent's and the model provider's terms, retention, data location, and settings before connecting business books.
Verified summaries, LedgerBeaver advisories, and optional hosted chat are separate. Eligible Pro and Max workspaces can separately enable TypeSafe Jev suggestions for AP exceptions, employee-expense accounts, purchase-order evidence, possible vendor-name matches, and selected CRM note follow-up. Each purpose requires its own workspace consent and an explicit request. General product-help model requests and cloud document transcription remain disabled. Ask Beaver book answers have a separate opt-in described below. Verified book summaries and documentation help do not send content to a model provider. A local external agent does not control these separate features.
LedgerBeaver never sells your data, never shares it for advertising, and never trains AI models on it. These commitments are in our Privacy Policy and Data & Compliance statement. Our published Data Processing Agreement remains a draft pending legal review, not an executed agreement.

LedgerBeaver AI advisories (TypeSafe Jev)

An owner or administrator can separately enable each purpose listed below for an eligible Pro or Max workspace. Enabling one purpose does not enable the others. A finance reviewer in LedgerBeaver, or an authorized connected agent, makes a separate AP-exception request. An expense owner or authorized approver makes a separate expense request. Expense requests contain only expense date, amount, normalized category, description, business-purpose/use, fixed server-resolved allowed account candidates, and approved coding-policy fields including legal entity type; employee identity fields, receipt contents, bank account and payment settlement records, reimbursement status, private notes, and credentials are excluded.

Purchase-order evidence guidance sends the eligible PO status, bounded line descriptions, and a flag that supporting evidence is incomplete. Possible vendor-name comparisons send bounded display/legal names and available country codes for a small candidate set; tax IDs, contact details, bank details, and vendor record IDs are not sent. Exact email or tax-ID matches are checked locally.

CRM follow-up guidance uses one note you select. LedgerBeaver prepares a limited excerpt, filters sensitive patterns, and shows the exact proposed excerpt before you confirm a request. Review that preview: automatic filtering cannot guarantee removal of every confidential or personal detail. Only the confirmed excerpt and allowed follow-up categories are sent. The result can direct you to an existing relationship or accounting view; it does not contact anyone or create a follow-up commitment.

Responses are advisory. An expense suggestion cannot change an account until a person explicitly applies it, and approval remains a separate decision bound to the displayed coding revision. It cannot approve, reimburse, post, or pay the expense. PO guidance does not create a receipt or approve a purchase. A possible vendor-name match does not verify legal identity or merge records. TypeSafe states that Jev customer requests and responses are not used to train its models. That commitment does not mean no processing or zero retention: processing and retention follow LedgerBeaver's agreement and TypeSafe's applicable terms. TypeSafe assessment tasks outside these listed purposes remain disabled.

2. Choose your privacy level

What an AI service does with the data it receives depends on its product, your account, contract, and settings. The connection method alone does not establish a privacy guarantee:

⚠️ Consumer subscription (Claude Pro, ChatGPT Plus, Gemini)

Depending on the applicable terms and settings, consumer AI products may use conversations and other content to train their models. Check whether training is enabled, what any opt-out covers, and how long content is retained before sharing business records. A paid consumer subscription does not by itself establish business-data protections.

✅ Business or API account

API and business plans often provide stronger data-use controls than consumer subscriptions, but terms, defaults, retention, and opt-out settings vary by provider and can change. Verify your provider's current business-data terms and account settings before sending financial data. Using a separately governed business/API account is the recommended default for business use.

🔒 Local model (Ollama, LM Studio, vLLM, llama.cpp)

Running your connected agent's model entirely on your hardware can avoid sending that agent's model inputs to a cloud AI provider. This depends on the whole setup: the agent, model endpoint, extensions, telemetry, synchronization, backups, and any cloud fallback. A desktop app or locally installed runtime may still use cloud services. LedgerBeaver continues to host your workspace, and separate in-app features may use their own providers.

No training does not mean no processing. A provider may retain content, process it for security, or permit human review under its applicable terms even when model training is disabled. Check these separately.

3. Running LedgerBeaver with a local model

Use an MCP-capable agent configured to run inference through your local model endpoint and connect to LedgerBeaver. Verify the agent's model selection and outbound services:

your local model → MCP-capable agent → LedgerBeaver
Ollama / LM Studio / vLLM / llama.cpp → Continue · Cline · OpenCode → our MCP endpoint

Install a local runtime. Ollama (ollama pull a model that fits your RAM — a capable tool-use model is strongly recommended for accounting work) or LM Studio (download a model, start the local server).
Pick an MCP-capable agent that supports custom / local model endpoints — for example Continue, Cline, or OpenCode. Point its model setting at your local runtime (Ollama and LM Studio both expose an OpenAI-compatible endpoint on localhost).
Connect the agent to LedgerBeaver. Generate an API key in the app (Connect AI page) and paste the ready-made MCP config snippet for your agent. Every snippet on that page auto-fills your key.

Hosted workspace, local inference: your books are still transmitted to and stored in LedgerBeaver's cloud workspace and processed by its service providers as described in the Privacy Policy. Local inference can avoid sending the connected agent's model inputs to a cloud provider only when the agent and related components do not send content there. It does not disable separate in-app provider features or change your existing rights in Customer Data or other individuals' privacy rights.

4. Practical ways to minimize exposure

Choose a business or API account with suitable contractual protections for business books; verify the terms that apply to your account.
Review training, retention, human-review, and data-location controls for both your agent and its model provider.
Use read-only periods: LedgerBeaver's write-authority controls let owners keep AI money-writes off until explicitly enabled.
Sensitive fields (email, phone, tax IDs) are masked by default in the views your AI reads.
For local inference, verify cloud fallback, telemetry, extensions, and backups (Section 3).
Revoking a connection blocks new access to LedgerBeaver through that connection after revocation, but does not retrieve information already received by the agent or its providers.

5. Your provider's policies

Services you independently select process information under their applicable terms and policies. Provider-wide privacy pages are a starting point; also check the terms for your specific product and account:

Anthropic privacy policy (Claude)
OpenAI privacy policy (ChatGPT, GPT API)
Google privacy policy (Gemini)
OpenAI business-data commitments
Anthropic commercial terms

6. Our commitments

As between you and LedgerBeaver, you retain the rights you have in Customer Data; we claim no ownership. Data held by LedgerBeaver is protected in transit and at rest, never sold, never used for advertising, and never used by LedgerBeaver to train models. See our Privacy Policy, Terms, and Data & Compliance statement. Our published DPA remains a draft pending legal review. Questions: [email protected].

Using automated processing or model inference for an enabled feature is separate from training models. Providers LedgerBeaver engages to process Customer Data on our behalf remain subject to our applicable data-protection and sub-processor obligations. We do not treat a provider we engage as your independently selected service merely because it powers a LedgerBeaver feature.

For the product workflow behind these disclosures, see connected AI for accounting, including scoped access, source checks and separate human approval.

Optional Ask Beaver book answers

When enabled, Ask Beaver offers optional, read-only book answers using Apodex Mini through OpenRouter and Novita. After you opt in for the current conversation, your question, recent conversation and a limited set of authorized book facts are sent to those providers. Relevant invoice references, account labels or counterparty names may be included; raw documents, payment credentials, contact fields, internal record IDs and agent memory are excluded. Filtering cannot guarantee that every sensitive detail typed into a question or label is removed. The model selects supporting facts; LedgerBeaver validates those selections and displays source-backed amounts and dates. It cannot change books, approve transactions, send messages or move money. Requests are restricted to endpoints OpenRouter designates as no-training and zero-data-retention, with no paid or alternate-provider fallback. LedgerBeaver excludes its production key from OpenRouter data-discount logging and keeps prompt/completion storage off. Provider policies and contractual obligations still apply; processing is US/global, not guaranteed to stay in one region. Turn AI book answers off to use internal verified summaries. The choice resets when you change user or company or reload.

AI selection can be mistaken or incomplete. Check the supporting ledger or report; a displayed answer is not certification of reconciliation, close readiness or accounting accuracy. Provider routing policies and OpenRouter data handling.