> ## Documentation Index > Fetch the complete documentation index at: https://docs.dualentry.com/llms.txt > Use this file to discover all available pages before exploring further. # AI Data Handling: Frequently Asked Questions > How DualEntry's AI changes models, stores and isolates your data, avoids training on it, and what flux analysis needs to explain variance drivers. This page answers the questions security reviewers, auditors, and investors most often ask about DualEntry's AI features: how the underlying models can change, where your data lives, how it stays separate from other customers, whether it trains models, and what variance analysis needs from you. For the platform-wide controls these answers build on, such as encryption, SOC reports, and access policies, see [Security and Compliance](./security-and-compliance). ## Can DualEntry change the underlying AI model without changing my skills or agents? Yes. DualEntry's AI features call language models through a provider-agnostic layer, so the model behind a feature can change without any change to your skills, scheduled agents, or saved work. Your configuration is stored as plain instructions, not as anything tied to one model. A [skill](../ai-automation/accounting-intelligence/skills) is the wording your admins wrote, and a [scheduled agent](../ai-automation/accounting-intelligence/scheduled-agents) is that wording plus a cadence and an accounting basis. When the model changes, the same instructions run against the same ledger data under the same permissions. DualEntry chooses the model for each AI feature by the level of reasoning the task needs, rather than letting each organization pick one. You don't select or supply a model inside DualEntry, and nothing in your organization's setup has to change when DualEntry moves to a different one. DualEntry checks changes before they reach you in two ways: * **Regression benchmark:** Changes to DualEntry's AI code, including prompt changes, run against a benchmark of accounting questions with known answers. The results are compared with the last passing run, so a drop in accuracy is visible before the change ships. * **Guardrails:** Conversational agents screen incoming messages for abuse and for text that tries to override their instructions. Database queries an agent writes are checked to stay inside your organization before they run. Responses are scanned to redact patterns such as card numbers and Social Security numbers. Numbers in an answer come from your ledger rather than from the model's general knowledge, which is why a model change affects wording more than figures. See [how Accounting Intelligence grounds its answers](../ai-automation/accounting-intelligence/ai-accounting-intelligence#data-grounding-and-compliance). ## Where is my data stored, and how does it reach the AI model? Your accounting data stays in DualEntry's encrypted cloud database, and an AI feature sends a model the records a request needs, plus context such as your company list and, if memory is on, your saved memories. Conversations, skills, and memories are stored in the same encrypted database as the rest of your organization's data. Prompts and responses, including any records they contain, are also processed by DualEntry's monitoring providers (Datadog and LangSmith) under the same security controls. When you ask Accounting Intelligence a question or a scheduled agent runs, DualEntry retrieves the relevant records under your permissions and includes them in the request to the model provider. The response comes back to DualEntry, which shows it to you. Any change an AI feature makes to your records is logged in the [audit trail](./audit-trail-and-compliance) and attributed to Accounting Intelligence. The conversation itself is not written to the audit trail. The AI model providers DualEntry uses are listed on the [Trust Center subprocessor list](https://trust.dualentry.com/subprocessors). Some features match records by meaning rather than exact text, for example when suggesting an account for a transaction. DualEntry builds those comparisons from your organization's own records when the suggestion is made. They are not added to an index shared with other customers. DualEntry does keep one searchable index for AI features: a knowledge base of DualEntry's own product documentation and in-app navigation, which Accounting Intelligence uses to answer how-to questions. It contains no customer data. For encryption in transit and at rest, see [Security and Compliance](./security-and-compliance#how-dualentry-protects-your-data). ## How is my data kept separate from other customers? Every request in DualEntry, including every AI request, is scoped to your organization, and access is further narrowed to the companies and permissions of the person or agent making it. Isolation is enforced in the application on every query. AI features inherit that scoping rather than having their own: * **Organization:** An AI request carries your organization's identity, and every tool it calls is bound to it. Queries that an agent writes must filter to your organization and pass a check before they run. * **Companies:** If a user is restricted to certain companies through [entity scoping](./user-roles-and-permissions#entity-scoping), Accounting Intelligence sees only those companies for that user. * **Permissions:** Accounting Intelligence can only reach records the user could open themselves. A scheduled agent runs under the permissions of the user who owns it. ### Multi-entity groups and portfolios A DualEntry organization can hold many companies, and how you divide a group between organizations decides what is shared. Companies inside one organization share users, roles, and skills, and can be consolidated. Separate organizations share nothing: a private equity firm with several portfolio companies on DualEntry keeps each portfolio company's data, AI context, and skills separate when each one is its own organization. See [How to Configure Multi-Entity and Consolidation](./multi-entity-consolidation). ## Does DualEntry train AI models on my data? No. DualEntry does not train or fine-tune AI models on your data. Your records, corrections, confirmed matches, and conversations do not feed a training pipeline. Some features improve as you use them, but they do it by reading your own records as context at the moment they run, not by changing a model. That learning stays inside your organization. The three kinds are: * **Skills:** Instructions your organization's admins write and save. Every skill is visible to your whole organization and to no other organization. See [Skills](../ai-automation/accounting-intelligence/skills). * **Feature context:** Bank Match AI, automated categorization, and anomaly detection reuse your past matches and decisions as context. See [How AI features use your data](./security-and-compliance#how-ai-features-use-your-data). * **Memories:** Short facts Accounting Intelligence keeps about an individual user so it can tailor its replies to them. ### How memories work Memories belong to one user, not to the organization, and other users never see them. Memory is on by default. While it is on, a daily process reviews that user's recent messages and keeps up to 30 short, durable facts, such as the user's role. A fact that is not reconfirmed within 15 days is removed automatically. Each user can view what has been remembered, delete individual facts, or turn memory off in **Settings** in the [Agents panel](../ai-automation/accounting-intelligence/agents-panel). Turning memory off stops new messages from being reviewed for memories. ## How much setup does AI variance analysis need to explain drivers? Very little: set a materiality threshold, and record vendors and customers on your transactions. DualEntry's flux analysis (period-over-period variance review) attributes variances to drivers by calculation, and the AI writes the explanation from those calculated drivers rather than guessing at causes. ### What DualEntry calculates for you For an account that moved more than your threshold, DualEntry breaks the variance down by counterparty: * **Expense and payable accounts:** Ranked by vendor, showing the vendors that contributed most to the change. * **Receivable accounts:** Ranked by customer. * **Large transactions:** Any single transaction worth more than 25% of the account's variance is called out. * **New vendors:** Vendors with no activity in the prior period are flagged when they account for 10% or more of the change. * **Recurring revenue:** Monthly recurring revenue (MRR) movement is broken down by customer into new, expansion, contraction, and churn. ### What you need to set up * **Materiality threshold:** Set an amount, a percentage, or both, separately for the income statement and the balance sheet. Without one, DualEntry uses a 5% threshold. See [Configure materiality thresholds](../core-financials/close-management/flux-analysis#configure-materiality-thresholds). * **Vendors and customers on transactions:** Driver attribution depends on them. Entries with no counterparty, such as payroll or depreciation journals, show the variance without a vendor breakdown. * **Contracts for MRR:** Customer-level MRR movement is calculated from contracts, so recurring revenue tracked only in journal entries can't be attributed to customers. * **Descriptive memos:** The AI's explanation is clearer when journal entry memos say what the entry was for. Flux drivers are attributed to vendors and customers only, not to departments, locations, or other classifications. To run an analysis and review flagged variances, see [Flux Analysis](../core-financials/close-management/flux-analysis). ## Related reading * [Security and Compliance](./security-and-compliance) * [How Accounting Intelligence Works](../ai-automation/accounting-intelligence/ai-accounting-intelligence) * [User Roles and Permissions](./user-roles-and-permissions) * [Trust Center](https://trust.dualentry.com) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.