--- title: "Private Finance AI — Whitepaper" description: "How finance teams generate validated spreadsheets whose figures tie out, and analyse sensitive data read-only, on an agent running on their hardware." url: "https://royalsoftworks.com/resources/finance-ai-on-device/" source: "https://royalsoftworks.com" format: "markdown" note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL." --- [All resources](https://royalsoftworks.com/resources/) # Private Finance AI ## Spreadsheets That Tie Out, Generated On-Device **A Royal Softworks whitepaper · AssistantGeneral** --- ## Abstract Finance teams handle material non-public information and live in spreadsheets — two facts that make cloud AI a poor fit. The data can't be pasted into a third-party SaaS box, and the generic chat tools are unreliable at producing a workbook whose formulas actually work. AssistantGeneral keeps financial data on the analyst's machine and generates **validated** spreadsheets and reports using a sandboxed code engine, with read-only access to the systems where the numbers live. --- ## 1\. Why cloud AI fails finance **Confidentiality.** MNPI, deal data, and client financials are exactly the material that cannot transit a vendor cloud — for legal, fiduciary, and competitive reasons. Audit and SOX-adjacent controls also demand a defensible answer to "where did this data go?" **Correctness.** Ask a generic chat assistant for a financial model and you frequently get a spreadsheet with broken references, circular calculations, or numbers that don't reconcile. In finance, a workbook that doesn't tie out is worse than no workbook. ## 2\. Data stays on the device AssistantGeneral runs the finance stack locally: the model, the document index, the embeddings, and the document generation engine are all on-device. A cloud model is attached only when policy permits. Financial documents ingested for analysis are indexed locally; nothing is sent to a vendor to be embedded or processed. ## 3\. Spreadsheets that actually tie out The document engine doesn't approximate a spreadsheet in Markdown — it generates a real `.xlsx` by writing Python (openpyxl, XlsxWriter, pandas, NumPy) that executes in a **sandboxed Pyodide (WebAssembly) worker** on the device. The design is built for numerical reliability: - **Layered construction** — the workbook is built in stages rather than one fragile pass. - **Workbook validation** — the math is checked; the engine enforces sensible, one-way calculation flow rather than emitting circular or hand-waved formulas. - **Surgical edits** — changes are applied as targeted line-level edits, not full rewrites, so a small correction doesn't risk the rest of the model. - **The source file is never overwritten** — edits are delivered as a separate copy, so an original model or template is always preserved. The same engine produces `.docx` reports and `.pptx` board decks from a brief. ## 4\. Analyze where the numbers live — read-only AssistantGeneral connects to the systems finance teams already use — SQL databases (Postgres/MySQL), cloud storage (S3, Azure Blob, Google Cloud Storage, Dropbox, OneDrive), and business systems — through a connector layer that is **read-only and SELECT-only by design**. The agent can discover sources, inspect schema, run a read query, and pull figures into a reconciliation or model — without any ability to write back. For internal systems that aren't a built-in connector, a **custom-connector builder** wires them up. ## 5\. Grounded narrative Financial commentary — MD&A, board narrative, variance explanations — is generated _grounded in the actual figures pulled_, with the groundedness gate keeping claims tied to the data and the retrieved documents. Contract and filing analysis reuses the same clause-extraction and expiry-tracking capabilities the legal library provides, applied to financial agreements, leases, and covenants. ## 6\. Repeatable, scheduled The monthly close digest, the weekly KPI pack, the recurring reconciliation — these are assembled once in the **visual workflow builder** and run on demand or on a **cron schedule**. What starts as a few custom workflows becomes the finance function's reusable internal library. ## 7\. The interface matches the work A **Finance UI preset** presents a data-and-connector-forward layout with voice surfaces hidden — built around the connectors, the knowledge base, and document generation. ## 8\. For the firm: governance When a finance function standardizes on AssistantGeneral, the governance plane (on the firm's own server) provides a shared knowledge base of policies and templates, single sign-on, an audit trail, and a policy that pins seats to the on-device model and the approved connectors — so the controls story is as clean as the data-residency story. ## 9\. Scope AssistantGeneral accelerates the mechanical and repetitive parts of financial work — model and report generation, document extraction, read-only data pulls, reconciliations, and grounded narrative — under the analyst's review, on data that never leaves the machine. It is a productivity tool for finance professionals, not an unsupervised decision-maker. --- _Evaluate it on a real model: generate a multi-tab budget from a brief, open it in Excel, and confirm the formulas are live and correct — offline. Contact Royal Softworks._ ## Evaluate AssistantGeneral See the product this paper describes — a private, local-first AI agent with an enterprise control plane that runs on your own infrastructure. [Explore AssistantGeneral](https://royalsoftworks.com/products/assistant-general/) [Talk to us](https://royalsoftworks.com/contact/)