# AI for Finance, Accounting & FP&A URL: https://justuju.org/role/finance/ For your role AI now drafts reconciliations, variance notes and forecasts, and reads contracts and filings in bulk. Finance keeps its oldest rule: a number counts only if you can trace it and someone signs for it. ## How AI is changing this work - The close gets agents that match transactions, prepare reconciliations and explain variances for a preparer to review. - Analysis shifts from building the spreadsheet to checking one an assistant built. - Research across filings, contracts and transcripts takes minutes, with a citation on each claim. ## Key layers, most important first ### 1. Layer 9 · Applications & Interfaces URL: https://justuju.org/layer/applications-interfaces/ Why it matters: Finance lives in spreadsheets, the ledger and the planning tool. AI arrives as an assistant inside each, so the first gains need no new system. What to do here: - Spreadsheet assistant: Describe the analysis; the assistant builds formulas, pivots and charts in the workbook and shows each step. (see https://justuju.org/concept/copilots/) - Ask the plan a question: Query budgets and forecasts in plain words and get the answer with the source model and date. (see https://justuju.org/concept/embedded-ai/) - Cited research: Search filings, transcripts and broker notes; accept only answers that link to the passage. (see https://justuju.org/concept/chat-assistants/) Examples as of 2026-11-11: Agent Mode in Excel (Microsoft 365 Copilot) (https://support.microsoft.com/en-us/office/frequently-asked-questions-about-agent-mode-in-excel-1cfd906d-40b4-46be-8e2d-65b893e28a02); Anaplan CoPlanner (https://www.anaplan.com/platform/anaplan-coplanner/); AlphaSense (https://www.alpha-sense.com/solutions/market-intelligence-platform/) ### 2. Layer 8 · Agents URL: https://justuju.org/layer/agents/ Why it matters: The close and payables are long, rule-bound, multi-step jobs: the kind agents do well. They are also where an unchecked action becomes a misstatement. What to do here: - Prepare, then sign: An agent gathers support, flags odd items and builds the reconciliation; a named person reviews and signs. (see https://justuju.org/concept/agent-loop/) - Invoice handling: Agents code invoices, check them against orders and flag fraud signals before approval. (see https://justuju.org/concept/agent-loop/) - Same controls as staff: Agents get roles, limits and separation of duties like any user, and every action is logged. (see https://justuju.org/concept/agent-permissions/) Examples as of 2026-11-11: BlackLine Verity (https://www.blackline.com/products/verity-ai/); Ramp Bill Pay agents (https://ramp.com/accounts-payable); Finance Agent for Excel (Microsoft) (https://learn.microsoft.com/en-us/copilot/finance/welcome) ### 3. Layer 7 · Orchestration & Tooling URL: https://justuju.org/layer/orchestration-tooling/ Why it matters: A language model should never do your arithmetic from memory. This layer makes it fetch figures from the ledger and compute with real tools. What to do here: - Numbers come from tools: The model calls the ledger, the warehouse or a calculator for every figure instead of recalling it. (see https://justuju.org/concept/tool-use/) - Documents in a grid: Run the same questions across hundreds of contracts or filings and check answers cell by cell against the source. (see https://justuju.org/concept/rag/) - Tie-out tests: Keep a set of questions with known answers from last quarter and re-run it before trusting a new setup. (see https://justuju.org/concept/guardrails-evals/) Examples as of 2026-11-11: Hebbia Matrix (https://www.hebbia.com/blog/how-private-equity-teams-use-hebbia); Model Context Protocol (MCP) (https://modelcontextprotocol.io/specification/2026-07-28/) ### 4. Layer 5 · Foundation Models URL: https://justuju.org/layer/foundation-models/ Why it matters: Models differ most on exactly what finance needs: multi-step reasoning and care with figures. Picking the right kind of model changes the error rate. What to do here: - Reasoning models for analysis: Use a model that works through steps before answering for variance analysis and policy questions. (see https://justuju.org/concept/reasoning-models/) - Know how it fails: A model predicts likely text. A plausible wrong number looks exactly like a right one. (see https://justuju.org/concept/next-token-prediction/) - Where the model runs: For material non-public data, check where the model is hosted, or use open weights you run yourself. (see https://justuju.org/concept/open-weights/) Examples as of 2026-11-11: Claude extended thinking (https://platform.claude.com/docs/en/build-with-claude/extended-thinking); DeepSeek-R1 (open weights) (https://huggingface.co/deepseek-ai/DeepSeek-R1) ## High-leverage use cases - Reconciliation preparation: Agents match transactions, attach support and draft the explanation; preparers review exceptions only. - Variance commentary: Draft budget-versus-actual notes from ledger data, with each figure linked to its source line. - Contract and lease review: Pull payment terms, renewal dates and covenants from a document set into one table. - Invoice coding and checks: Code, match and flag invoices before approval, with a person releasing payment. - Forecast scenarios: Ask what-if questions of the planning model and compare scenarios side by side. ## Start here: the next 30 to 60 days - Week 1 to 2: Pick one recurring report. Have an assistant rebuild it in a copy of the workbook and tie every number back by hand. - Week 3 to 4: Pilot agent-prepared reconciliations on two low-risk accounts, with the usual reviewer. - Day 30 to 60: Write the control: who reviews AI-prepared work, what evidence is kept, and which accounts are off limits. - Day 30 to 60: Brief your auditors early on where AI touches the close. ## Risks and pitfalls - Plausible wrong numbers: A model can state a figure that looks right and is not. One maker's own guidance warns against relying on its spreadsheet assistant for decisions in sensitive areas such as finance. Source: https://support.microsoft.com/en-us/office/frequently-asked-questions-about-agent-mode-in-excel-1cfd906d-40b4-46be-8e2d-65b893e28a02 - Broken audit trail: If you cannot show where a number came from and who approved it, the time saved is lost at audit. Keep sources and sign-offs. - Automated credit decisions: In the EU, AI that scores a person's creditworthiness is high-risk, with duties for the firm that uses it. Source: https://eur-lex.europa.eu/eli/reg/2024/1689/oj - Leaking material information: Unreleased results pasted into a public chat tool are a disclosure risk. Use approved tools with contractual limits on data use. - Agents with payment rights: An agent that can both approve and pay removes a control. Keep separation of duties for agents too. ## Sources - Agent Mode in Excel (Microsoft 365 Copilot) (support.microsoft.com): https://support.microsoft.com/en-us/office/frequently-asked-questions-about-agent-mode-in-excel-1cfd906d-40b4-46be-8e2d-65b893e28a02 - Anaplan CoPlanner (anaplan.com): https://www.anaplan.com/platform/anaplan-coplanner/ - AlphaSense (alpha-sense.com): https://www.alpha-sense.com/solutions/market-intelligence-platform/ - BlackLine Verity (blackline.com): https://www.blackline.com/products/verity-ai/ - Ramp Bill Pay agents (ramp.com): https://ramp.com/accounts-payable - Finance Agent for Excel (Microsoft) (learn.microsoft.com): https://learn.microsoft.com/en-us/copilot/finance/welcome - Hebbia Matrix (hebbia.com): https://www.hebbia.com/blog/how-private-equity-teams-use-hebbia - Model Context Protocol (MCP) (modelcontextprotocol.io): https://modelcontextprotocol.io/specification/2026-07-28/ - Claude extended thinking (platform.claude.com): https://platform.claude.com/docs/en/build-with-claude/extended-thinking - DeepSeek-R1 (open weights) (huggingface.co): https://huggingface.co/deepseek-ai/DeepSeek-R1 - Regulation (EU) 2024/1689 (the AI Act): https://eur-lex.europa.eu/eli/reg/2024/1689/oj - NIST AI Risk Management Framework (AI RMF 1.0): https://www.nist.gov/itl/ai-risk-management-framework --- Justuju · https://justuju.org/ · Content CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote with a link to the page.