Top 3 Implementations Today
1 | Spendesk CFO: MCP Directly Connects to Claude, Compressing Three Core Finance Reports from Hours to Under 8 Minutes
- Scenario: Travel compliance checks, budget vs. actual OPEX reviews, AP aging analysis.
- Actionable Steps: Spendesk CFO Pauline Babell shared three specific prompt workflows. Travel compliance: “Generate travel policy compliance report, April 2026”, outputting violation list, cost center allocations, and recommended corrections. Budget comparison: “Generate budget report for Q4 2025”, outputting variances, unplanned expenditures, risk and action plans by cost center. AP aging: “AP aging report — overdue exposure this week”, outputting overdue exposure, invoice counts by aging tier, and priority action list.
- Review Controls: MCP connection set to read-only (cannot create payments, approve invoices, modify records); data permissions aligned with Spendesk user permissions; Claude Enterprise does not use company data for training. Each query result must be reviewed by a controller or AP owner before reporting to management.
- Outputs: Compliance report, OPEX review deck, AP aging report—all can be set to run automatically monthly.
- Key Prerequisite: Data must first be consolidated into a clean platform. Pauline Babell’s original words: “Consolidation comes first. You cannot connect what is not consolidated.”
- Source: CFO Connect — MCP and Claude for Finance (CFO practice sharing published by Spendesk community, 2026)
2 | Cube Founder: AI vs. Automation Decision Framework—Two Questions to Determine Which to Use
- Scenario: Finance team selection—whether to use rule-based automation or AI for a process.
- Actionable Steps: Use two dimensions to draw a decision matrix. (1) “Can I clearly write the rules?”—if yes and covering all exceptions → automation; if judgment, context, or pattern recognition is needed → AI. (2) “Is audit requirement high?”—output for auditors, board, or capital allocation → high auditability, must have manual review and confidence thresholds; internal analysis/scenario planning → light governance. Highest risk quadrant: uncertain rules + high audit requirements.
- Review Controls: High-auditability AI outputs must include documented assumptions, confidence thresholds, and sign-off by designated reviewers.
- Outputs: Decision matrix table. Cube’s suggested specific targets: rule-determined → automation scenarios include recurring journal entries, period-end accruals, intercompany eliminations, variance threshold alerts, report distribution, data validation, balance sheet reconciliation; AI scenarios include natural language cross-system queries, large transaction anomaly detection, scenario narrative/commentary drafting, multi-department forecast input synthesis.
- Key Warning: Automation risk is “solidifying current understanding flaws”; AI risk is “no auditable reasoning path”.
- Source: Cube Software — AI vs. Automation in FP&A (Vendor founder article with reusable framework, updated May 2026)
3 | New CFO First 90 Days AI Onboarding Playbook—Week-by-Week Breakdown
- Scenario: New CFO or finance leader quickly onboarding and establishing AI work foundation.
- Actionable Steps: Before taking office: request recent 24 months’ board decks, investor materials, strategic plans, use AI to synthesize in advance. Weeks 1-2: Use AI to synthesize historical documents, identify recurring themes, risks, and gaps; first thing is clarify data lineage—where each KPI comes from, who owns source data. Weeks 3-4: Build process checklist, map each manual process’s owner, time cost, downstream dependencies, avoid touching any “load-bearing” processes. Month 2: Audit all periodic reports—three questions: (1) Who reads it, what decisions does it drive? (2) If stopped for 48 hours, would anyone notice? (3) Is there more accurate data elsewhere? Goal to cut about 50% of “zombie reports”. Month 3: Deliver forward-looking financial model with scenario analysis, not just historical review.
- Review Controls: Every number must be traceable to audit-grade, connected data sources. AI output only adopted if traced to trusted infrastructure.
- Outputs: Process checklist, zombie report cleanup log, forward-looking financial model.
- Source: Cube Software — The New CFO’s First 90 Days (Vendor article with actionable playbook, 2026)
Accounting / Close / Controls
1 | Rule-Determined Accounting Processes Should Prioritize Automation Over AI
Based on the Cube decision framework (see Top 3 Implementations Today item 2), the following processes fall into automation targets with “writeable rules, determined data sources, consistency priority”: recurring journal entries and period-end accruals, intercompany eliminations with clear logic, variance alerts triggered by thresholds, report distribution and scheduling, data validation checks, balance sheet reconciliation. Judgment standard: if a junior analyst can write the steps in 10 minutes and the process is the same monthly, it should be automated.
- Risk Control: Before automation, audit existing rules to ensure they are still correct, processes have changed, and outputs are reviewed by someone. The biggest hidden risk of automation is “solidifying current understanding flaws”.
- Source: Same Cube article as above.
2 | StackAI: Capital Account Reconciliation AI Agent (Vendor Demo)
Case study for a wealth management fund: capital account reconciliation originally required analysts to manually verify thousands of records across multiple entities, statements, and ledgers. StackAI demonstrated building an Agent to automate matching and initial screening. Input: multi-entity capital account statements and ledger data. AI action: cross-entity automatic matching, flagging anomalies. Manual review: analyst confirms each anomaly flagged by AI. Output: reconciliation report.
- Note: This is a vendor demo; actual deployment effectiveness needs self-verification. But the process architecture (multi-source data → Agent matching → human review of anomalies → output report) can be referenced.
- Source: YouTube — StackAI: Build a Capital Account Reconciliation AI Agent (Vendor demo video, released late 2025, date unspecified)
FP&A / Planning / Reporting
1 | Budget vs. Actual OPEX Review: From One-Week Cycle to 4 Minutes
See Top 3 Implementations Today item 1. Prompt “Generate budget report for Q4 2025”, output including variances, unplanned expenditures, cost center tiers, risk and action plans. Pauline Babell’s original words: “No more Google Slides for OPEX review where it’s pending review. That’s powerful.” Key: can be set to run automatically monthly; the fifth workflow’s setup cost is much lower than the first—compounding effect.
2 | High-Value Scenarios for AI in FP&A
Based on the Cube framework, AI is suitable for “problems requiring advanced analysts to truly think to answer”: natural language queries across financial systems, large transaction set anomaly detection, scenario narratives and executive commentary drafting, synthesizing multi-department forecast inputs, predictive analysis of non-linear historical data. Key condition: data must be trustworthy first for AI to be useful. “AI doesn’t make data trustworthy. The data has to be trustworthy first.”
- Source: Same Cube article as above.
Treasury / Cash / Risk
1 | AP Aging Report: From Hours to 7.3 Minutes
See Top 3 Implementations Today item 1. Prompt “AP aging report — overdue exposure this week”, outputting total overdue exposure, invoice counts by aging tier, cash curve visibility, priority action list. Can be run automatically weekly.
Data not available. This period did not find new AI implementation cases or practical methods within the last 365 days on cash forecasting, bank transaction automatic matching, DSO/O2C risk monitoring.
Tax / Compliance / Audit
Data not available. This period did not find new AI implementation cases or practical methods within the last 365 days on tax research, SOX/internal controls, or audit evidence management.
CFO / Leader Team Building Experience
1 | cjgustafson (Startup CFO): You Cannot Vibe Code a Public Company
X user cjgustafson self-identifies as Startup CFO, recently published two related posts: “You Cannot Vibe Code a Public Company: Where AI Actually Works in Finance” and “How do you build an AI-first finance team? And maybe more importantly, how do you trust the outputs?” The former discusses the real applicable boundaries of AI in finance; the latter focuses on how to build an AI-first team and trust outputs.
- Lead to Verify: X content could not be fully extracted; unable to confirm specific process details. But from bio “Startup CFO” and topic positioning, it is a valuable lead for tracking AI-native finance team organizational methods.
- Source: X — cjgustafson / X — cjgustafson (Operator signal, to be verified)
2 | AI Ops / Chief of Staff Hiring Signals: AI-Native Companies Scaling BizOps Recruitment
AI Operators Newsletter (Evan Lee) published Vol. 026 on May 13, 2026, summarizing 50 AI company BizOps/Chief of Staff positions. Over the past year, 693 positions published, with 3,300+ subscribers. Notable signals: Variance (AI agents for compliance) hiring Business Operations Manager; Paradigm (AI-powered spreadsheet) hiring Business Operations; multiple Series A-B companies hiring Head of Operations and Strategy & Operations Manager.
- Interpretation: AI-native companies are using BizOps/CoS roles to take over finance/operations work that originally required more specialized positions, which is an organizational signal of headcount substitution.
- Source: AI Operators Newsletter — Vol. 026 (Hiring signal summary, May 2026)
Open Source / AI Engineering Insights
1 | Deja Vu: Local-First AI Memory Layer for Cross-Tool Context Sharing
GitHub project JSingletonAI/dejavu, 204 stars, Python implementation, Apache-2.0 license. Core architecture three layers: interface layer (Python SDK / CLI / REST / MCP Server) → memory engine (uses Venice API for memory extraction and ranking) → local SQLite storage (~/.dejavu/memories.db).
- Finance Team Insights: If teams use Claude Desktop / Cursor and other MCP clients for financial analysis, Deja Vu allows AI to remember company accounting policies, account structures, historical variance explanations, etc., across different sessions without re-entering each time. All data stays in local SQLite, not sent outbound.
- Applicable Pilot: First try in FP&A scenarios—store company cost center structure, common KPI definitions, historical commentary into Deja Vu, observe if Claude reduces hallucinations and repeated questions when generating monthly commentary.
- Note: Depends on Venice API for memory extraction (API key required); 204 stars indicates early-stage project, stability needs assessment for production environment.
- Source: GitHub — JSingletonAI/dejavu (Open-source project, continuously active)
Small Experiments for This Week
1 | Run AP Aging Report with Claude + MCP
- Data: Export recent one week’s AP aging data (vendor, invoice number, due date, amount, payment status).
- Action: Configure Spendesk or similar tool’s MCP connection, use prompt “AP aging report — overdue exposure this week” to query.
- Review: AP owner compares AI output overdue amount and invoice count against ERP raw data item by item.
- Output: AP aging report + variance record.
- Judgment Standard: If accuracy >95% and generation time <10 minutes, worth expanding to monthly automatic run.
2 | Draw an AI vs. Automation Decision Matrix
- Data: List finance team’s current 10 most time-consuming monthly processes.
- Action: Score using two dimensions—(1) Rule certainty (1-5, 5=rules completely clear); (2) Audit requirement (1-5, 5=for auditors/board). Draw 2x2 matrix, place 10 processes into it.
- Review: Controller reviews if classification is reasonable.
- Output: Decision matrix table + automation priority ranking.
- Judgment Standard: Upper right quadrant (rule-determined + high audit) processes should complete automation POC within 30 days.
3 | Use Deja Vu to Store Company Accounting Policy Context for Claude
- Data: Organize company accounting policy manual (revenue recognition policy, expense classification standards, intercompany elimination rules, etc.).
- Action: Install Deja Vu (
pip install dejavu && dejavu serve), use CLI to store policy entries into memory bank, then connect via Claude Desktop with MCP. - Review: Senior Accountant verifies if Claude references correct policy entries when answering accounting policy-related questions.
- Output: Deja Vu memory bank + test Q&A record.
- Judgment Standard: If Claude’s answer accuracy improves from baseline by >30%, consider expanding to more contexts.
4 | Audit and Cut 3 “Zombie Reports”
- Data: Export team’s current all periodic report list.
- Action: For each report, ask three questions (see Top 3 Implementations Today item 3): who reads it, if stopped for 48 hours would anyone notice, is there more accurate data elsewhere?
- Review: FP&A owner and business head confirm which reports can be discontinued.
- Output: Zombie report cleanup log (report name, discontinuation reason, last reader confirmation).
- Judgment Standard: If 30%+ reports can be cut, team time release is significant, worth continuing.