# Automating the finance function: RPA, AI, and the future of FP&A
In 2017, JPMorgan Chase quietly retired 360,000 hours of legal review work per year. The killer wasn't a layoff, it was COIN, the bank's Contract Intelligence platform, which reads commercial loan agreements in seconds rather than the weeks it took junior associates. By 2026, COIN's descendants ingest not just contracts but ISDA master agreements, KYC files, and intercompany transfer pricing documentation across 60+ jurisdictions affected by OECD Pillar Two. The lesson for CFOs is not that machines are coming for accountants. It's that the finance function, the cost center every CEO has tried and failed to compress for thirty years, is finally compressible. McKinsey's 2024 benchmark estimates 42% of finance activities are now fully automatable with existing technology, and another 19% can be substantially augmented. Companies executing well are reporting 25-50% cost reductions in transactional finance and, more importantly, 30-40% faster close cycles.
This lesson is about how to actually capture that value, not the slide-deck version.
Most CFOs inherit a confused vocabulary from their consultants. "AI," "RPA," and "intelligent automation" get used interchangeably in board decks, which is how you end up with $20 million pilots that automate the wrong process. Be precise.
RPA is rule-based screen scraping. UiPath, Automation Anyway, and Microsoft Power Automate dominate this space. A bot logs into SAP, copies a vendor invoice, validates it against a PO in Coupa, and posts it. It does not "think." It does not handle exceptions well. The economic case is straightforward: a typical RPA bot costs $5,000, $15,000 per year fully loaded and replaces roughly 2-3 FTEs of repetitive work in accounts payable, intercompany reconciliations, or bank statement processing.
Siemens deployed over 4,000 bots across its shared service centers between 2018 and 2024, automating a reported 2 million hours annually. The CFO organization, then led by Ralf Thomas, treated RPA as plumbing, not transformation. That framing matters. RPA is not strategic; it's a productivity floor.
This is where finance starts to look genuinely different. ML models classify transactions, predict cash collections, detect anomalies in expense reports, and extract structured data from unstructured documents, the contracts, invoices, and emails that defeat pure RPA.
BlackRock's Aladdin is the canonical example at scale. Originally built as a risk system in the 1990s, Aladdin now processes roughly $21.6 trillion in assets under management or administration as of 2025, running tens of thousands of Monte Carlo simulations daily on portfolios that include nearly 10% of the world's stocks, bonds, and loans. What makes Aladdin instructive for corporate CFOs isn't the asset management use case, it's the architecture. Aladdin is a single source of truth that the front, middle, and back office all touch. There is no reconciliation between "the trading book" and "the accounting book" because there is only one book. Most corporates run 8-15 finance systems that all disagree with each other at month-end. Aladdin's lesson: the automation problem is fundamentally a data architecture problem.
By 2026, the frontier has shifted again. Microsoft Copilot for Finance, Workday's Illuminate, and Oracle's Fusion AI AgentsAI AgentsAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.Voir la définition complète → now sit on top of the ERP, drafting variance commentary, building first-pass forecasts, and generating board-ready narratives from raw GL data. Unilever's FP&A team, under CFO Fernando Fernandez, publicly disclosed in early 2025 that generative AI cut its monthly management reporting prep from 9 days to 3.
The shift from copilots to *agents*, autonomous systems that take multi-step actions, is the 2026 story. An agent doesn't draft a journal entry for review; it posts the entry, alerts you only if confidence drops below a threshold, and learns from your correction. The control implications are non-trivial, which we'll return to.
The contrast between two well-documented transformations tells you everything about *how* to sequence this.
Siemens started with discrete, high-volume processes: accounts payable, travel expense audit, intercompany netting. By 2023, the company reported €100M+ in annualized run-rate savings from finance automation alone, with payback periods under 12 months on most bot deployments. The discipline was relentless process mining, using Celonis (a Munich-based company Siemens both uses and partially owns) to identify the actual, not theoretical, paths transactions took through SAP.
The critical insight: Siemens automated *after* it remediated process variance. Roughly 60% of the work in their early projects was eliminating unnecessary process steps before deploying any bot. Automating broken processes just gets you broken outcomes faster.
Microsoft's finance organization under CFO Amy Hood took a different path. Beginning around 2019, Microsoft rebuilt its forecasting capability around ML models that ingest 160+ external data signals, semiconductor lead times, hyperscaler capexcapexCapital Expenditure (CapEx) is money spent to acquire, upgrade, or extend long-lived assets like equipment, property, or software that deliver value over multiple years.Voir la définition complète → announcements, sovereign cloud regulations, even satellite imagery of partner data centers. The result, publicly discussed at multiple investor events: Microsoft's revenue forecast accuracy at the segment level is within 1.5% on a 90-day basis, materially better than analyst consensus.
The lesson is that Microsoft didn't start with cost takeout. It started with forecast accuracy because that's where finance creates strategic leverage. The cost savings followed, but they were a byproduct of building decision-quality data infrastructure.
Most CFOs should do *both*, Siemens-style cleanup on transactional processes funds the Microsoft-style investment in forecasting and scenario intelligence. The mistake is to do one without the other.
Vérification des acquis
1. According to McKinsey's 2024 finance automation benchmark cited in the lesson, what percentage of finance activities are fully automatable with existing technology?
2. A CFO is evaluating an RPA business case for AP invoice posting. Based on the economics described in the lesson, what is the realistic fully-loaded annual cost and FTE displacement per bot?
3. JPMorgan's COIN platform, referenced as a benchmark for finance automation, originally eliminated roughly how many hours of work annually, and in what domain?
4. Select ALL correct answers about Robotic Process Automation (RPA) as described in the lesson.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about the benefits companies are reportedly capturing from well-executed finance automation programs, per the lesson.
Sélectionnez toutes les réponses correctes.
Here is the uncomfortable truth that most VPs of FP&A have not yet internalized: roughly 60-70% of what their teams do today, pulling data, formatting variance reports, reconciling cubes, building decks, will be done by software within 36 months. The roles that survive and thrive shift from production to interpretation.
The "monthly close" is a 19th-century artifact built around when ledger books could physically be reconciled. By 2026, leaders like BlackRock, Stripe, and Shopify operate something close to a continuous close: financial position is known within hours, not weeks. Shopify's CFO Jeff Hoffmeister has discussed the company's "live P&L" capability, which allows merchant economics to be re-modeled daily as payment processing rates and ad CACCACCustomer Acquisition Cost (CAC) is the total sales and marketing spend divided by the number of new customers gained in a period. It measures how efficiently you grow.Voir la définition complète → fluctuate.
The CFO's monday-morning question: *What would it take to know my company's economic position by Tuesday of each week, not the 8th business day after month-end?* The answer almost always involves three things: collapsing the chart of accounts, automating intercompany eliminations, and pre-loading accrualsaccrualsAccrual accounting records revenue and expenses when they are earned or incurred, not when cash changes hands, giving a more accurate picture of financial performance.Voir la définition complète → from contract metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). rather than manual calculation.
Traditional budgeting allocates dollars. AI-augmented FP&A allocates *capacity to respond*. The mature pattern: rather than a single annual budget with quarterly reforecasts, leading finance organizations now maintain a rolling 18-month model with 50-200 driver variables, against which they run continuous scenario simulations.
This matters acutely in 2026 because of three converging pressures: OECD Pillar Two has made effective tax rate modeling a quarterly fire drill in many MNCs; CSRD reporting in the EU now requires forward-looking transition plan financials with quantified climate scenarios; and IFRS 16 lease modifications driven by hybrid-work footprint changes continue to generate balance sheet volatility. No human team manually models all of this fast enough. The companies that have invested in scenario engines, Anaplan, Pigment, and Oracle EPM with embedded AI, are simply faster decision-making organizations.
Every CFO reading this should pause on one risk. When an agentic AIagentic AIAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.Voir la définition complète → posts journal entries, who is the "preparer" for SOX purposes? When a generative model drafts the MD&A, who certifies that it isn't hallucinating a non-existent acquisition? The PCAOB's 2025 guidance and the SEC's amendments to Item 408 disclosures have started to require explicit disclosure of material AI use in financial reporting processes.
Practically, this means three controls every CFO should insist on before any production AI deployment in finance:
1. Deterministic logging, every model decision is reproducible from a stored input and model version.
2. Confidence thresholds with human routing, actions below X% confidence escalate to a human reviewer, with thresholds calibrated by materiality.
3. Drift monitoring, model performance is tracked against a holdout, and degradation triggers retraining, not silent failure.
If your CIO can't explain how your finance AI does these three things, you are not ready to deploy at scale.
Monday morning, here is what you do:
1. Commission a process mining diagnostic on your top 5 transactional processes within 30 days. Use Celonis, Microsoft Process Mining, or UiPath Process Mining. You will discover that your AP process has 47 variants, not the 3 your team described. Variance is the enemy of automation. Fix process before deploying bots.
2. Mandate that every finance system procurement decision pass a "single source of truth" test. If a new tool requires a new reconciliation, it should be rejected unless the strategic value clearly exceeds the integration tax. The Aladdin lesson is architectural discipline.
3. Move your FP&A team from a monthly to a rolling-forecast cadence within two quarters, even if the underlying tooling isn't perfect yet. The behavioral shift, talking about the future more than the past, matters more than the technology stack. The technology can follow.
4. Establish an AI Use Inventory and a tiered governance model in the next 60 days. Tier 1 (informational only, e.g., draft commentary): light review. Tier 2 (decision-supporting, e.g., forecast inputs): documented validation. Tier 3 (actioning, e.g., autoposting entries): formal control with SOX-grade evidence. This is what your auditors will ask about in 2027; build it now.
5. Reskill, don't replace, your top FP&A talent on prompt engineering, data modeling, and scenario design. The half-life of an Excel modeler is shrinking. The half-life of a finance professional who can interrogate an AI's assumptions and translate model output into a decision recommendation for the CEO is expanding. Budget $5,000, $10,000 per head per year for this. It is the single highest-ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → line item in your function.
The finance function that emerges from this decade will be smaller, faster, and considerably more strategic, or it will be diminished and outsourced. The CFOs deciding which way their organization goes are doing so right now, in the budgets they're approving for FY2026. Choose accordingly.