The Finance Workflows Anthropic Automates With Claude (And the Exact Prompts to Copy)

Anthropic's finance team uses Claude to automate five production finance workflows: AR reconciliation, FP&A reporting, ERP data validation, master data governance, and recurring scheduled reporting. This article explains each workflow, provides the exact prompt structures used, and shows how any finance team can implement the same approach using Claude Cowork, Claude Code, and Claude Skills.

Key Takeaways

• Anthropic's finance team runs over 150 reusable Claude Skills stored in a shared, version-controlled library

• Every automated workflow includes a mandatory human review step before anything touches the ledger

• Meta-prompting — using Claude to write your Claude prompts — is the step most teams skip and the one that most improves output quality

• The five workflows to automate first: AR reconciliation, FP&A variance commentary, ERP migration validation, master data governance, and scheduled reporting

• Skills encode institutional knowledge, so process expertise survives turnover and retirements

• Finance teams with no coding experience can run complex multi-step workflows using plain-language descriptions in Claude Code

What Are Claude Skills, Claude Cowork, and Meta-Prompting?

Before getting into the workflows, here are the definitions you need to follow along.

What are Claude Skills? Claude Skills are saved, reusable instructions that tell Claude how to perform a specific task. They can be stored in a shared library — Anthropic uses a GitHub repository — and distributed across a team as workspace plugins. A Skill for monthly variance commentary, for example, runs the same way for every analyst, every time, with no manual setup.

What is Claude Cowork? Claude Cowork is Anthropic's agentic AI environment for knowledge work. It allows finance teams to run multi-step automated workflows — connecting to data sources, processing information, and producing formatted outputs — without writing code. Think of it as the layer between a prompt and a production automation.

What is Claude Code? Claude Code is a coding environment within Claude that lets users build tools and automations using natural language. Crucially, Anthropic's internal case study documents that finance team members with no coding experience can describe workflows in plain text ("query this dashboard, get information, run these queries, produce Excel output") and Claude Code executes the entire sequence.

What is meta-prompting? Meta-prompting means using Claude to write your Claude prompts before you build an automation. Instead of drafting a prompt from scratch, you describe the finance task to Claude in plain language and ask it to generate a well-structured prompt for you. Three separate CFO Connect speakers — Christian Sanford, Sherilyn Kamga, and Kevin Steel — each arrived at this approach independently. It is the most-skipped step and the one with the highest impact on output quality.

What is Layer 2 vs. Layer 3 finance AI? Layer 1 is using Claude as a reactive chat tool. Layer 2 is building automated pipelines using Claude Cowork — where Claude takes actions across tools and data sources without manual triggering. Layer 3 is using Claude Code and Skills to build custom tools and encode workflows as reusable institutional assets. Anthropic's finance team operates at Layers 2 and 3. The progression is covered in detail in this CFO Connect session recap.

Why Is Finance AI Adoption Stalling — and What Does Anthropic Do Differently?

Finance teams are under more pressure than they have been in years. 84% of senior finance leaders now report an acute talent shortage. The average company has 17 open finance and accounting roles today, up from just 2 in 2024. 75% of current CPAs are nearing retirement age, taking decades of process knowledge with them.

At the same time, the work keeps growing. 73% of finance teams say the business is growing faster than they can keep up. Yet 45% of FP&A time is still spent cleaning and reconciling data — a ratio that EY's FP&A Trends research found has barely moved in seven years. Only 9% of FP&A teams report acting as true strategic partners to the business.

The AI tools exist to change this. 59% of finance leaders now use AI in some capacity (Gartner, 2025), up from 37% in 2023. But scaling that use is where teams consistently get stuck. BCG found that the median GenAI ROI in finance is just 10% against a 20% target, and only 33% of finance teams have managed to scale AI beyond individual experiments.

The teams that do scale share a common pattern: they stop treating AI as a chat tool and start encoding their workflows as reusable, shareable automations.

That is exactly what Anthropic's finance team has done. In June 2026, Tim Ross (Finance AI Product Lead) and Lisa To (Head of Finance Systems and Transformation Office) gave a rare live demonstration of their production finance workflows during an Anthropic webinar — with CFO Connect producing the session write-up. You can read the full breakdown here.

What follows takes that write-up and pairs each workflow with the exact prompt structure that powers it, drawing on prompts from CFO Connect's 25 Claude Prompts for Finance Teams session.

Before the Workflows: Why Does Meta-Prompting Matter?

The single most common reason finance AI automations underperform is a poorly structured prompt. Before building any of the five workflows below, use this prompt to generate a well-structured starting point:

"You are an expert prompt engineer specialising in finance workflows. Help me create an optimal prompt for [specific finance task]. Include: a clear role definition, specific output format requirements, handling of edge cases and exceptions, and the four elements of a well-structured finance prompt: input format, output structure, calculation logic, and exception handling."

Run this in Claude Chat, iterate two to three times, then take the output into Cowork or Code to build the automation. This step alone typically cuts debugging time in half.

What Finance Workflows Does Anthropic Automate With Claude?

Anthropic's finance team has identified five workflow categories where Claude Skills deliver the highest impact: AR reconciliation, FP&A variance commentary, ERP migration validation, master data governance, and scheduled reporting. Here is how each one works, with the prompt structure to replicate it.

How Does Claude Reconcile AR and the General Ledger?

The problem: Reconciliation is the single biggest bottleneck in month-end close. It consumes 20 to 50 hours per month for most finance teams. 50% of teams still take more than six working days to close the books (Ledge / CFO.com, 2025). Konica Minolta's bank reconciliation used to take 30 days; with AI-assisted matching across 45,000 monthly line items, it now takes 7.

For finance teams still running reconciliation manually, Spendesk's guide to spend control covers how automated transaction-tagging and reconciliation reduces manual entries and creates a clear audit trail — a useful complement to the Claude workflow above.

What Anthropic does: Tim Ross demonstrated a live AR sub-ledger to GL reconciliation where Claude ingests both data sets, identifies exact matches, flags timing differences likely to resolve, and escalates unexplained variances — producing a three-tier output before a human reviews. Nothing touches the ledger without sign-off.

The prompt:

"Compare these two data sets [paste AR sub-ledger] and [paste GL extract] and identify: 1. Exact matches, 2. Timing differences likely to resolve within [X days], 3. Unexplained variances requiring investigation. Format as a three-tier reconciliation report. Highlight unexplained items. Flag any items over [£/€/$X threshold] for immediate escalation."

Why the output format matters: Specifying "three-tier" forces Claude to categorise rather than summarise. Finance teams running this in Claude Cowork report processing a full month's transactions in the time it previously took to open the source files.

How Does Claude Automate FP&A Variance Commentary?

The problem: Revenue variance commentary is one of the highest-value deliverables finance produces and one of the most time-consuming to format. Most analysts spend the majority of their time on the mechanics, not the analysis. EY's research found that AI-enabled FP&A teams achieve 25% higher forecast accuracy and spend measurably more time on insight generation than their non-AI peers.

What Anthropic does: Lisa To demonstrated a workflow where Claude connects to BigQuery, pulls month-over-month revenue data, identifies the top drivers of variance, quantifies their impact, and outputs leadership-ready commentary — formatted for a board pack, not a raw data dump. The analyst reviews and approves; Claude builds.

The prompt:

"You are an FP&A analyst preparing month-over-month revenue variance commentary for a CFO presentation. Analyse this data: [paste data]. Identify the top 3 drivers of variance. Quantify their impact in both absolute and percentage terms. Draft concise commentary suitable for a CFO. Use professional financial language. Lead with the most significant driver. Output format: one executive summary paragraph, followed by a three-row table (driver / impact / status)."

External benchmark: BCG documented a consumer goods company using a comparable GenAI FP&A workflow that cut report generation time by 50% and produced forecasts 30% faster. Allianz Technology reduced their data integration workload by 60% applying the same principle.

Can Claude Validate ERP Migration Data?

The problem: ERP migrations are among the highest-risk finance projects a team undertakes. Data quality failures often only surface months after go-live. Manual validation of even one year of transactional data typically takes hundreds of hours and still misses edge cases.

What Anthropic does: Tim Ross shared a live validation of their NetSuite to Workday migration. Claude compared field-by-field mapping across one full year of financial data:

One year of data validated in approximately 20 seconds. Doing that manually would have taken hundreds of hours. — Tim Ross, Finance AI Product Lead, Anthropic

The prompt:

"You are a data migration validation specialist. Compare [source ERP export] against [target ERP import] for the period [date range]. Check for: 1. Field-level mapping accuracy across [list key fields], 2. Record count variances by entity and period, 3. Rounding or currency conversion discrepancies, 4. Missing or duplicated transactions. Output a validation report with a pass/fail status per field, a summary of exceptions by severity (critical / warning / informational), and a recommended resolution action for each exception."

For non-technical finance staff: Anthropic's own internal case study documents that finance team members with no coding experience can describe this workflow in plain text to Claude Code and have it execute the full sequence. No engineering involvement required.

How Does Claude Manage Finance Master Data Governance?

The problem: Cost centre changes, account name updates, and entity restructuring are low-frequency but high-stakes events. A single misapplied change can corrupt months of reporting and break downstream automations. Most teams manage this through spreadsheets, email chains, and tribal knowledge that disappears when the person who built it leaves.

What Anthropic does: Their team runs a Skills-based master data governance workflow that validates any change request against existing hierarchies, checks downstream report impacts, and routes for approval — before anything is applied in the system.

The prompt:

"You are a master data governance analyst. A change request has been submitted: [describe change]. Before approving: 1. List all reports and dashboards that reference this entity, 2. Identify any naming conflicts with existing structures, 3. Flag all child accounts or cost centres affected, 4. Confirm whether this change requires CFO sign-off under our governance policy [paste policy excerpt]. Output a structured impact assessment ready for reviewer approval."

How Can Finance Teams Automate Recurring Reports With Claude?

The problem: Daily cash positions, weekly KPI packs, and monthly board summaries are high-frequency, low-variance work. They should not require a skilled analyst's time. But in most teams, they do — consuming hours that could go to analysis.

What Anthropic does: Several of their 150 Skills are scheduled automations running on a fixed cadence. Claude pulls the data, applies the standard template, formats the output, and routes it to the right recipients. The team receives the report; they do not produce it.

The prompt:

"You are an automation architect for a finance team. I need to automate our [daily / weekly / monthly] [report type]. Design a step-by-step automation including: 1. Input data sources and access method, 2. Processing logic and calculations required, 3. Output format and distribution list, 4. Error handling if data is missing or out of range, 5. A validation checkpoint before sending. Present this as a Skill specification a non-technical finance analyst could hand to Claude Cowork to execute."

Sherilyn Kamga's CFO Connect session goes further, covering how Zapier integrations can trigger Claude automations from calendar events, Slack messages, or email alerts — removing even the manual trigger from the loop.

How Should Finance Teams Organise Their Claude Skills Library?

Running each workflow once is useful. Scaling it across a team of ten, twenty, or fifty people is the goal. Anthropic's approach:

• A version-controlled GitHub repository of Skills, each written in plain language

• Skills distributed as workspace plugins so every analyst sees the same library

• Every update tracked so the library improves over time, rather than drifting

The five-step execution model Tim Ross described applies to every Skill:

1. Claude fetches the relevant data

2. Claude processes and analyses it

3. Claude formats the output to specification

4. A human reviews and approves

5. The output is deployed or shared

Step 4 is non-negotiable. Nothing in Anthropic's finance function goes to the ledger, to leadership, or to an external party without human sign-off. The automation handles the mechanical work. Judgment stays with the team.

Your best preparer writes the procedure down once in plain language and it runs the same way for everyone, every time. This knowledge gets versioned, not lost when somebody moves on. — Tim Ross, Finance AI Product Lead, Anthropic

In a profession where 75% of current CPAs are nearing retirement and 84% of finance leaders report they cannot fill open roles, that Skills library is not a productivity tool. It is an institutional resilience strategy.

How Should Finance Teams Start Using Claude This Quarter?

Weeks 1–2: Run the meta-prompting prompt on your most painful recurring workflow. Most teams start with reconciliation or variance commentary. Refine the prompt in Claude Chat before touching Cowork.

Weeks 3–4: Run that one workflow in Claude Cowork. Time the before and after. Document the prompt in plain language, as if explaining it to a new team member.

Month 2: Add two more workflows. Start a shared Skills repository. Distribute to the full team.

Month 3: Review what is running and what is still manual. Set a Skills count target for the following quarter.

Anthropic did not build 150 Skills in a week. They built one, learned from it, and compounded.

Frequently Asked Questions

What finance workflows can Claude automate? Claude can automate a wide range of finance workflows, including AR sub-ledger to GL reconciliation, FP&A variance commentary and forecasting, ERP migration data validation, master data governance (cost centre and account changes), and recurring scheduled reporting. Anthropic's finance team currently runs over 150 automated workflows using Claude Skills across these categories.

Do you need to know how to code to use Claude for finance workflows? No. Anthropic's own internal documentation shows that finance team members with no coding experience can describe workflows in plain text to Claude Code and have it execute multi-step automations. Claude Cowork also allows non-technical users to run agentic pipelines without writing code. The main requirement is a well-structured prompt, not programming knowledge.

What is a Claude Skill and how is it different from a prompt? A prompt is a single instruction. A Claude Skill is a saved, reusable, shareable instruction set that runs a complete workflow the same way every time. Skills can be stored in a shared library (Anthropic uses GitHub), distributed across a team, and version-controlled so the library improves over time without losing prior knowledge. They are the difference between an individual using AI and a team running AI at scale.

Is it safe to use Claude for sensitive finance data and workflows? Anthropic's finance team includes a mandatory human review step in every automated workflow before any output touches the ledger, goes to leadership, or is shared externally. This human-in-the-loop design is central to their approach and is recommended for any finance team adopting Claude. Controls, audit trails, and approval routing can be built directly into Skill specifications.

How long does it take for a finance team to see results from Claude? Teams that start with one high-frequency workflow — typically reconciliation or variance commentary — typically see measurable time savings within the first two to four weeks. BCG found that finance teams using GenAI for FP&A workflows reduced report generation time by 50% and accelerated forecasts by 30%. Anthropic validated one full year of ERP migration data in approximately 20 seconds; the same task manually would have taken hundreds of hours.

What is the difference between Claude Chat, Claude Cowork, and Claude Code for finance teams? Claude Chat is Layer 1: interactive, reactive, useful for drafting and analysis but not automated. Claude Cowork is Layer 2: agentic pipelines that connect to data sources and produce formatted outputs automatically. Claude Code is Layer 3: the ability to build custom tools and automations using natural language — no coding skills required. Anthropic's finance team uses all three layers, with the majority of their production workflows running in Cowork and Code via Skills.

How do you use meta-prompting to improve Claude finance prompts? Meta-prompting means using Claude itself to write your prompt before you build an automation. Describe the finance task in plain language to Claude Chat, ask it to generate a well-structured prompt including role definition, output format, calculation logic, and exception handling. Iterate two to three times, then take the resulting prompt into Cowork or Code. Three separate experts at CFO Connect arrived at this technique independently. It consistently produces better outputs than drafting prompts from scratch.

Can Claude help with month-end close specifically? Yes. Month-end close is one of the highest-impact areas for Claude automation. The biggest bottleneck in most closes is reconciliation, which consumes 20 to 50 hours per month for the average finance team. Claude's AR reconciliation workflow — comparing sub-ledger to GL, categorising matches, flagging variances, and escalating exceptions — directly addresses this. Reddit reduced their close from 15 days to 6 using AI-native finance automation. Konica Minolta cut their bank reconciliation cycle from 30 days to 7.

These workflows are drawn from three live CFO Connect sessions. Read the full recaps: Claude for Finance Teams | 25 Prompts for Finance Teams | How Anthropic's Finance Team Uses 150 Claude Skills

Also from CFO Connect: Microsoft Copilot for CFOs: Real Workflows That Save Time