AI in Finance: Where to Start and How to Leverage the Finance Function
It's D+6 of the close. The CFO asked for the board pack yesterday. A controller is exporting the trial balance to Excel for the third time this week, because someone booked a correction after the first export. Somewhere in the company a Claude or Copilot licence sits unused, bought after a conference talk about "AI in finance".
We see this scene in almost every finance team we meet. The problem is rarely the AI. It's the order in which companies try to adopt it.
Short answer: start AI in finance with the foundations, not the tools. Clean your source data, document your workflows and connect one governed finance dataset to AI. Then automate the 20% of tasks that eat 80% of your team's time - starting with a skill, growing it into an agent, and only then building an app.
Below is the order that works, based on what we've built at incro for finance teams in Poland, Germany and the US: three foundations, a lightweight way to build, and six use cases where AI already saves measurable hours - AP, reconciliations, variance analysis, flux analysis, month-end commentary and forecasting.
Foundation 1: Clean the source before you add AI
AI reads your ledger literally. Give it a messy ledger and it will produce confident, well-formatted, wrong answers. So the first project is not an AI project at all. It's housekeeping, and it pays off long before any model touches the data.
Processes. Every recurring process needs one owner, one cut-off and one definition of done. When two accountants book the same accrual in two different ways, AI learns both versions and trusts neither.
Chart of accounts (CoA). The most common blocker we find is a CoA with 1,200 accounts, half of them dormant and a quarter called "other". Consolidate unused accounts, name them by economic substance and map every one to a reporting line in the P&L, balance sheet and cash flow. Without that mapping AI cannot explain anything above transaction level.
Cost centre allocation. Allocation keys - headcount, revenue, square metres, hours - should be defined and written down. If someone changes the key manually in a spreadsheet every month, AI cannot tell whether a cost moved because of the business or because of the key.
Consolidation models. Intercompany eliminations, FX rules and entity mappings belong in one model, not in five files. Group-level analysis is where AI gives the most leverage, and where a broken consolidation does the most damage.
A quick test we use with clients: can a new analyst trace any number in the management report back to its source in under 10 minutes? If not, AI won't be able to either.
Foundation 2: Map your finance workflows
AI can only automate work that is written down. In most finance teams the real process lives in one senior accountant's head and in last month's email thread. When that person is on holiday, the close slows down. When you try to hand the work to AI, it stops completely.
Two documents fix this.
A comprehensive month-end close checklist. Not 15 headlines, but every task with its owner, input, output, deadline (D+1, D+3) and dependency. A typical mid-size company ends up with 80-150 close tasks. Once they're listed, the candidates for AI become obvious: repetitive, rule-based and data-heavy.
Outcome-oriented workflow documentation. Describe each workflow by what "good" looks like, not by which buttons to click. Compare:
- Click-based: "Export the AP aging, paste it into tab 3, filter by vendor."
- Outcome-based: "Every invoice above 10k PLN older than 60 days is flagged with an owner and a reason by D+2."
The second version is a specification an AI agent can execute and be tested against. The first one breaks the day your ERP moves a menu.
Foundation 3: Build one finance dataset and connect it with MCP
AI needs direct, governed access to your numbers. Uploading a CSV to a chat window gets you an answer. It doesn't get you an answer you can trust, because nothing guarantees the data is complete, current and traceable to source.
One finance dataset, on infrastructure you already use. Bring the GL, AP/AR subledgers, VAT register, cash and bank, budget and key operational drivers into one model, with the same account mapping and cost centre structure you cleaned in Foundation 1. Where it lives matters less than how well it's governed:
- Microsoft - a data warehouse on Microsoft Fabric or Azure (Azure SQL, Synapse), a natural choice for companies already on Microsoft 365 and Power BI.
- AWS - a warehouse or relational database on Amazon Web Services (for example Redshift or PostgreSQL on RDS), common in tech companies whose engineering teams already run on AWS.
- A lighter setup, such as a PostgreSQL database, works well for mid-size companies that don't need a full warehouse yet.
How this looked at Base.com. Base's financial data lived in a legacy ERP (Comarch Optima) that was never designed to be queried by dashboards, let alone AI. The finance team wasn't short on data - it was short on a dataset. We built a PostgreSQL database fed directly from the ERP, merging five data sources (GL, opening balance, VAT register, cash and bank, AP/AR) into one governed system, validated table by table against the ERP and deployed on Base's own AWS infrastructure. Today it's the foundation for both reporting and Claude working directly on the numbers. Read the full Base.com case study.
An MCP connector on top. The Model Context Protocol (MCP) is an open standard, introduced by Anthropic, that lets AI models such as Claude query your systems directly and securely. Instead of pasting data into prompts, the model asks the dataset "show GL movements on account 4010 for October by cost centre" and gets the real answer. What it gives finance teams:
- Fresh data - every answer comes from the current ledger, not last week's export.
- Access control - the connector exposes only the tables and roles you allow.
- Traceability - each answer can be traced back to the query and source transactions, which auditors will ask for.
- Reusability - one connector serves every use case in this article, from AP to forecasting.
We wrote more about this in Why your finance agents need MCP.
Why AI in finance fails without strong foundations
Without clean data, mapped workflows and a governed dataset, AI doesn't fail loudly. It fails quietly, with outputs that look professional and are not useful.
We've seen it many times:
- A variance commentary explains a 300k PLN cost increase with "higher activity", because the cost was booked to the wrong cost centre.
- A reconciliation agent matches 95% of items and silently misses the 5% that matter.
- A forecast is built on 24 months of history that include a one-off restructuring nobody tagged.
In finance, a plausible wrong answer is worse than no answer. It costs the CFO's trust after the first board pack, and the project dies there. AI amplifies whatever you give it: clean foundations turn into leverage, messy ones into errors at scale.
Build: automate the 20% of work that takes 80% of the time
With the foundations in place, the temptation is to start with the most impressive use case. Don't. Start with the most painful one. Go back to your close checklist and score each task on three questions:
- How many hours per month does it take?
- Is it repetitive and rule-based?
- Is the input data already in your finance dataset?
Tasks that score high on all three are your first AI projects. In almost every team we work with, the same handful comes up: invoice processing, reconciliations, variance explanations and commentary writing. Together they often take 60-80% of close-related hours, while being the least strategic part of the job.
Pick one, and measure how long it takes today before you build anything. Without a baseline you can't prove the saving later - and a proven saving is what gets the next use case funded.
Start light: skill first, then agent, then app
The fastest way to burn a finance AI budget is to start by building an app. Build in three steps, and move to the next one only when the previous one works.
1. Skill (days to build). Written instructions, rules and examples the AI follows for one task, e.g. "how we write variance commentary". An analyst runs it inside Claude. Move on when the output is right 9 times out of 10.
2. Agent (weeks to build). The skill plus tools: the MCP connector, the ability to query data, check its own results and run several steps on its own. It runs on a trigger or schedule and a person reviews the output. Move on when it saves measurable hours for 2-3 close cycles.
3. App (months to build). The agent with a user interface, permissions, logs and integrations, used by the whole finance team as a product.
Anthropic has built this progression into Claude. Skills (Agent Skills) package a team's instructions and examples so Claude applies them consistently. Claude Projects give a team a shared workspace with its own context. And the Claude Agent SDK and Claude Code let you turn a proven skill into an agent that runs on a schedule against your MCP-connected data.
A skill is cheap to test and cheap to throw away. More importantly, it forces your best people to write down their judgement - and that written judgement is the real asset. Most of the value of the agent and the app comes from the quality of that first skill.
Document every skill and agent, and keep them in one library
A skill that lives in one analyst's chat history is just a clever prompt. When that analyst leaves, it leaves with them. Treat skills and agents like any other finance process: documented, owned and versioned.
What to document for each skill or agent:
- Purpose and the outcome it delivers (the outcome-oriented description from Foundation 2)
- Owner in the finance team and who reviews its output
- Data sources and the MCP tools it can use
- Rules, thresholds and examples of good output
- How it was tested, and a change log with version numbers
Where to keep the library. Pick one place the finance team can actually open, browse and update without asking IT:
- Claude - if your plan allows sharing skills and projects across the organisation, this is the simplest option: the library sits where the work happens.
- GitHub - the best choice for agents with code, because you get version control, reviews and a clear history of every change. Base.com's data foundation, for example, was handed over with full documentation and codebase in a GitHub repository.
- SharePoint - often the most practical home for finance teams on Microsoft 365: familiar, permission-controlled and easy for non-technical users.
The tool matters less than the rule: one library, one owner per skill, and nothing goes into production undocumented.
Decide to scale only after you see results
Scaling is a decision, not a default. After two or three close cycles with a working agent, sit down with the team and answer four questions:
- Time: how many hours did it save against the baseline you measured?
- Quality: how many outputs needed correction, and were the errors caught in review?
- Adoption: does the team use it without being reminded?
- Reuse: can the same dataset, connector and skill pattern serve the next use case?
If the answers are good, scale: turn the agent into an app, roll it out to other entities or add the next use case. If they aren't, go back to the foundations. In our experience the problem is almost always data or process definition, not the model.
6 AI use cases in finance that work today
Each use case follows the same pattern: clean data, a documented workflow, a skill, then an agent with access to the finance dataset through MCP.
1. Accounts payable (AP) automation
The AP team at a growing company often spends most of its week typing: vendor, amount, VAT, cost centre, repeat. An AI agent reads incoming invoices, extracts the data, matches them to purchase orders and receipts, and proposes the GL account and cost centre based on history. Exceptions - a price difference, a missing PO, a new vendor - go to a person with an explanation attached.
Foundation it needs: a clean vendor master, coding rules per cost category, an approval matrix.
Typical result: the AP team moves from data entry to exception handling. More in How AI agents can run the AP process.
2. Account reconciliation
An agent compares GL balances with bank statements, subledgers and intercompany counterparties, matches items with fuzzy logic (partial payments, different references) and lists unmatched items with a likely cause.
Foundation it needs: consistent account mapping, intercompany rules defined in the consolidation model.
Typical result: reconciliations prepared before the accountant opens the file, with time spent only on real differences.
3. Variance analysis (budget vs actual)
AI breaks budget-vs-actual variances into drivers - volume, price, mix, FX, timing, one-offs - by drilling from P&L line to cost centre to transaction. It also separates real business changes from booking errors.
Foundation it needs: a budget built on the same CoA and cost centres as actuals, defined materiality thresholds.
Typical result: a driver-based explanation of every material variance on D+3 instead of D+8.
4. Flux analysis
Flux analysis compares balances period over period (month on month, year on year) and explains significant movements. An agent scans every account against your thresholds, pulls the transactions behind each movement and drafts the explanation - including balance sheet accounts that rarely get attention.
How this looked at Paessler. Paessler, the Nuremberg-based maker of the PRTG monitoring platform, has a 13-person finance team across Warsaw and the US. Before working with us, most of their monthly workflows ran by hand in Excel, and AI use ranged from daily Claude users to people who had never opened it. In a four-week AI-in-finance bootcamp, the team built more than 10 working Claude workflows on their own data. One of them was an actual-vs-actual flux analysis that flags variances at account, vendor and department level, built for their July close. On the tasks participants automated, the first re-run was 50-65% faster. Read the full Paessler case study.
Foundation it needs: documented thresholds (e.g. above 10% and above 50k PLN), tagged one-off transactions.
Typical result: full account coverage instead of sampling, and an audit-ready flux file.
5. Month-end commentary
Variance and flux analysis feed straight into management commentary: a written narrative for the CFO, board or investors, in your company's tone, with numbers that tie to the reports. A skill holds the format, style and rules of what counts as material.
Foundation it needs: variance and flux analysis already working, examples of good past commentary.
Typical result: the controller edits a draft instead of starting from a blank page at 11 pm.
6. Forecasting
AI builds rolling forecasts from historical patterns combined with business drivers - pipeline, headcount plans, contracted revenue, seasonality - and runs scenarios on request ("what if EUR/PLN rises 5%?"). The finance team owns the assumptions; the agent does the recalculation.
Foundation it needs: clean history with one-offs tagged, a driver-based model, agreed forecast assumptions.
Typical result: forecasts updated monthly or weekly instead of quarterly, with scenarios in minutes.
FAQ: AI in finance
Where should a CFO start with AI in finance?
With data and process foundations: a clean chart of accounts and cost centre structure, a documented month-end close checklist, and one finance dataset connected to AI through an MCP connector. Only then automate the most time-consuming tasks.
What are the best first AI use cases for a finance team?
Accounts payable, account reconciliation, variance analysis, flux analysis, month-end commentary and forecasting. They are repetitive, data-heavy and take up most of the close calendar.
Which data warehouse should a finance team use for AI?
The one your company already runs on. Microsoft Fabric or Azure suits companies on Microsoft 365 and Power BI; AWS suits tech companies whose engineering runs there; a PostgreSQL database is often enough for mid-size firms. Governance and validation against the ERP matter more than the platform.
What is an MCP connector in finance?
A connector built on the Model Context Protocol, an open standard from Anthropic, that gives an AI model secure, permission-controlled access to your finance data, so it queries live numbers instead of exported files.
Should we build an AI app or an agent first?
Neither. Start with a skill - documented instructions the AI follows for one task. Turn it into an agent when it works reliably, and into an app when the agent proves measurable savings.
Where should finance teams store their AI skills and agents?
In one shared, documented library the finance team can manage: in Claude if your plan supports organisation-wide sharing, in GitHub for agents with code and version control, or in SharePoint for teams on Microsoft 365.
Why do AI projects in finance fail?
Mostly because of poor data quality and undocumented processes. AI then produces answers that look right but aren't usable, and the team loses trust.
Summary: foundations first, then build light, then scale
AI in finance is not a tool purchase. It's a sequence:
- Clean the source - processes, chart of accounts, cost centre allocation, consolidation.
- Map the workflows - a full close checklist and outcome-oriented documentation.
- Connect the data - one finance dataset on Microsoft, AWS or a lighter setup, with an MCP connector on top.
- Build light - pick the 20% of tasks that take 80% of the time, start with a skill, grow it into an agent.
- Document and share - every skill and agent in one library the finance team owns.
- Scale on evidence - only when hours saved and output quality are proven.
Go back to the controller exporting the trial balance on D+6. With this sequence in place, the reconciliations are prepared before she opens them, the flux file is drafted overnight and the commentary is waiting for her edits. Teams that skip the sequence get impressive demos and no change in the close calendar.
Want to know where your finance function stands? At incro we build finance data foundations and deploy Claude-based skills and agents for AP, reconciliations, variance and flux analysis, commentary and forecasting - for companies like Base.com and Paessler. Book a 30-minute call → or start with our AI Finance Readiness Assessment to find the first use cases worth automating in your company.
Your financial data won't fix itself.
30 minutes. We'll tell you exactly where your data is costing you money — and what AI can do about it.


