# AI Agents for Accounting and Controlling in Mid-Sized Companies

> AI agents take routine work off payables, receivables, closing, controlling and cash planning: reading documents, matching, suggesting account codes and explaining variances. Posting, approval and tax advice stay with people, and German GoBD rules require traceability and unaltered originals.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-agents-accounting-controlling  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-11-14

Finance teams in mid-sized companies handle a lot of work that needs neither judgement nor creativity: keying in invoices, matching payments, preparing dunning runs, explaining why a cost centre is over budget. That is exactly where AI agents fit. At the same time, few business functions are as tightly regulated as bookkeeping. This guide maps the finance areas where agents work well, explains how a finance agent is built, and sets out the limits that German bookkeeping rules, tax law and the role of the external tax adviser impose.

The examples refer to the German framework because many mid-sized companies with German entities work under it. Other jurisdictions have their own rules on audit trails and record retention, so check the local equivalents; the principles carry over. This is orientation, not legal or tax advice.

## What is a finance agent and what is it used for?

A finance agent is an AI system that completes accounting and controlling tasks in several steps on its own: it reads documents, queries the ERP or ledger, compares results and files a proposal. Unlike plain OCR, it checks context and explains its reasoning instead of just extracting fields.

The difference from a chat assistant matters. A chat assistant answers when asked. An agent works through a case. A supplier invoice arrives by email; the agent identifies the supplier, the purchase order and the goods receipt, suggests the account, cost centre and tax code, and places the proposed entry in the approval queue. A person reviews and confirms.

Common goals in mid-sized companies:

- **Less manual data entry** for documents that always look the same.
- **Faster resolution of discrepancies**, because the agent has already gathered the paperwork.
- **Variance reports people can read**, with explanations in plain sentences rather than bare numbers.
- **Lighter month-end and year-end close**, because reconciliations are prepared in advance.

How agents work across business functions, and which building blocks they share, is covered in the overview of [AI agents for business teams](https://www.ruebstudio.de/en/ai-agents).

## How does a finance agent work? The building blocks

A finance agent has four parts: a language model that understands documents, tools that give it controlled access to finance systems, a rule set for coding and approvals, and an audit log that records every step. In finance, the audit log is what makes an agent usable at all.

**1. The language model.** It reads PDF invoices, emails, bank statements and contracts, understands line items and payment terms, and writes explanations. Whether the model runs in the cloud or on your own hardware is a real decision for financial data; the trade-offs are discussed on the page about [sovereign AI on your own infrastructure](https://www.ruebstudio.de/en/sovereign-ai).

**2. Tools.** The agent calls narrowly defined functions: fetch open items, look up the purchase order for an invoice number, read bank transactions, create a draft entry. An agent that can only read and draft is much easier to sign off than one with full write access.

**3. Rules.** Chart of accounts, cost centres, approval limits, the four-eyes principle and house rules ("travel always goes to the traveller's cost centre") are fixed. The model works inside those rules and escalates when it is unsure.

**4. Audit log and storage.** Every proposal is stored with its inputs, reasoning, timestamp and confirmation. German law makes this non-negotiable: under [Section 146(1) of the Fiscal Code](https://www.gesetze-im-internet.de/ao_1977/__146.html), entries must be recorded individually, completely, correctly, promptly and in order, and Section 146(4) forbids changing an entry in a way that hides its original content.

Connecting to the finance system is the least glamorous part and often the most work. How an agent reaches DATEV, the German accounting software, and when a file export is enough, is covered in [connecting an AI agent to DATEV](https://www.ruebstudio.de/en/guides/ai-agent-integrations/connect-ai-agent-datev). This guide sits one level above that: which finance tasks justify an agent in the first place?

## Where agents fit: five finance areas

Five areas suit agents well: payables, receivables, closing, controlling and cash planning. They differ in how structured the input is and how costly a mistake would be. The more structured the data, and the more likely an error is caught before posting, the earlier an area is worth starting with.

### Accounts payable: capture, coding, approval

Payables is the most common starting point. The agent reads the invoice, matches it against the purchase order and goods receipt, suggests account, cost centre and tax code, and routes it for approval. It flags mismatches such as changed prices, missing receipts or duplicate invoice numbers, and says why.

E-invoicing changes the picture. Since 1 January 2025, businesses in Germany must generally be able to receive e-invoices; from 2027 companies with prior-year revenue above EUR 800,000 must issue them for domestic B2B sales, and from 2028 the obligation generally applies to everyone, according to the [IHK Cologne summary of the finance ministry's October 2025 letter](https://www.ihk.de/koeln/hauptnavigation/recht-steuern/steuern/bmf-schreiben-zur-e-rechnung-vom-15-10-2025-6774012). Under [Section 14(1) of the VAT Act](https://www.gesetze-im-internet.de/ustg_1980/__14.html), an e-invoice is issued, transmitted and received in a structured electronic format. With structured invoices, the agent spends less effort on recognition and more on matching, coding and plausibility checks. Paper and PDF invoices from foreign suppliers will still arrive, so recognition does not disappear.

### Accounts receivable: cash application and collections

On the receivables side, the agent matches incoming payments to open items even when the reference is incomplete, the customer pays in instalments or settles several invoices at once. For collections, it prepares dunning proposals and drafts clarification emails for disputed items, for example when a customer withholds payment over a complaint.

Be careful with credit decisions. Under [Annex III of the EU AI Act](https://artificialintelligenceact.eu/annex/3/), systems that evaluate the creditworthiness of natural persons are high-risk, with an exception for fraud detection. If you sell to consumers, that is a strong reason not to hand credit decisions to an agent.

### Closing: reconciliations and supporting documents

During month-end and year-end close, the agent prepares reconciliations: trial balance against sub-ledgers, clearing accounts, accruals and deferrals, comparison with the prior year. It collects documents for unusual items and writes a list of open questions for the finance team or the tax adviser. Closing entries and valuation decisions stay with people.

### Controlling: explaining variances

Controlling is where language models are most useful, because they connect numbers with words. The agent compares actuals against budget per cost centre, finds the entries behind a variance and writes an explanation ("materials over budget, driven by three consolidated invoices from one supplier in October"). The arithmetic should come from a database query or script, not the model; the model interprets the result.

### Cash planning: forecasts from open items and patterns

For cash planning, the agent combines open items, payment terms, recurring expenses and each customer's payment history into a forecast. It states its assumptions explicitly, such as a large customer continuing to pay late. The forecast supports management decisions; it is not a guarantee.

Deeper guides on individual areas, such as supplier invoice processing and process documentation under the GoBD, are planned; the build lists them at the bottom of this page once they are published.

## Prerequisites: data, systems, organisation

A finance agent needs three things: clean master data, technical access to the ledger, and an organisation that defines who reviews and approves proposals. Without all three, even a strong model will not produce reliable results.

**Data.** Chart of accounts, cost centres and supplier and customer master data should be maintained. Duplicate suppliers, outdated bank details and inconsistent cost centres lead to wrong proposals. Historical postings serve as examples; if they contain systematic errors, the agent learns those too.

**Systems.** The agent needs read access to open items, purchase orders and bank transactions, plus a place to file proposals: a posting batch, an invoice register entry or a task in the approval workflow. Where no API exists, a structured export is often enough to start.

**Retention.** Under the German GoBD, documents must be kept in the format in which they were received. After the amendment of 14 July 2025, keeping the structured part of an e-invoice is generally sufficient; the PDF part of a hybrid invoice only needs keeping if it holds tax-relevant extra information, and data extracted by OCR must also be kept once checked, as [KPMG summarises](https://kpmg.com/de/de/themen/2025/07/bmf-gobd-zweite-aenderung.html). For an agent, that means it never replaces or alters the original: it works on a copy and points back to the source.

**Retention periods.** Under [Section 147(3) of the Fiscal Code](https://www.gesetze-im-internet.de/ao_1977/__147.html), books, financial statements and the organisational documents needed to understand them must be kept for ten years, and accounting vouchers for eight. An agent's rules and logs explain how entries came about, so agree with your tax adviser how long to keep them.

**Organisation.** Decide who reviews the agent's proposals, up to which amount a single approval suffices, and who may change the rules. The process documentation must describe the agent as part of the workflow. An internal reference of coding rules and special cases helps; how that kind of knowledge becomes usable for agents is shown on the page on [knowledge management with AI agents](https://www.ruebstudio.de/en/ai-agents/knowledge-management).

## A step-by-step approach

The approach that works is gradual: pick one narrow process, let the agent only propose at first, compare its proposals with actual postings, and widen the scope only after that. Trust then rests on evidence from your own data rather than on vendor claims.

1. **Pick a process.** Choose something frequent and uniform with a clear check before posting, such as invoices from a stable group of suppliers or cash application.
2. **Document the current flow.** Who does what today, which systems are involved, what are the exceptions? This write-up doubles as the first draft of your process documentation.
3. **Write the rules down.** Coding rules, approval limits and escalation cases. Anything not covered goes to a person.
4. **Run in shadow mode.** The agent drafts proposals alongside the normal process, but nothing is posted. The team compares proposal and actual posting and records differences.
5. **Evaluate.** How often is the proposal right, which error types occur, where is master data missing? Only then decide whether the agent joins the live approval flow.
6. **Go live with approval.** The agent files proposals, a person confirms. How to build such approval steps is covered in the guide to [human-in-the-loop design for AI agents](https://www.ruebstudio.de/en/guides/ai-agent-architecture/human-in-the-loop-ai-agents). Spot checks and a monthly review of corrections remain mandatory.
7. **Expand.** More suppliers, more processes, later controlling reports. Each expansion starts in shadow mode again.

## Limits and risks

Finance agents hit three limits: legal responsibility, professional judgement and technical failure modes. Responsibility for proper books stays with the company, tax advice stays with authorised professionals, and language models can produce answers that sound right but are wrong.

**Responsibility and audit trail.** An agent does not change who is responsible for the books. It has to be built so that every entry remains traceable: who proposed it, who confirmed it, on what basis. An agent that silently adjusts entries after the fact conflicts with the immutability rule in Section 146(4) of the Fiscal Code.

**The tax adviser's role.** In Germany, only authorised professionals may provide tax assistance on a commercial basis, under [Section 5 of the Tax Advisers Act](https://www.gesetze-im-internet.de/stberg/__5.html). Companies that outsource bookkeeping or year-end work to an external tax adviser feel this directly. An agent can prepare data, collect questions and assemble documents. Whether an expense is deductible or how a transaction is treated for VAT belongs to the adviser or your in-house tax team. Involve the adviser early, because they will have to work with the output.

**Judgement calls.** Valuing receivables, setting provisions and handling accruals on ambiguous contracts are judgement calls. An agent can gather arguments; it cannot decide.

**Technical errors.** Language models sometimes confuse similar amounts, misread dates or draw overconfident conclusions from incomplete data, a failure mode known as [AI hallucination](https://www.ruebstudio.de/en/glossary/ai-hallucination). Countermeasures: do arithmetic outside the model, validate mandatory fields, match against purchase orders and receipts, and escalate rather than guess when uncertain.

**Privacy and confidentiality.** Finance data includes salaries, terms and personal data. Settle where the model runs, who can see the logs and whether a data processing agreement is needed. Answer those questions before launch and the agent can be operated in line with the GDPR.

**When it does not fit.** With very low document volumes, constantly changing suppliers without purchase orders, or neglected master data, the effort for rules and controls often outweighs the benefit. In that case, clean up master data or improve invoice intake first.

## How do I get started?

Start with an honest look at a single process: how many documents, how many exceptions, how good is the master data? The answer tells you whether an agent in shadow mode makes sense or whether the data foundation needs work first.

Three questions help you choose:

- **Where does routine eat the most time today?** Often supplier invoices or cash application.
- **Where would an error be caught before posting?** That is where an agent carries the least risk.
- **Who needs to be on board?** Finance, controlling, your tax adviser and, where relevant, data protection should be involved from day one.

If your team first needs a shared understanding of AI agents, a [hands-on workshop on AI agents](https://www.ruebstudio.de/en/workshop) that walks through a real finance process is a good first step. For comparison with other functions, from [quotation agents](https://www.ruebstudio.de/en/ai-agents/quotation) to sales, and for definitions of key terms, see the [AI glossary](https://www.ruebstudio.de/en/glossary).

## Frequently asked questions

### Should an AI agent post entries to the ledger on its own?

It can, but for most companies it should not. Responsibility for correct books stays with the company, so the agent proposes and a person confirms.

### Can a finance agent replace the tax adviser?

No. In Germany, providing tax assistance on a commercial basis is reserved for authorised professionals under Section 5 of the Tax Advisers Act. The agent prepares data; the tax assessment stays with the adviser.

### What data does a finance agent need first?

A clean chart of accounts, supplier and customer master data, open items, bank transactions and past postings as examples. Poor master data simply teaches the agent existing mistakes.

### Is it fine to let an agent score customer credit?

Not for consumers: assessing the creditworthiness of natural persons is a high-risk use under Annex III of the EU AI Act. For business customers the rules differ, but a person should still make the call.

## Sources

- [Section 146 German Fiscal Code (AO) – bookkeeping requirements](https://www.gesetze-im-internet.de/ao_1977/__146.html)
- [Section 147 German Fiscal Code (AO) – retention of records](https://www.gesetze-im-internet.de/ao_1977/__147.html)
- [KPMG: Second amendment to the GoBD (14 July 2025)](https://kpmg.com/de/de/themen/2025/07/bmf-gobd-zweite-aenderung.html)
- [IHK Cologne: Federal Ministry of Finance letter on e-invoicing, 15 October 2025](https://www.ihk.de/koeln/hauptnavigation/recht-steuern/steuern/bmf-schreiben-zur-e-rechnung-vom-15-10-2025-6774012)
- [Section 14 German VAT Act (UStG) – issuing invoices](https://www.gesetze-im-internet.de/ustg_1980/__14.html)
- [Section 5 German Tax Advisers Act (StBerG)](https://www.gesetze-im-internet.de/stberg/__5.html)
- [EU AI Act, Annex III – high-risk AI systems](https://artificialintelligenceact.eu/annex/3/)
