The finance function can no longer keep up. What AI changes about that, and what it does not.
Most finance functions in fast-growing mid-sized companies have not started with AI. This page describes what worked in two mandates since July 2026, where it failed, what it costs and who should judge it. Tools are neither recommended nor named. Every figure comes with its source and date.
Where is the blind spot?
Among German mid-sized companies, 20 % use AI, according to KfW Research in February 2026 for the years 2022 to 2024, based on over 6,800 responses in the KfW SME Panel. For medium-sized companies the ifo Institute reported 47.2 % in May 2026, for large companies 67.2 %. A finance function that has not started yet is in the majority.
The first blind spot sits inside the department. Individual employees use public chat tools with real figures, without rules and without the management knowing. That brings no measurable benefit, but a data protection risk and a duty of AI literacy under Art. 4 of the EU AI Act, in force since February 2025 and unknown to anyone in the building.
The second blind spot sits outside. In a 2026 mandate, the knowledge of how the monthly report to the parent company is produced sat in a single head at the external tax adviser. No written description existed; it was produced only on request. If that person is absent, the report is not produced, with or without AI.
Which stage is your finance function at?
Three stages are enough to place every finance function I have seen in mandates. The stage decides what makes sense next. Introducing a tool at stage 1 costs money and delivers nothing, because it works on data nobody trusts. The table names the marker of each stage and the next step.
| Stage | How you recognise it | What makes sense next |
|---|---|---|
| 1 · Unregulated | Figures sit in several Excel files between ERP and report; reconciliations are done by hand; individual employees use chat tools at their own discretion | Build the data base: inventory, cleanse master data and chart of accounts, rules for tools, documentation |
| 2 · Data base | ERP and report agree without manual reconciliation; someone maintains the chart of accounts; the rules for AI tools are written down; procedures are documented | Select one process, fix the metric before the tool, automate, keep the evidence |
| 3 · Automation | At least one process runs by machine with a metric and evidence; your own team operates it; every automated figure is traceable | Next process by the same rule; measure operating values; keep the procedural documentation current |
Why does it usually fail on the data?
AI does not repair a data base by itself. It finds duplicates, mispostings and deviations faster than any person. The rule for what is correct has to be set by someone who understands the chart of accounts. So in every mandate the data question comes before the tool question, and four questions usually answer it.
| Self-test: four questions you can answer without asking anyone | What a no means |
|---|---|
| Do ERP and reporting agree without manual reconciliation? | Every reconciliation is a source of error that no tool can automate without carrying the error along |
| Does someone maintain the chart of accounts? | Without maintenance, accounts grow that nobody can assign any more; every classification by a tool fails on them |
| How many Excel files sit between accounting and the board report? Do you know? | Every file is a break in the data flow; in one case there were 11 without a single link |
| Do you know what share of incoming invoices passes on the first run? | Without a first-pass rate there is no metric against which any automation has to prove itself |
Two noes mean stage 1. In practice, automation fails on master data, not on technology. In a 2026 mandate the automatic invoice recognition had been bought and failed on the master data: three systems, three spellings of the same supplier. Invoice numbers were altered when typed in; no later automation could have matched those documents.
In both cases since July 2026 the data base was the problem: 11 unlinked files for a monthly report and three spellings of the same supplier, two lists without a common identifier for an auction of 239 lots. The tools came in only once the data stood; before that they would have worked on figures nobody could check.
What concretely works today?
Finance functions of this size bring me five tasks. For each one the table below gives the starting point, what a machine can do, which order of magnitude is documented and which precondition applies. Where I have no mandate figure of my own, it says so. A figure from a vendor brochure does not replace a measured one.
The month-end close takes too long
| Starting point today | Delivery on calendar day 18, regularly on the last day; 1.5 of 2.5 positions tied up with emailing, renaming, forwarding and uploading around 5,000 documents a year (2026 mandate, measured) |
|---|---|
| What a machine can do | Invoice recognition, accrual routines, reconciliations, a check sheet with automatic controls |
| Documented order of magnitude | No measured reduction yet; the prototype has been running since September 2026, operating values follow after the first month |
| Precondition | Vendor master data in one system; e-invoice reception in place |
The forecast is never right
| Starting point today | Rebuilt every month, assumptions in the controller's head, deviations explained only afterwards |
|---|---|
| What a machine can do | Calculate driver logic, run scenarios, show deviations per driver; the assumptions stay with people |
| Documented order of magnitude | No mandate figure of my own. So no figure here |
| Precondition | A driver model the finance lead can explain |
Liquidity over 13 weeks
| Starting point today | Excel template, bank data and open-item lists typed in by hand, two hours of upkeep every week |
|---|---|
| What a machine can do | Read bank data and open-item lists automatically, roll due dates forward, show the change against the prior week |
| Documented order of magnitude | No mandate figure of my own. So no figure here |
| Precondition | Bank access via interface, maintained payment terms |
The report for parent, board or investors costs days
| Starting point today | 11 files without a single link, around 3,000 rows of personnel cost typed from PDF, four control cells for over 20,000 values |
|---|---|
| What a machine can do | Read the files, recognise the data flows, generate links, check sheet and log |
| Documented order of magnitude | Prototype in one day (2 September 2026): one linked workbook, over 250 manual transfers turned into formulas, 18 automatic checks, both balance sheets reconcile to zero |
| Precondition | A control sheet stating which inputs deliberately remain inputs |
An old case nobody understands any more
| Starting point today | Auction of around 240 lots, neither settled nor booked ten months after the sale; 228 files, 598 pages |
|---|---|
| What a machine can do | Read documents, link every figure to its document, build a register, recalculate every settlement line |
| Documented order of magnitude | Measured (PeriGuard, August 2026): continuous, auditable document and data package covering 239 lots, booked correctly, reconciled, closed by mutual agreement |
| Precondition | A finance lead who decides what to claim |
On the report in one day: the machine read the 11 files, recognised the data flows and generated the links, the check sheet and the log. Which inputs deliberately remain inputs and which balancing items need clarification was financial judgement. The prototype is built and reviewed but not yet in monthly operation; that is why it is called a prototype.
On the old case: the machine read around 600 pages and linked every figure to its document. What to claim and what to hold back, it did not decide. The negotiation route, the clarification meeting and the proposal to the other side were financial judgement. All figures with dates are on the PeriGuard mandate page.
What does not work?
The data base remains manual work; AI does not repair it by itself. When the 11 files were linked, 12 items surfaced that still carried June values in the July report, and four balancing items nobody could explain any more. The machine found them. Whether they stay, get corrected or dissolved was for someone who understands the balance sheet.
The judgement stays with a person. A set of figures can be internally consistent and still wrong. That is why nothing goes to the client before a second, separate instance has taken it apart. At PeriGuard the proposed asset disposal was arithmetically sound and still circular: a result-neutral back-calculation that confirmed itself. Only someone trained in accounting recognises that.
The responsibility stays with a person too. Every figure visibly carries how hard it is: documented, arithmetically confirmed, as stated by others, assumption, proposal, decided, open. These seven status labels appear in every report because a tool phrases an assumption as convincingly as a documented fact. Leaving the status out shifts the responsibility onto the reader, who does not notice.
What does it cost?
It costs days above all. An assessment takes 2 to 3 days, the data base 4 to 6 weeks, one automated process 60 to 90 days, billed by days worked within the corridor named in the guide on daily rates. Tool costs come on top and are named in advance.
A new ERP, a large suite or a data team are not part of it. For companies with EUR 4m to 100m revenue, ERP exports and lightweight tools are enough. The large platforms are oversized for this size and tie up capacity that is missing in the close. What remains after the mandate is operated by your own team.
Where does the data stay?
No company data in public chat tools; ending that is usually the first step of the work. I process client data only on an enterprise platform with a data processing agreement under Art. 28 GDPR, hosting in the EU and a contractual exclusion of use for training. Where confidentiality requires it, models run as needed locally at the client.
Every automated figure remains traceable, with procedural documentation under German GoBD rules and auditability for the statutory auditor. Where a works council exists, it is involved before the tool, because a works council involved early carries the introduction. The complete five rules with reasons are on the page AI automation in finance.
What does the EU AI Act require of the finance function?
Regulation (EU) 2024/1689 has applied in large parts since 2 August 2026; the duty of AI literacy under Art. 4 has applied since 2 February 2025. It covers every company that uses AI systems, even when that use happens only in an employee's chat window.
For the finance function this means: know which tools are in use, train the users, document the use, label AI-generated content where Art. 50 requires it. Fines for the most serious breaches reach EUR 35 million or 7 % of turnover. This is not legal advice; the duties in detail are in the guide on the EU AI Act.
How do you start?
Five steps in a fixed order, because each step provides the basis for the next. That is why the assessment comes first: after two to three days it is clear whether the rest is worth it. A tool at the start would automate a process nobody has measured, on data nobody trusts.
- Data and process inventory: where the figures live, how many files sit in between, which tools are in use officially and unofficially. Without this picture every decision is a guess.
- Set rules: no company data in public tools, one enterprise platform with a data processing agreement, training under Art. 4 of the EU AI Act. That ends the risk from the blind spot before any benefit arises.
- Metric before tool: at PeriGuard it was fixed before the register was built that every figure of the final statement must lead back to a document. In the end 513 links led to the documents and 374 cross-references were checked by machine. Without a metric there is no evidence.
- One process first: the first automated process has to deliver the evidence with which the management releases the second.
- Handover to your own team with documentation that removes the dependency on one person. The handover is complete only when the process runs without the consultant.
Who should judge this?
Someone without a product of their own, independent of the vendor, who understands the chart of accounts. The vendor can say whether its tool fits, the IT consultant whether the technology runs. Whether the figure is right afterwards and the auditor accepts it can only be judged by someone with accounting and VAT expertise who carries responsibility for the accounts.
In my mandates this judgement is part of interim financial leadership: first the data base, then one process with metric, evidence and handover to your own team. What that looks like as a service, with formats, data rules and the two cases, is on the page AI automation in finance.
Frequently asked questions about AI in finance
Do I need a new ERP for this?
No. In both cases since July 2026 the ERP stayed as it was; the work used exports and lightweight tools. Changing the ERP because of AI is the most expensive answer to a data question. If the change is due anyway, the cleansed data base carries it, see ERP implementation in a scale-up.
Is this only for large companies?
The two cases come from companies with EUR 20m and around EUR 40m revenue; in one case the finance function had 2.5 positions. Large companies use AI more often, 67.2 % against 47.2 % for medium-sized ones (ifo, May 2026). The difference lies in the tool: ERP exports and a register suffice where a group buys a platform.
Do we have to hire someone for this?
Not for the data base and the first process. Both are a fixed-term mandate with handover to your own team. What remains is operated by the finance lead with what is there: documentation, a control sheet, a process with a metric. A company of this size needs a data team only once several processes are running.
Our data is a mess. Do we have to clean it up first?
Yes, and that is the actual work. Master data and chart of accounts cleansing take 4 to 6 weeks, with tools that find duplicates and mispostings and propose assignments. Decisions are made by hand. At PeriGuard, 228 files became a register in which each of the 239 auctioned lots sits with its invoice chain.
Can I cut positions with this?
Rarely in the first year, and whoever starts there gets no process handed over. In a 2026 mandate, 1.5 of 2.5 positions were tied up with tasks a machine can take over. What frees up is time for the close, the forecast and talks with bank and board that currently get left behind. That is the benefit management feels.
How do I know whether my team already uses AI?
Ask, without reproach: who has used a chat tool for a task involving company figures in the last four weeks? Usually several people answer. Then the blind spot is confirmed: the benefit is unmeasured, the data is outside the building, the Art. 4 duty is open. The first step is a written rule.
- KfW Research, Fokus Volkswirtschaft No. 533, 11 February 2026: 20 % of German mid-sized companies use AI (reference years 2022 to 2024, KfW SME Panel, over 6,800 company responses).
- ifo Institute, business survey May 2026, published 5 June 2026: 54.5 % of companies use AI; large companies 67.2 %, medium-sized 47.2 %, small 51.2 %; 16 % plan to use it.
- PwC, The Digital CFO 2026, press release 26 May 2026: 207 CFOs surveyed, 157 of them in Germany; just over a fifth describe their finance function as strongly digitalised.
- Mandate page PeriGuard, data base and automation; anonymised 2026 mandate (reporting prototype of 2 September 2026).
- Regulation (EU) 2024/1689 (EU AI Act): Art. 4 since 2 February 2025, full application 2 August 2026, Art. 99 fines.
Which stage is your finance function at?
30-minute initial call, no obligation. You answer the four self-test questions, I tell you which stage you are at and what the next step costs. Reply within 24 hours.
Request an initial call