AI automation in finance: clean up first, then automate.
In practice, automation in finance fails on master data, not on technology. So I work in two stages: first the data base, then one process with a metric set before the tool. For companies with EUR 4m to 100m revenue whose finance function runs on Excel and an ERP and can no longer keep up. No prior AI knowledge is needed.
Who is this service for?
For companies with EUR 4m to 100m revenue whose finance function runs on Excel and an ERP and can no longer keep up: the close drags, the board report costs days, and nobody knows which figure is right. You need no AI knowledge. You need someone who understands the chart of accounts and knows the tools.
What happens in stage 1, the data base?
Stage 1 takes 4 to 6 weeks and can be booked on its own. 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 must be set by someone who understands the chart of accounts. That is this stage, in five deliverables.
| Deliverable | What you get | Why it matters |
|---|---|---|
| Data maturity assessment | Where the figures live, how many Excel files sit between ERP and report, which reconciliations are done by hand | Every manual reconciliation is a source of error and an obstacle to automation |
| Master data and chart of accounts cleansing | AI tools classify and propose, the finance lead decides | In practice, automation fails on master data, not on technology |
| Process picture from system data | Process mining: how invoice processing, approvals and the close actually run, in numbers (cycle time, first-pass rate, rework loops) | Only the measured actual state shows which process is worth automating |
| Documentation | Procedures that used to live in one head are written down and executable | Removes the dependency on one person, including one outside the company |
| Prioritised list | What can be automated and what cannot, with a reason for each item | Prevents a tool being bought before the process is in place |
For this size class, ERP exports and lightweight tools are enough; the large suites are oversized. For many companies stage 1 is the whole mandate, because a clean data base also carries the ERP change and the next due diligence. What gets automated afterwards is decided by the prioritised list, not by a tool catalogue.
What happens in stage 2, the automation?
Stage 2 takes 60 to 90 days and automates one process. One, because one process with evidence is worth more than five half-finished ones. The metric is fixed before any tool is touched, the evidence follows afterwards, and at the end your own team takes over. Four steps, each with a result you can check.
| Step | Content | Result |
|---|---|---|
| Inventory | Which tools are in use, officially and unofficially, and which duties under Art. 4 of the EU AI Act follow from that | List of tools, data paths and duties |
| Process selection | One process from the prioritised list, with metric and target value | A written target the result is measured against |
| Implementation | Build with the team, procedural documentation under German GoBD rules, review by a second instance before delivery | A running process with evidence against the target |
| Handover | Training, documentation, ownership inside your own team | The process runs without me |
Where does your data stay?
The first question every decision-maker asks is: what happens to our data? Five rules apply in every mandate, and each has a reason. They apply before the first tool is used. Anyone who cannot keep one of these rules should not put company figures into an AI tool.
| Rule | Reason |
|---|---|
| No company data in public chat tools. | This is usually the current state, and ending it is 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 any use of the data for training. | Without these three commitments the data path cannot be controlled, and the data protection officer is right to object. |
| Where confidentiality requires it (investor reports, M&A, payroll data), models run as needed locally on the client's own hardware or in the client's own cloud. Nothing leaves the building. | That costs computing power and support, and I say so in advance. |
| Every automated figure remains traceable. | Procedural documentation under GoBD and auditability for the statutory auditor are part of the delivery, not an extra. |
| Where a works council exists, it is involved before the tool. | A tool that changes workflows is subject to co-determination, and a works council involved early carries the introduction. |
The enterprise platform I use is Langdock, a German provider that makes AI models from several vendors available in a protected environment for companies. It looks like a chat window. The difference lies in the contract and the data path: data processing agreement, servers in the EU, no use of inputs for training, a log of every step.
What principles do I work by?
Figures can be consistent and still wrong. So nothing goes to the client before a second, separate instance has taken it apart. Before delivery, a second, separate AI system reviews the work critically, without knowing how it was produced. It is not an auditor and replaces none. It finds what one's own eye misses after four weeks in a case.
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 I deliver. The reader does not have to guess whether a figure comes from a document or from an assumption, and the auditor sees it without asking.
Documentation is a deliverable, not a by-product. Every mandate produces a document that removes the dependency on one person: a control sheet with the monthly routine, or a knowledge document with chronology, pitfalls and an archive whose cross-references are checked by machine. Whoever comes after me can read the case without calling me.
In which formats is this available?
Three formats, depending on how far along you are and what you need. Billing is by days actually worked, within the corridor described in the guide on daily rates. You receive a concrete quote after a 30-minute initial call in which we clarify which format fits.
| Format | Duration | Result | Order of magnitude |
|---|---|---|---|
| Assessment | 2 to 3 days | Data maturity assessment, tool list, prioritised list, recommendation for stage 1 or 2 | 2 to 3 daily rates within the EUR 1,200 to 2,500 corridor |
| Implementation | 60 to 90 days | Stage 1, stage 2 or both, with evidence against the target value and handover | Daily rate for days worked, scope depends on the process |
| Component of an interim CFO mandate | Term of the mandate | Data base and automation as part of interim financial leadership | Included in the mandate, full-time or 3 to 5 days per week |
What has come out of it in mandates?
Two cases from two mandates since July 2026. One is measured, one is a prototype and is called that. For every figure it says what the machine did and what was financial judgement, because that is the difference between a tool and a result. Details with all figures are on the PeriGuard mandate page.
Auction settled after ten months open
Inventory list and lot list shared no common identifier; 31 of 239 lots could be matched. From 598 pages a register emerged with every invoice chain and 513 document links; 374 cross-references were checked by machine. The package was booked correctly, reconciled and closed by mutual agreement. Machine: read, link, calculate. Financial judgement: what to claim.
11 files, one linked workbook, one working day
The monthly report to the parent company came from 11 files without a single link. In one day over 250 manual transfers became formulas, 18 checks run automatically, both balance sheets reconcile to zero. Along the way 12 items still carrying prior-month values surfaced. Machine: read, link, check. Financial judgement: which inputs deliberately remain inputs. Anonymised; operating values to follow.
Frequently asked questions about AI automation in finance
Do I have to know in advance what I want to automate?
No. The prioritised list of what can and cannot be automated is a deliverable of stage 1. Whoever fixes a process in advance usually picks the most visible one, not the one with the largest effect. The process picture from the system data shows where the rework loops actually sit, and the selection follows from that.
Is stage 1 enough on its own?
For many companies, yes. A cleansed data base with documentation carries the monthly close, the next ERP change and the next due diligence, even without automation. Whether stage 2 follows is decided by the prioritised list at the end of stage 1: only a process with a measurable effect and clean master data is worth the effort.
What does AI automation in finance cost?
The assessment costs 2 to 3 daily rates within the corridor of EUR 1,200 to 2,500. The implementation over 60 to 90 days is billed by days worked, and the scope depends on the process. As a component of an interim CFO mandate the service is included in the mandate. Tool costs come on top and are named in advance.
Who decides what is booked correctly?
The finance lead, in stage 1 together with me. AI tools classify documents and accounts and propose assignments, often faster and more completely than a person. The rule for what is correct is set by someone who understands the chart of accounts and the accounting standards. This split appears in every report: what the machine did, what was financial judgement.
- Mandate page PeriGuard, section Data base and automation (auction, closed August 2026).
- Reporting package prototype of 2 September 2026, anonymised mandate; operating values after the first month.
- Daily rate corridor: What does an interim CFO cost?
- Art. 4 Regulation (EU) 2024/1689 (EU AI Act), AI literacy, applicable since 2 February 2025; Art. 28 GDPR; German GoBD.
Is your finance function no longer keeping up?
30-minute initial call, no obligation. You describe where the figures live and what does not add up; I tell you whether stage 1 is enough and what an assessment delivers. Reply within 24 hours.
Request an initial call