For a small business, AI in controlling means this: you hand an AI the figures from your management report, your trial balance or your spreadsheet and let it explain, compare and project. It translates ratios into plain language, finds the biggest gaps between plan and actual, builds a draft cash flow forecast from your inputs, runs scenarios and drafts questions for your tax advisor. What it does not do is guarantee correct figures. Language models can copy values wrongly, miscalculate or fill gaps with invented numbers. So you recalculate every total you need for a decision yourself, and the decision stays with you.
This post is about analysis and planning. Bookkeeping itself is covered in AI Bookkeeping: What Actually Works Today, working inside the spreadsheet in AI for Excel: Formulas, Analysis, Limits, and data analysis in general in AI Data Analysis Without a Data Scientist. In my business, Peter, my AI employee for receipts and bookkeeping, has been matching receipts against the credit card statement since August 19, 2026, and reports what doesn't add up. He doesn't book entries: his process only knows matching and reconciling, and final approval stays with me and my accountant. So controlling is not part of his process. The tool features below are as described on the vendors' help pages as of October 31, 2026, not from a test of my own.
What controlling means in a small business
In a small business, controlling is a handful of questions you answer regularly:
- Liquidity: Is there enough money in the account for the coming weeks, even if a client pays late?
- Contribution margin: What is left of a job or a service once you subtract the costs that only arise because of it, such as materials and subcontracted work?
- Plan versus actual: Where does the year deviate from the plan, and why?
- Planning: What do you expect for the next months in revenue, costs and payments?
- Ratios: How are the materials ratio, staff costs or gross profit developing compared with last year?
The basis usually comes from bookkeeping. In Germany that is the BWA (betriebswirtschaftliche Auswertung), a condensed monthly management report, and the trial balance (Summen- und Saldenliste), which shows every account that has postings with its balance. If your tax advisor sends you both monthly, you already have what an AI needs.
Where AI in controlling helps
Have the management report explained. Upload the report as a PDF or spreadsheet and ask: "Explain every line in one sentence and tell me which three figures changed most compared with last year."
Find variances. Put plan and actual side by side and let the AI sort the variances in euros and percent. It won't find the cause, but it shows you where to look.
Cash flow forecast as a draft. From your account balance, expected income and fixed expenses, the AI builds a monthly overview. It is a draft: only you know which invoice gets paid when.
Run scenarios. "What happens if my biggest client pays 30 days later?" or "What if material prices rise by ten percent?"
Prepare questions for your tax advisor. Have the AI turn the anomalies into a short list: "Why are other costs twice as high in September?" That saves time in the meeting, and the answer comes from someone who knows your bookings.
How to structure instructions like these is shown in Prompt Templates: 12 Copy-and-Paste Briefs for Small Businesses.
An example with made-up figures
An example: a landscaping business compares its third quarter with its plan. All figures here are made up; they only show what the analysis looks like.
A useful analysis states: the result is €14,000 below plan, mainly because revenue fell short while more was spent on materials and subcontracted work. The materials ratio rose from 30.0 to 34.3 percent of revenue. So the question for the business is: higher purchase prices, more rework, or jobs that were priced too tightly? No AI answers that; the jobs do.
Then liquidity, figures also made up: account balance on October 1 is €24,000. In October €62,000 comes in and €58,000 goes out, in November €48,000 in and €61,000 out, in December €41,000 in and €52,000 out. At year end, €4,000 is left in the account. Now the scenario: a client payment of €15,000, planned for December, only arrives in January. Then the account stands at −€11,000 at the end of December. This is where a cash flow forecast shows its value: you see the gap in October, not in December.
All totals in this example were recalculated with a small Python script. Even if the AI builds the table, you check at least the result and the closing balance yourself.
Where AI in controlling does not help
Inventing figures. If a value is missing, a language model sometimes fills the gap with a plausible number without saying so. So ask the AI explicitly to mark missing values as missing. How such errors happen is explained in AI Hallucinations: Why AI Makes Things Up.
Calculation errors. When a model answers without code, it calculates "in its head", and mistakes happen. According to their help pages, ChatGPT and Claude can write and run Python code; OpenAI itself advises reviewing the code, outputs and assumptions before relying on the result. When the AI calculates with code, the risk drops, but a wrong column or a duplicated month is still an error. Recalculate every total you pass on.
The wrong method. OpenAI writes that ChatGPT may choose a method that does not match your intent on the first try. So say exactly what you want: which columns, which period, net or gross.
For the same reason, Peter works by a fixed rule: amounts, dates and receipt numbers always come from the screen, never from memory. You can adopt that rule for every controlling analysis.
Tools for AI in controlling
Spreadsheets with AI. Microsoft explicitly describes Copilot in Excel for budgets, financial models and data analysis; Copilot works with tables, charts, PivotTables and formulas in the file and shows insights as charts, summaries, trends or outliers. At the same time, Microsoft's FAQ says Copilot can make mistakes and advises against using it for decisions in sensitive areas such as finance. If you already plan in Excel, start here, using Copilot for the analysis, not for the decision.
Claude or ChatGPT with file upload. OpenAI names .xls, .xlsx and .csv explicitly for ChatGPT. Anthropic writes that with Claude you can create financial models in Excel with working formulas and analyze uploaded data.
Your bookkeeping and tax advisor software. Which reporting or AI features your bookkeeping software or your advisor's software offers is in their product description. Ask your tax advisor which reports they can give you besides the monthly management report before you build something of your own.
Financial data: privacy and trade secrets
This isn't legal advice, just the operator's perspective. Clarify two questions before the first upload.
Are there people in the data? A condensed management report shows totals per line. Open-items lists or payroll reports, on the other hand, contain names. If an AI service processes such data on your behalf, Art. 28(3) GDPR requires a contract with the provider, the data processing agreement. Simpler still: replace names with numbers before uploading. More on this in AI and Privacy: What the AI Gets to See.
Are they trade secrets? Margins, cost calculations and client revenues interest competitors. Under the German Trade Secrets Act, information is only protected if it is the subject of secrecy measures that are appropriate under the circumstances (Section 2 No. 1 GeschGehG). A written rule on which financial data may go into which AI service is an obvious building block for that.
Anthropic also warns that with network access enabled, hidden instructions in files could trick Claude into sending data to external servers. Analyze files of unknown origin without it.
AI reporting in controlling: the same routine every month
The biggest benefit comes when one analysis turns into a fixed monthly report: every month the same management report, the same saved prompt. How to turn any analysis into a fixed report is covered in AI Data Analysis Without a Data Scientist. For controlling, three points come on top:
- Set thresholds. Write into the prompt which ratios, which comparison and from which variance something counts as "notable".
- Recalculate. Check at least revenue, result and closing balance against the source.
- Collect questions. Whatever stays unclear goes to your tax advisor as a list.
Frequently asked questions
Does AI replace the controller?
No. It takes over the legwork: explaining figures, sorting variances, calculating scenarios. Judging why a figure is what it is, and what follows from it, needs someone who knows the business.
Can I just upload my management report to ChatGPT or Claude?
Technically yes, both accept files. Check first whether names are in it, and which plan and contract apply. Treat your business figures as trade secrets; they are only protected if you take appropriate steps to keep them secret.
How accurately does AI calculate?
When it calculates with code, the risk of arithmetic errors drops compared with an answer without code. Errors still happen, for example from a wrong column, so recalculate every total you need for a decision.
What is the difference between AI in bookkeeping and AI in controlling?
Bookkeeping records what happened: receipts, entries, reconciliation. Controlling analyzes those figures and plans ahead. AI in controlling is only as good as the bookkeeping underneath.
Do I need dedicated software for this?
Not to start with. A spreadsheet with an AI feature or an AI chat with file upload is enough for explaining the management report, plan versus actual and a simple cash flow forecast.
Where to go next
Take your latest management report and have it explained line by line. Then ask for the three biggest changes compared with last year and recalculate one of them yourself. If it matches, build a simple three-month cash flow forecast next and play through one scenario: what happens if your biggest client pays a month later? Which other tasks you can hand off to AI is covered in Which Tasks You Can Hand Off to AI. We work through workflows like this step by step in my community AI Employee Framework.
