Blog · 5 November 2026 · 6 min read
Does local AI work on Excel files or just read the text?
Anyone who has tried asking an AI assistant for a total knows this: sometimes the figure is right, sometimes it is made up with great confidence. With company data that is not acceptable. What makes the difference is how the AI treats the file: as text to read or as data to run the numbers on.
In short: yes, a local AI really can work on the data in an Excel file. Instead of "reading" the cells and estimating the result, it writes a small program and runs it on the file: sums, filters, period-on-period comparisons and charts come from a real calculation, carried out on the company's computer. It does, however, have to be planned for and configured: it is not the default behaviour. On very large files or files with many linked sheets you have to tell it where to look, and the result should always be reread.
What is the difference between reading an Excel file and working on it?
They are two different functions, and they give very different results. An Excel file can go into the AI in two ways:
| Aspect | Read as a document | Used as data |
|---|---|---|
| What the AI does | It looks for the rows most relevant to the question and reads them | It writes and runs a calculation on the whole file |
| Good for | Finding a piece of information: "what discount does this customer have?" | Doing calculations: "revenue per sales agent in the third quarter" |
| Risk | Wrong totals: it sees only part of the rows | A badly set-up calculation, which can nevertheless be checked |
| What you need | The document archive | An environment in which the AI can run the calculations |
The first way is document search: excellent for text, unsuited to calculations. A language model does not add up ten thousand rows "in its head": either it sees only part of them, or it guesses. The second way gets round the problem, because the calculations are done by a program and not by the model.
How, in practice, does it do the calculations?
It writes the instructions, runs them and reads the result. Open WebUI allows the model to write and run Python code during the conversation. You upload the file and ask the question in plain English; the AI opens the file, looks at how the columns are laid out, writes the calculation, runs it and reports the result back to you, with a chart if you ask for one.
All of this takes place on the company's machine: the file is not sent to any external service.
There are two levels, and it is worth knowing which one you are being offered:
- The built-in engine, ready with no configuration, which runs the code with a fixed set of tools. It is enough for simple analyses, but the documentation itself describes it as a solution that is now outdated and limited.
- A dedicated working environment (Open Terminal), which Open WebUI recommends for real work: the AI has a small isolated computer at its disposal in which to load spreadsheets and CSVs, install whatever is needed, correct its own mistakes and return a finished report.
To work on the company's Excel files you need the second one. It is one of the things to have written into the quote.
What can I ask it?
What you would ask a colleague who is good with Excel, provided the data is in the file:
- Sums and counts: total orders per customer, number of open lines, average per month.
- Filters: only items below minimum stock, only invoices more than 60 days overdue.
- Comparisons: this year against last year, one price list against another, two exports of the same list to see what has changed.
- Charts: trend over time, breakdown by category.
- Clean-up: finding duplicates in a master data list, standardising dates and codes written in different ways.
- A new file with the result, to open in Excel.
Tip: always ask it to explain how it did the calculation: which columns it used, which rows it excluded. If the method is right, the number is right. It is a check you cannot make with an answer that was merely "read".
Where it gets things wrong or needs help
- Very large files or files with many linked sheets. Faced with twenty sheets that refer to one another, the AI may start from the wrong one. It needs pointing in the right direction: "use the Orders sheet, columns A to H".
- Sheets designed for the eye rather than for data. Multi-row headers, merged cells, totals in the middle of the table and colours that "mean something" confuse any program. A clean table, with one row per item, works much better.
- Formulas and macros. The AI works on the values it finds in the file. It doesn't run macros, and the logic hidden in complex formulas may escape it.
- Ambiguous questions. Is "turnover" with or without VAT? By order date or invoice date? If you don't say, it chooses for itself, and not always as you would have chosen.
- The model. Writing a correct calculation takes a model capable of reasoning and of using tools: small ones get it wrong more often.
- The final check. For a figure that ends up in a set of accounts or a quote, a spot check remains mandatory.
Is it safe to let the AI run programs?
Yes, if the environment is isolated. An AI that runs code must do so in a closed space, separate from the rest of the computer and from the data it does not need. The Open WebUI documentation states exactly this: run the workspace in an isolated container, so that what the AI does stays confined inside it.
It is a choice made by whoever installs it, and should be asked about explicitly: where the calculations are run, which folders that environment has access to, which users can use it.
How we configure it
We include the isolated computing environment in the project from the start, test it on your real files during installation, on site at the client's premises and throughout Italy, and in the training we show the two things that make the difference: how to prepare a sheet so that the AI understands it and how to have the calculation explained to you. If the data comes from the management system, we also set up the exports in the most suitable form.
The system remains the company's, which uses it independently. You will find the full picture on the Local AI for businesses page; if you would like to try it on one of your own files before deciding, write to us.
About the author: Dabryx
Dabryx is an Italian software house that installs local AI in companies, on site across Italy, on servers or Macs. The basic installation starts at €2,900, hardware not included, and consulting at €100 per hour. The office is in Rocca Priora, near Rome.
We also install local AI in Rome, Milan, Parma and Reggio Emilia, and we develop custom management systems and custom software.
Frequently asked questions
Can a local AI do calculations on an Excel file?
Yes, if it is configured to run code: it writes a small program, runs it on the file on the company's computer and reports the result. Sums, filters, comparisons and charts come from a real calculation.
Why does the AI sometimes get the totals in an Excel file wrong?
Because in that case it is reading the file as text: it sees only part of the rows and estimates the result. For calculations it has to work on the file as data, running the calculation.
Are my Excel files sent to an external service?
No. With a local AI both reading the file and running the calculations take place on the company's machine.
Can the local AI create a new Excel file with the results?
Yes, in the dedicated workspace it can save the result in a new file to open in Excel. Elaborate formatting and macros need reviewing by hand.
Want an estimate for your company?
Tell us how many people will use the AI and what for: we'll tell you which configuration you need.
Response times
Within 2 hours on business days (9:00–18:00) for clients with a support contract.
AI consulting
Basic installation from €2,900, consultancy from €100 per hour, on your company's server or Mac.
Where we work
Throughout Italy: we come on site to the client's premises and also support you remotely.