We Built an AI File Converter – Here's How

The problem: manual data shuffling

Here's a scenario that's more common than it should be: export from a WMS arrives as XLSX, needs to become XML for Pohoda accounting. Someone opens it in Excel, manually maps columns, reformats dates, saves as XML, imports into Pohoda. Repeat every week. Nobody codes a custom integration because it's "too small" – but over a year it adds up to days of someone's time.

The same pattern shows up everywhere: CRM exports that need to match ERP import templates, supplier price lists that need to be normalized to your product catalog format, logistics reports that need to feed a BI dashboard. Every system has its own format, and the glue between them is a person with Excel.

The solution: describe the output, AI handles the mapping

The idea behind Adaptine File Converter is simple: upload your source file, upload a sample of what the output should look like (or describe the target format), and let AI propose the column mapping. No coding, no IT ticket.

The user reviews the proposed mapping, adjusts anything that needs it, and runs the conversion. The result is a ready-to-import file in the target format. Supported formats: CSV, XLSX, XML, JSON.

How it works technically

The core is the Claude API. When a file is uploaded, we send the column headers and a few sample rows to Claude with a structured prompt asking for a JSON mapping between input and output fields.

Claude analyzes field names, sample values, and the target format definition, then returns the mapping. Take the WMS-to-Pohoda case: source columns are item_code, qty_shipped, unit_price – target XML expects <KodPolozky>, <Mnozstvi>, <JedCena>. Claude maps these correctly even across different naming conventions, because it understands the semantics of the values, not just the names.

The JSON mapping Claude returns drives the actual file transformation. We keep Claude out of the conversion step itself – it only proposes the mapping, our code does the actual file work. This keeps outputs deterministic and easy to audit.

What we learned building it

• Full mapping in one prompt beats iterating column by column. Asking Claude for the complete mapping in a single call produces more consistent results than building it incrementally – the model can use the full context to resolve ambiguities.
• Sample values matter more than headers. Sending a few rows of actual data (not just column names) dramatically improves mapping accuracy, especially when names are cryptic abbreviations like ARTNO or QTY_S.
• Date and number format conversions come for free. Claude correctly infers that 2026-06-04 and 04.06.2026 are the same date in different formats, and handles the conversion without explicit instructions.
• Mixed-row exports are the hardest case. Files where the same column means different things depending on the row type – aggregated totals vs. line items in the same export – require preprocessing logic that pure column mapping can't solve.

Try it

If you're spending time every week manually reformatting data between systems, try it for free – and get in touch if you need a custom integration or have a more complex case.