Where AI actually pays off in business software

Most businesses do not need a custom model. They need one repetitive task handled reliably. These are the cases we see working.

Pulling data out of documents

Invoices, delivery notes, purchase orders and application forms still arrive as PDFs and photographs, and somebody types them into a system. Modern models handle this well, including messy scans that defeat traditional OCR.

The saving is straightforward: if two people spend half their day on data entry, that is a full role recovered. Keep a review step for anything below a confidence threshold.

First-draft responses to routine enquiries

A large share of inbound support is a small number of questions repeated. Drafting a response from your own documentation, for a human to approve and send, captures most of the time saving while keeping a person accountable for what goes out.

Search that understands intent

Keyword search fails when the user's words differ from your content's words. Semantic search matches meaning, which matters most on large document sets — policies, manuals, historical records — where staff currently ask a colleague instead.

Forecasting from your own history

Stock levels, staffing, seasonal demand. If you hold several years of clean transactional data, a model will usually beat a spreadsheet extrapolation. If your data is patchy, fix that first — it is the cheaper project and everything else depends on it.

Where it still does not pay

• Tasks that must be exactly right with no review. Anything legally or financially binding needs a human signature.
• Problems a rule would solve. If an if statement handles it, use the if statement — it is cheaper, faster and auditable.
• Fully autonomous customer conversations. The failure cases are visible, public and disproportionately damaging.

How to start

Pick one task, measure the hours it consumes today, and benchmark accuracy against your real data before committing. If it does not clearly beat the current process on that one task, it will not survive contact with the rest of the business.