Most AI features in business software fail for an unglamorous reason: nobody identified the decision they were supposed to change. A dashboard predicting something no one acts on is a cost with a demo attached. So the first question is never which model — it is which decision, made by whom, how often, and what it costs to get wrong today.
Forecasting: beat the rule of thumb, consistently
Reorder quantities, staffing levels and cash position are the forecasting problems that recur in operational software. They usually need classical models rather than deep learning, and their value comes from being slightly better than the current heuristic across hundreds of small decisions rather than dramatically better once.
The baseline to beat is not zero. It is the experienced manager's rule of thumb, and you should measure against exactly that. A model that cannot outperform the person who has been doing this for ten years is a finding worth having early, not a project to continue funding.
Document extraction: remove typing, keep judgement
Supplier invoices, delivery notes and receipts arrive as PDFs and photographs, and somebody retypes them. Extraction with per-field confidence scoring, plus a review queue for anything uncertain, removes most of that work while keeping a human on the genuinely ambiguous cases.
This is the easiest category to justify commercially because the outcome is directly measurable: hours removed and error rate, both verifiable by the client against their own before-and-after.
Assistants: the engineering is permissions, not prompting
Retrieval-based assistants answering questions over company records are useful, and the hard part is not the model. An assistant must retrieve only what the asking user is already allowed to see, cite the records it answered from, and refuse when those records do not contain an answer.
Get that wrong and you have built a data-leak with a chat interface. Permission-aware retrieval is a design constraint from the first line, not a filter applied to results afterwards.
What we avoid
Features where a confident wrong answer is expensive and invisible. Models trained on data too thin to support the claim. Fine-tuning as an opening move rather than a last step.
We also insist on an evaluation set before launch — a fixed list of inputs with known-correct outputs — so accuracy is a number both sides can check, and so a regression six months later is detected rather than reported by a customer.
Sequence matters more than sophistication
Clean data models and documented APIs come first; the model is the last component and often the smallest. Businesses whose records are already coherent can add these features incrementally.
Businesses whose data lives in three disconnected systems should fix that first — and that fix is worth doing whether or not any AI is ever added on top of it.
Working on something this touches on? Tell us what you are building — we will say honestly whether we are the right team for it.