Most AI projects begin with a technology and go looking for a problem. It works better the other way round. Four questions, asked in order, settle most cases, and the cheapest suitable answer comes first.
1. Can you write it down as a rule?
If so, you don’t need AI. Approval limits, tax calculations, reorder points and routing by region belong in your ERP’s configuration. A rule is cheap to run, simple to test and behaves the same way every time. If your rules have more exceptions than cases, read on.
2. Are you predicting a number or a category from structured data?
Use classical machine learning. Demand forecasts, churn, failure risk, late payment and fraud flags all fit, particularly when you have years of ERP history. Models such as gradient boosting are usually cheaper and more consistent for this work than a large language model, and you can measure them directly against your current method.
A useful warning sign: if someone proposes pasting a sales table into a chatbot to forecast next quarter, ask for a backtest against last year’s method first.
3. Is the input messy text, or is the output language?
Use generative AI. Reading supplier bills, searching policy documents, drafting quotes and summarising tickets all fit, because the work involves language no rule can parse reliably. Retrieval lets the model cite your own documents, and an evaluation set tells you how often it gets things right. Keep calculations out of it: prices, totals and stock figures should come from your system, not from the model.
4. Does the task need several steps that change each time?
Consider an agent. Checking stock, comparing vendors, drafting an order and asking for approval is a fair example, because the path depends on what each step finds. If the path is the same every time, you want a workflow that calls a model at one step, and that will be cheaper and easier to test.
They work best in layers
These aren’t rivals. A good system often uses all of them on one process. Take reordering stock:
- The ERP holds the stock and vendor records and enforces approval rules.
- A machine learning model forecasts demand per product.
- Generative AI explains the forecast in plain language and reads vendor emails.
- An agent ties the steps together and prepares a draft order for the buyer.
The order matters. Each layer depends on the one beneath it, which is why we build from the records up. You can see the same idea for five business modules in the drawing on our home page, or browse twelve use cases by problem.