Applied AI for Business: Where Machine Learning Actually Pays Off

Enthusiasm for AI has outrun evidence of return in most companies, and the reason is consistent: the modelling was never the hard part. Projects stall because the data needed to train on does not exist, is not labelled, or is spread across four systems that disagree with each other.

Four problem shapes that reliably work

Forecasting is the first. If you have three years of clean transaction history with seasonality in it, demand and inventory forecasting is well-trodden and delivers measurable savings. Second is anomaly detection: fraud, unusual invoices, equipment behaving out of pattern. You need a lot of normal examples and very few abnormal ones, which is exactly the data most businesses already have.

Third is document extraction. Invoices, purchase orders, forms and reports still get keyed in by hand at enormous cumulative cost, and modern extraction handles this well enough that humans only review exceptions. Fourth is classification and routing: sorting incoming tickets, enquiries or applications to the right queue.

"If you cannot describe the decision a human currently makes and the information they use to make it, there is nothing to automate yet."

The data question, answered honestly

Before scoping any model, check three things. Do you have at least a couple of years of the relevant history? Is the outcome recorded, so the model has something to learn from? And is it consistent, or did the process change halfway through in a way nobody logged? A model trained on inconsistent history will confidently reproduce that inconsistency.

# Baseline first. If a model cannot beat the simple rule the business
# already uses, it is not ready to replace it.
def seasonal_naive(history, season_length=52):
    """Last year, same week. Surprisingly hard to beat."""
    return history[-season_length]

def evaluate(model_predictions, baseline_predictions, actuals):
    model_error = mean_absolute_error(actuals, model_predictions)
    baseline_error = mean_absolute_error(actuals, baseline_predictions)
    return {"model": model_error, "baseline": baseline_error,
            "worth_shipping": model_error < baseline_error * 0.9}

Keep a human in the loop at the start

Deploy the first version as a recommendation rather than an automatic action. Let staff accept or override it, and record which they chose. That gives you a real accuracy measurement on live data and a labelled dataset for the next iteration, and it means an early mistake costs a correction rather than a customer.

How this fits an existing system

The usual architecture is unglamorous and works well: your existing application carries on running, and a small Python service exposes the model behind an internal API. Nothing about your ERP or CRM has to be rewritten, and the model can be retrained and redeployed without touching the product.