Labeling Data for Machine Learning in SMEs
Where do labels for supervised learning come from? Inconsistent labels, small data and steps to set up labeling at SME scale; Floemdata's real status.

You want to teach a model “which customer pays an invoice late?” Past invoices alone are not enough; next to each invoice the model also needs the answer “paid late / paid on time”. That answer is called a label. In most SMEs the hard part is not choosing a model but finding a reliable label. The example below is fictional.
Where labels come from
- An outcome the system already records: an invoice paid more than 30 days after its due date, a returned order, a closed quote. This is the cheapest and most reliable source because nobody writes the label by hand.
- Manual marking: an expert, for example a collections officer, classifies each record. It is slow but the only way for outcomes that are not in the system.
- Rule-based pre-label: a simple rule assigns labels automatically and a person only corrects the doubtful ones.
Inconsistent labels
Two people can label the same record differently: is “late payment” 30 days or 15? If the label definition is not written down, the model learns the inconsistency between the people. Run a small test: have two people label the same 50 records independently and see how many they agree on. If agreement is low, the problem is not the amount of data but the label definition.
Steps for small data
- Define the label in one sentence and give examples of borderline cases.
- Label a few hundred records first and compare against a simple rule (“overdue and high amount”); if the model does not beat it, do not continue.
- Do not feed in information from after the label is known: fields such as the payment date train the model on a prediction it could never make in practice.
- Look at the label distribution: if a class is a 2 percent minority, accuracy alone is misleading.
Where Floemdata stands
The alerts in Floemdata's product definition are rules computed from history: the rule is defined as data, it stays silent when history is insufficient, and no invented thresholds are set. We could not find a machine-learning model trained on labeled data in the product documents, and the AI assistant is on the roadmap. The only source Floemdata's warehouse runs on real data is SAP Business One; which SAP field is suitable as a label depends on the customer environment and is not verified here.
To start, write the outcome you want to predict in one sentence, have two people label the same 50 records and note the agreement rate. Questions are welcome at info@floemdata.com or via Floemdata.