AdAI
Definition

Classification is the machine learning task of sorting things into categories. A classifier looks at an input (an email, a customer record, a transaction) and assigns it to one of a fixed set of labels: spam or not spam, churn risk or stable, fraudulent or legitimate, high priority or low. The model learns from labelled examples and then applies what it learned to new inputs it has not seen before.

Key Takeaways

  • Classification sorts inputs into discrete categories. It is what powers spam filters, lead scoring, ticket routing, fraud detection, and most other 'smart sorting' features in business software.
  • It is one of the two main types of supervised machine learning, alongside regression (which predicts numbers rather than categories).
  • Almost every business tool you already use runs classifiers in the background. Gmail, HubSpot, Stripe, Zendesk, Salesforce, Intercom, all of them.
  • You do not need a data scientist to use classification. The value already lives inside the tools. Custom classifiers (training one on your own data) are now achievable with a few hundred examples and a few hours.
  • Accuracy depends on training data, not on the algorithm. Gmail hits 99.9% on spam because it has billions of examples. Lead scoring hits 75% because the signal is weaker and the data is messier.
99.9%
of spam, phishing, and malware blocked by Gmail using classification models, across roughly 15 billion unwanted messages filtered every day
Source: Google Workspace Security Report, 2024
+38%
improvement in Gmail spam detection (with false positives down 19.4%) after Google deployed its RETVec text classifier in 2023
Source: Google Research, RETVec deployment paper, 2023
82.6%
of phishing emails analysed between September 2024 and February 2025 contained AI-generated content, putting classification models in an arms race with generative AI
Source: Hoxhunt Phishing Trends Report, 2025

In Simple Terms

A classifier is a sorter. You hand it a thing. It tells you which bucket the thing belongs in. The buckets are fixed in advance. The decision is the model's best guess based on what it has seen before.

Spam filtering is the textbook example. Gmail looks at every incoming email and decides: inbox, promotions, spam, or trash. That decision is a classification. Behind it is a model that has been trained on billions of emails, each labelled by humans or by user behaviour (clicked, ignored, marked as spam). The model learns the patterns and applies them to new mail in milliseconds.

Inside an SMB, classifiers are doing the same kind of work in dozens of places. Your support tool tags incoming tickets as billing, technical, or sales. Your CRM scores incoming leads as hot, warm, or cold. Your payment processor flags transactions as legitimate or possibly fraudulent. None of it looks like AI from the outside. All of it is.

Where Classification Shows Up in an SMB

A handful of concrete places.

Email and inbox tools. Gmail and Outlook classify every message as inbox, promotions, social, spam, or junk. Google's published numbers put Gmail's classification accuracy on spam at 99.9%, across about 15 billion unwanted messages per day.

Support tools. Zendesk, Intercom, and Help Scout classify incoming tickets by topic and urgency, then route them to the right person. Most have this on by default. The classification gets better the more tickets your team has labelled.

CRMs. HubSpot, Salesforce, and Pipedrive run lead-scoring classifiers that tag every new contact as a hot, warm, or cold lead based on the patterns of past customers. The model is trained on which leads in your history converted and which did not.

Payment processors. Stripe Radar runs a classifier on every transaction and assigns a fraud risk score. The model is trained on Stripe's view of billions of transactions across millions of businesses, which is why it sees patterns no individual SMB could spot.

Review and feedback tools. Customer feedback platforms classify open-ended responses as positive, negative, or neutral, and by topic (price, support, product quality), so you can see trends without reading every comment.

When to Build Your Own Classifier

Most SMBs never need to. The classifiers built into your existing tools cover the most common needs and improve with time.

A custom classifier earns its place when your business has a sorting problem that no existing tool covers. A few patterns where this happens.

Product categorisation. An ecommerce business with 50,000 SKUs that need to be tagged by category, material, season, and audience. A custom classifier trained on a few hundred labelled examples can tag the rest at a fraction of manual cost.

Domain-specific document sorting. A small law firm sorting incoming case documents as contracts, correspondence, evidence, or court filings. Off-the-shelf classifiers do not know your domain. A fine-tuned one does.

Custom lead scoring. When the built-in scoring in your CRM does not capture what actually predicts a customer for your specific business. A custom classifier trained on your actual closed-won versus closed-lost deals can outperform the generic scoring.

Training cost in 2026 is low. OpenAI fine-tuning, Google Vertex AI AutoML, and Hugging Face AutoTrain all let an SMB train a custom classifier with a few hundred labelled examples and tens of dollars in compute. The expensive part is now labelling your data, not running the model.

“With TensorFlow, we are now blocking around 100 million additional spam messages every day.”

Neil Kumaran, Group Product Manager, Gmail Security & TrustGoogle Workspace Blog, 2019

Frequently Asked Questions

What is the difference between classification and regression?
Classification predicts categories. Spam or not spam. Hot lead or cold lead. Regression predicts numbers. House price, predicted revenue, days until churn. Same machine learning family, different outputs. Most SMB use cases (filtering, tagging, routing) are classification problems.
What can classification actually do for my business?
Filter spam out of your inbox. Tag incoming support tickets by topic or priority. Score leads as hot, warm, or cold. Flag customers at risk of churning. Sort product reviews as positive, negative, or neutral. Detect fraudulent transactions. Most of this is already built into the tools you use (Gmail, HubSpot, Zendesk, Stripe Radar) and runs without you knowing it is there.
Do I need a data scientist to use classification?
No. The classification running inside Gmail, HubSpot, Zendesk, Intercom, and Stripe is configured by the vendor. You get the value without touching a model. If you want a custom classifier (sort your own product catalogue, score your own leads with your own rules), services like OpenAI's API, Google Vertex AI, or Hugging Face AutoTrain let you train one in a few hours without a data team.
How accurate is business classification in practice?
Highly accurate for well-defined tasks with lots of training data. Gmail blocks 99.9% of spam. Stripe Radar's fraud classifier prevents billions in losses. Lead scoring is more variable, typically 70-85% accurate because the signal is weaker and tastes vary by company. Accuracy depends almost entirely on how clean and representative your training data is.
Can I train a classifier on my own data?
Yes, and it is much easier than it was three years ago. OpenAI lets you fine-tune a model for classification with a few hundred labelled examples. Google Vertex AI AutoML handles tabular and text classification through a visual interface. Hugging Face AutoTrain runs the same way. Cost is typically tens of dollars in compute for a small SMB use case, not the thousands of dollars custom development used to cost.

Related Glossary Terms & Resources

Join 5,000+ SMB owners getting weekly AI agent insights

Subscribe Free