AI Basics for Business Owners — Lesson 2
Machine Learning, Models, and Training
Learning Objectives
- 1Explain training, inference, and model concepts in plain language.
- 2Understand why training data quality determines AI output quality.
- 3Evaluate AI vendor claims about model capabilities.
How AI systems learn
Machine learning is the process of teaching an AI system to recognize patterns by showing it examples. Training is the phase where the system processes large amounts of data and adjusts its internal parameters to get better at a task. The result of training is a model — the distilled knowledge that can be applied to new situations.
Think of it as apprenticeship. If you wanted to train someone to identify fraudulent transactions, you would show them thousands of examples of legitimate and fraudulent transactions, point out the patterns, and let them practice. Over time, they develop judgment about what looks suspicious. Machine learning does this mathematically, at scale, with far more examples than any human could review.
After training, the model enters production for inference — applying the patterns it learned to new data. When a spam filter evaluates your incoming email, it is running inference: applying its trained model to each new email to predict whether it is spam or legitimate.
Training data determines everything
The quality, size, and representativeness of training data determine the quality of the resulting model. A model trained on biased examples will produce biased results. A model trained on limited data will miss important patterns. A model trained on outdated data will make predictions based on patterns that may no longer apply.
For business applications, this means asking vendors: What data was the model trained on? How recent is it? Does it represent our customer base, industry, and use case? How is the training data updated? What biases might be present? These questions reveal whether the AI is likely to perform well for your specific situation.
Custom training — fine-tuning a model with your own data — can improve performance for specific tasks but requires sufficient quality data, technical expertise, and ongoing maintenance. It is not always necessary. Many business uses work well with general-purpose models guided by good prompts.
Model types and choosing between them
Large Language Models (LLMs) like GPT-4, Claude, and Gemini are trained on vast amounts of text and can generate, analyze, and transform text for a wide range of tasks. They are general-purpose and versatile but may lack domain-specific knowledge.
Specialized models are trained for specific tasks: image recognition, speech transcription, document classification, fraud detection, or medical imaging. They typically outperform general-purpose models on their specific task but cannot do anything else.
When evaluating AI for a business use case, consider: Is a general-purpose model sufficient, or does this task require specialized training? What accuracy is needed? What is the cost per use? What data needs to be shared with the model? Is the model hosted by a third party or can it run in our environment? These factors determine the right model choice.
Case Study
The biased hiring model
Situation
A company trained an AI model to screen resumes using historical hiring data from the past ten years. The model consistently ranked male candidates higher than female candidates with identical qualifications. Investigation revealed that the historical data reflected a decade of unconscious bias in human hiring decisions.
Analysis
The model did exactly what it was trained to do — it learned the patterns in the data. Since the data reflected biased human decisions, the model reproduced and amplified those biases. The training data was accurate but not fair.
Takeaway
AI trained on biased historical data reproduces bias at scale. Any AI used for decisions that affect people must be audited for bias and tested against fairness criteria.
Reflection Questions
- 1. If you were evaluating an AI vendor, what questions would you ask about their training data?
- 2. Are there decisions at your organization that are currently made by rules or intuition that might benefit from pattern-based AI?
Key Takeaways
- ✓Training teaches AI to recognize patterns; the result is a model used for inference.
- ✓Training data quality determines model quality — biased data produces biased results.
- ✓General-purpose models are versatile; specialized models excel at specific tasks.
- ✓Ask vendors about training data, recency, bias testing, and relevance to your use case.