What is generative AI?
Generative AI is a class of AI models that create new content, such as text, images, software code, audio, or synthetic data, by learning the patterns in large collections of existing examples. A large language model (LLM) that drafts an email, an image model that renders a product mockup, and a speech model that reads a report aloud are all generative AI. The output is new material that follows the patterns the model learned, and it is produced on demand from a prompt.
This sets generative AI apart from predictive AI, which produces a label, a score, or a forecast from existing data. A churn model says which customers are likely to leave. A generative model writes the retention offer. Many business systems combine the two.
How generative AI works
Most generative models are trained with self-supervised learning on very large datasets. A language model learns to predict the next token (a word or part of a word) given everything that came before it. Repeated across trillions of tokens, that single objective teaches grammar, facts, writing styles, and the structure of code. Image models learn a different objective, often by removing noise from a scrambled image until a clear one emerges, guided by a text description.
After pre-training, a model is tuned to follow instructions and to refuse harmful requests, usually through instruction tuning and reinforcement learning from human feedback (RLHF). The result is a foundation model that a business can use as is, adapt with retrieval-augmented generation (RAG), or fine-tune on its own data.
Generation is probabilistic. The model samples each token from a distribution, so the same prompt can produce different answers, and the model can state things that are not true. That property, called hallucination, is the central operational risk of generative AI and shapes how it should be deployed.
Generative AI compared with predictive AI
| Aspect | Generative AI | Predictive AI |
|---|---|---|
| Output | New text, images, code, or audio | A label, score, or forecast |
| Typical input | A natural-language prompt plus context documents | Structured features from a database |
| Training data | Broad public corpora, with company documents supplied at retrieval time | The company’s own historical records |
| Main failure mode | Plausible but wrong content | Drift as conditions change |
| Typical business use | Drafting, summarizing, answering questions, writing code | Forecasting demand, scoring risk, detecting fraud |
Generative AI in practice
A regional property insurer receives thousands of claim emails a month with attached photos and repair estimates. A generative AI system reads each message, extracts the policy number, loss date, and claimed items into a structured record, and drafts an acknowledgment letter in the adjuster’s voice. The adjuster reviews the draft, corrects anything wrong, and sends it. The adjuster’s time per claim shrinks to a few minutes of review, and the extracted fields feed the existing claims system without retyping.
The same pattern appears in a freight broker that drafts carrier rate requests, a law firm that summarizes deposition transcripts, and a city clerk’s office that answers residents’ questions about permit requirements from the published code.
Generative AI and private AI
Generative models need the text they work on, which means the claim email, the contract, or the patient note travels to wherever the model runs. Companies with sensitive data or steady volume often choose to run open-weight models on infrastructure they control, so documents never leave their environment and usage is not metered per token. That is the private AI approach. The tradeoff is that the company must operate GPUs and an inference server, which is a sound choice at steady volume and a poor one for occasional use.
What to watch for
- Decide where a human reviews output before it reaches a customer, a regulator, or a system of record, and keep that review step until measured error rates justify removing it.
- Ground the model in your own documents with RAG rather than relying on what it memorized during training, which has a knowledge cutoff and contains nothing about your business.
- Ask vendors where prompts and documents are processed, how long they are retained, and whether they are used for training.
- Measure quality on a fixed set of real examples before and after every model or prompt change, because output quality shifts in ways that casual testing misses.
- Budget for token costs or GPU capacity based on measured volume, since generation cost scales with the length of both input and output.