Generative AI is artificial intelligence that creates new content, such as text, images, code, audio or video, by learning patterns from large amounts of existing data and producing output that fits a prompt. It is what sits behind AI chat assistants, writing and coding tools, image generators and the “AI” features now added to many business applications. Its outputs are probabilistic: often useful, sometimes wrong, and rarely identical twice. For buyers, it is less a single product than a capability that shows up in many products under different terms.
At a glance
- Generative AI creates content; much earlier business AI classified, scored or predicted instead.
- Large language models (LLMs) are the best-known type, behind text-based tools such as chat assistants.
- Outputs are probabilistic: the same input can give different results, and confident output can be wrong.
- It works best on well-defined tasks with clear quality criteria and human review in proportion to the stakes.
- What you buy is usually an application or feature on top of a model, and data handling and pricing terms vary by vendor and tier.
What problem it solves
For most of computing history software could only do what it was explicitly programmed to do. That made it good at structured, repeatable work and poor at anything involving interpretation: drafting a reply, summarizing a long report, answering a question in plain language or writing code from a description.
Generative AI handles those tasks. Teams use it to draft and edit documents and emails, summarize meetings and research, answer questions from internal knowledge, suggest responses for support agents, and generate, explain or review code. Where it works well it cuts time spent on routine reading and writing, and makes information easier to find. Where it is applied without a defined task, good data or review, it produces plausible output that someone still has to check, which can cost more time than it saves.
How it works
Training. A generative model is trained with machine learning on very large amounts of example data, such as text, code or images, and learns the statistical patterns in it. Language models learn how words and ideas tend to follow one another; image models learn how visual features relate to descriptions.
Generation. Given a prompt, the model produces output that fits those patterns, one piece at a time. It does not look up verified facts unless the application supplies them, which is why it can produce confident errors, often called hallucinations.
Grounding. Many business applications connect the model to your own content using retrieval-augmented generation (RAG): the system finds relevant documents and passes them to the model with the question. This tends to improve relevance and accuracy for company-specific questions, but it does not guarantee correct answers.
From content to action. Conversational AI often uses generative models for dialogue, and agentic AI uses them to plan and carry out multi-step tasks across systems, which raises the stakes for permissions and oversight.
Model, application or feature. You might buy access to a model directly, an application built on one, or a feature inside software you already own. The model sets the capability ceiling; the application decides the data connections, controls and contract terms. The LLM entry covers this distinction in more detail.
When it matters for buyers
- Vendor pitches and renewals. Many software vendors now add generative AI features, sometimes at extra cost; ask what the feature does and on what terms.
- Board or leadership pressure. Start from specific problems and measurable outcomes, not a mandate to “use AI.”
- Unapproved use. Staff using personal AI accounts with company data is a common reason to buy an approved tool and set an AI governance policy; see shadow AI.
- Pilots. Test with your own documents and workflows in a proof of concept; demos on curated examples often overstate results.
- Regulated work and liability. Legal, financial, health and customer-facing outputs need review steps, records, and contract terms on who is responsible for errors.
- Fraud and security. The same technology makes convincing phishing and deepfake fraud easier, which affects security training and payment controls.
Questions to ask vendors
- What exactly does the AI feature do, and which tasks has it been tested on?
- Which model or models does it use, who hosts them, and what happens when a model is updated or retired?
- Are our prompts, documents and outputs stored, for how long, and are they used to train any model?
- Where is our data processed, and can we restrict it by region?
- How does it respect our existing permissions when it reads internal content?
- How do you measure accuracy for our use case, and can we run a proof of concept on our own data?
- How is it priced at our expected volume, and what limits or overage charges apply?
- Who is responsible contractually if the output is wrong or infringes someone’s rights?
We help buyers scope use cases and compare AI platforms and terms in our artificial intelligence overview.
How it differs from predictive AI
Predictive AI, the long-established business use of machine learning, analyzes data to classify, score or forecast: whether a transaction looks fraudulent, which customers may leave, how much stock to order. Its output is a label, a score or a number, and it can be tested against known outcomes. Generative AI produces new content instead, which makes it suited to language and creative tasks but harder to measure, since there is often no single right answer. Both are built with machine learning and fall under artificial intelligence (AI). Many products combine them, for example a support platform that predicts a ticket’s category and then drafts a reply.
Problems this helps with
- Every vendor pitch now says "AI-powered" and we can't tell what that means
- Leadership wants an AI plan before we have a use case
- Staff using consumer AI tools with company data
- A pilot that looked great in the demo but struggles on our own documents
