Definition
Generative AI is artificial intelligence that produces new content rather than only classifying or predicting. Give it a prompt — a question, an instruction, a sketch, a few examples — and it generates text, images, audio, video or software code that did not exist before.
Traditional "predictive" AI answers narrow questions: is this email spam, will this customer churn, what is in this photo. Generative AI answers open-ended requests: write this email, summarise this report, draw this product concept, explain this spreadsheet formula. That shift from labelling to creating is why generative AI spread so quickly into everyday office work.
The best-known generative AI tools are chat assistants built on large language models, image generators built on diffusion models, and coding assistants that suggest and write software.
How generative AI works
Learning patterns from data
Generative models are trained on very large collections of examples. A language model reads huge amounts of text and learns, statistically, which tokens tend to follow which. Most modern language models use the transformer architecture introduced in the 2017 paper Attention Is All You Need.
Generating one piece at a time
When you send a prompt, a language model predicts the most plausible next token, adds it, and repeats until the answer is complete. Image models work differently — diffusion models start from random noise and gradually refine it into an image matching the prompt — but the principle is the same: generate output that fits the patterns learned in training.
Shaped by instructions and context
The same model can draft a legal summary or a birthday poem because its behaviour depends on the prompt and the information supplied in its context window. This is why prompt engineering and techniques such as retrieval-augmented generation matter so much in practice.
Why generative AI matters at work
The research picture is clear on two points: generative AI can deliver large gains on the right tasks, and it can make people worse on the wrong ones.
- Adoption. The 2024 Microsoft and LinkedIn Work Trend Index found that 75% of knowledge workers use generative AI at work, and 78% of those users bring their own tools.
- Productivity on suitable tasks. In the Harvard Business School and BCG field experiment with 758 consultants, those using GPT-4 completed 12.2% more tasks and worked 25.1% faster on tasks within the AI's capability.
- Errors on unsuitable tasks. In the same study, on a task outside the AI's capability frontier, AI users were 19 percentage points less likely to reach the correct answer.
- Gains for less experienced staff. The Stanford and MIT study of 5,179 customer-support agents measured a 14% average productivity increase, rising to 34% for novice and low-skilled workers.
Types of generative AI
| Type | What it creates | Typical work use |
|---|---|---|
| Text (LLMs) | Emails, reports, summaries, answers | Drafting, summarising, research starting points |
| Code | Functions, scripts, formulas | Spreadsheet formulas, automation, software |
| Image | Illustrations, mock-ups, photos | Marketing visuals, concepts, presentations |
| Audio | Speech, voice, music | Voice-overs, transcription and dubbing |
| Video | Short clips, animations | Training videos, social content |
| Multimodal | Mixed inputs and outputs | Reading charts, describing images, analysing documents |
Strengths and limits
Strengths: speed, fluency, breadth of general knowledge, tireless iteration and the ability to restructure information into almost any format.
Limits:
- Hallucination. Models can produce confident, fluent and false statements — see AI hallucination. In Mata v. Avianca, a US federal court fined lawyers $5,000 in June 2023 for filing a brief citing cases ChatGPT had invented.
- Accountability stays with the organisation. In Moffatt v. Air Canada (February 2024), a tribunal held the airline liable for wrong refund advice given by its website chatbot.
- Knowledge cut-off. A model only knows what was in its training data unless connected to current sources.
- Bias. Models can reproduce stereotypes present in their training data.
- Data exposure. Pasting confidential information into unapproved tools can breach company policy or law.
Example
Marta runs customer communications for a mid-sized logistics company. Every week, her team answers dozens of near-identical emails about delivery delays, and every quarter she writes a customer newsletter.
She starts with low-risk work. Using her company's approved assistant, she asks it to draft five reply templates for common delay scenarios, giving it context (B2B customers, a formal but warm tone), the company's actual delay-compensation policy pasted into the prompt, and a requirement to keep each reply under 120 words. The drafts are good, but one template promises a refund the policy does not allow. Marta catches it because she checks every commitment against the policy, then adds "never promise anything not stated in the policy" to the prompt.
For the newsletter, she uses an image generator to create a header illustration and the assistant to turn bullet points from operations into readable paragraphs. She edits the final text herself and has a colleague review it before sending.
After a month, her team reports — by their own timing — that drafting a delay reply takes a few minutes instead of a quarter of an hour, and that the newsletter takes one afternoon instead of two days. The policy check stays mandatory.
How to start using generative AI: step by step
- Read your organisation's AI policy. Find out which tools are approved and what data they may receive.
- Pick a low-risk, repeated task. Drafting, summarising, rewriting and brainstorming are good first choices.
- Prepare your input. Remove personal or confidential data you are not permitted to share.
- Write a structured prompt. Give context, a role, a clear ask, a format and a tone — the CRAFT framework is one way to remember this.
- Supply the facts. Paste the policy, data or document you want the answer based on, instead of relying on the model's memory.
- Add an example when format matters. One or two samples of the output you want improve consistency — see few-shot prompting.
- Iterate. Ask for specific changes: shorter, more formal, as a table, for a different audience.
- Verify everything that matters. Names, numbers, dates, quotes, citations and commitments.
- Keep a human accountable. Decisions about people, money and legal matters need a human in the loop.
- Save and measure. Store prompts that work and note the time saved and errors caught.
Common mistakes
- Treating generative AI like a search engine and trusting its facts without checking.
- Writing vague one-line prompts and blaming the tool for generic answers.
- Using personal accounts for company data.
- Applying AI to tasks it handles poorly — complex reasoning outside its competence, precise calculations without checks, or decisions about people.
- Publishing AI output without editing or disclosure where disclosure is required.
Key takeaways
- Generative AI creates new content from patterns learned in training data.
- It delivers measured productivity gains on suitable tasks and measurable errors on unsuitable ones.
- Hallucination, bias and data exposure are the main risks.
- Good results come from structured prompts, supplied facts, verification and human accountability.
- Start small, measure results and build a library of prompts that work.