Summarisation

Summarisation is the AI task of condensing longer text into a shorter version that captures the essential points. It is one of the most immediately useful AI capabilities — saving hours of reading time across emails, reports, articles, meeting transcripts, and documents.

Types of summarisation

• Extractive summarisation: Selects and combines the most important sentences directly from the original text. The summary contains only original sentences, rearranged and filtered.
• Abstractive summarisation: Generates new sentences that capture the meaning of the original text. The summary may use different words and phrasing than the source. This is what modern LLMs excel at.
• Query-focused summarisation: Summarises text with a specific question or angle in mind. "Summarise this contract focusing on termination clauses."

How LLMs summarise

Modern LLMs handle summarisation naturally because they understand language deeply. You provide the text and an instruction ("Summarise this in 3 bullet points"), and the model generates a concise version. The quality depends on:

• The length and complexity of the source text.
• The specificity of your summarisation instructions.
• Whether the text fits within the model's context window.

Effective summarisation prompting

• Specify length: "Summarise in 100 words" or "Summarise in 5 bullet points."
• Define audience: "Summarise for a non-technical executive" produces different results than "summarise for the engineering team."
• Specify focus: "Summarise focusing on financial implications" filters out irrelevant details.
• Request structure: "Provide a one-sentence summary, then 3-5 key points, then any action items."

Business applications

• Email triage: Summarise long email threads so you can prioritise without reading everything.
• Meeting notes: Summarise transcripts into key decisions, action items, and discussion points.
• Report digests: Condense lengthy reports into executive summaries.
• Research synthesis: Combine and summarise multiple articles on a topic.
• Legal review: Summarise contracts highlighting key terms and obligations.
• Customer feedback: Aggregate and summarise hundreds of reviews into themes and insights.

Challenges and limitations

• Hallucination risk: The model may add information not present in the source text, especially with abstractive summarisation.
• Context window limits: Very long documents may need to be summarised in chunks, then the chunks combined.
• Important detail loss: A summary inherently omits information, and the model may not always choose the best details to keep.
• Bias: The model's training may cause it to emphasise certain types of information over others.