Large language models (LLMs) — the technology behind ChatGPT, Claude, Gemini and open models like Llama and Qwen — are now good enough to use in real products. But "it can write" is not a business case. Here is where we see them earning their keep.
Answering questions from your own documents
The most valuable pattern is called retrieval-augmented generation (RAG). Instead of asking the model what it remembers from the internet, you let it search your documents first — manuals, policies, product sheets, past tickets — and answer only from what it finds, with references.
This is how companies build internal assistants that new staff actually trust: "How do I process a return for a wholesale customer?" gets the answer from your own procedure, not a guess.
Summarising long, boring things
Meeting transcripts, customer feedback, legal documents, research papers. LLMs are excellent at producing a first summary that a person then checks. The time saving is real because reviewing is much faster than writing.
Pulling structured data out of messy text
Emails, chat messages and free-text forms hide useful information: order numbers, complaints, dates, amounts. An LLM can turn them into clean rows in a database, which can then feed dashboards and alerts.
Drafting, not publishing
LLMs are great at first drafts: product descriptions, reply templates, job adverts, report sections. The rule we recommend to clients is simple — AI drafts, a human approves, at least until you have measured the quality over time.
Where LLMs still struggle
- Exact facts and numbers. Without a source to check against, models can still state wrong things confidently.
- Calculations. Let the model call a calculator, database or spreadsheet instead of doing maths in its head.
- Private data. Decide early what may leave your systems. For sensitive work, an open model hosted on your own server may be the better choice.
Looking ahead to 2027–2028
LLMs are moving from chat to action. Models increasingly use tools — search, databases, email, your CRM — to finish a job end to end. Costs are also falling fast, which makes it realistic to run AI on every support ticket or every document rather than only the important ones.
Our advice
Start with a narrow, well-defined task, connect the model to trusted data, and measure accuracy before you scale. Contact us if you want an assistant that answers from your own documents.