Last reviewed:
What is prompt engineering? Definition and business implications
Prompt engineering is the discipline of writing prompts methodically. It is not a rare profession but a transverse competence, which consists of formalising, testing, and standardising the instructions sent to an LLM to obtain reliable and reproducible outputs at the scale of an organisation.
Prompt engineering covers the set of techniques that turn a vague request into a structured instruction producing a reliable output. The main techniques include: assigning a role (“You are an expert in...”), providing examples (one-shot, few-shot), decomposing the task into steps (chain of thought), specifying the output format (JSON, markdown, numbered plan), and introducing negative constraints (“do not mention...”). In 2023, the term designated an emerging and well-paid profession (up to 300,000 dollars per year at some U.S. laboratories). Since 2024-2025, this segmentation has largely collapsed: models have become more tolerant of imprecise prompts, and best practices have spread. Prompt engineering is no longer a profession; it is a skill integrated into business roles (sales, legal, marketing), in the same way as Excel mastery or Google search.
Concrete example
The NBER study by Brynjolfsson, Li, and Raymond (2023, Generative AI at Work) observed 5,172 customer-support agents equipped with an AI assistant. The average productivity gain was 14%, but varied strongly by experience level and prompting quality: 34% gain for the least trained agents, and a near-zero effect for the most experienced on simple cases. The central lesson for executives: AI-driven productivity is not a property of the model, but of usage practice. Without a collective prompting discipline, gains remain uneven and not capitalisable.
A larger-scale case, in a customer-support context: Klarna, February 2024. The Swedish fintech group publicly announced that its AI assistant (based on OpenAI) handles two-thirds of customer-service conversation volume in one month — 2.3 million interactions in 23 languages, the equivalent of 700 full-time agents. Tickets are resolved in 2 minutes on average vs. 11 minutes previously, with a satisfaction score matching human agents. The key point: Klarna reported the performance does not come from the raw model, but from "fourteen months of prompt engineering and orchestration" — a versioned prompt library, systematic A/B tests, routing to humans on high-stakes cases. Prompt engineering didn't replace the agents; it changed the nature of the work: fewer repetitive tasks, more supervision and complex cases.
Comparison
| Technique | What it changes | Typical cost | When to choose it |
|---|---|---|---|
| Prompt engineering | Instructions and examples given to the model at each call — the model itself is unchanged | Low — a few hours to a few weeks per use case | The default: 80–90 % of enterprise use cases are solved this way. Fast to iterate, reversible, no retraining. |
| RAG (Retrieval-Augmented Generation) | The model receives relevant excerpts from a document base with each question | Medium — vector search infrastructure + source curation | When domain knowledge lives outside the model (internal documents, product catalogue, an evolving knowledge base). |
| Fine-tuning | The model is retrained on a proprietary corpus — behaviour and style change durably | High — annotated dataset + compute cost + maintenance on every new base model | Highly specific editorial style, niche vocabulary the model doesn't cover, or latency drop on a single well-scoped task. |
FAQ
Do you need an expert to do prompt engineering?
No. In business, prompt engineering is primarily a writing discipline: clear context, explicit output format, short examples. Documented gains (14 % average productivity uplift, NBER 2023 study of 5,172 support agents) come from good practices that can be shared, not from technical expertise.
How do you structure a good prompt?
Four building blocks: role (who the model is speaking to), context (data, constraints, expected tone), task (action verb and success criterion), output format (explicit structure). Adding one or two examples (few-shot) improves format fidelity. Iterate on 3-5 versions before fixing a production prompt.
Prompt engineering or fine-tuning: what's the difference?
Prompt engineering adjusts what you ask the model at inference: fast, no model training cost, reversible. Fine-tuning retrains the model on proprietary data: expensive, slower, but useful for a very specific style or expertise. In 2026, 90 % of enterprise use cases are solved by prompt engineering + RAG.
What does a bad prompt cost?
On a high-volume generation use case (support, extraction, summarisation), a poorly calibrated prompt can multiply per-request cost by 3 to 10 (inflated input tokens, off-format responses needing rework) and drop quality by 20 to 40 points on internal benchmarks. Auditing a production prompt takes a few hours; savings are measurable in weeks.
Is prompt engineering becoming obsolete?
No — it's evolving. Recent models (Claude 4/5, GPT-5) tolerate vague prompts better, but the performance gap between a mediocre and a well-crafted prompt remains significant in production. Techniques evolve (chain-of-thought, tool use, structured output); the discipline stays.
Should companies document their prompts?
Yes — in a versioned internal library. Each production prompt should have: an identifier, a date, a use case, the target model version, the measured quality metric, and an example of expected output. Without this discipline, prompts duplicate, drift and become un-auditable — treat them as code.
See also
Further reading
Generative AI at Work, Brynjolfsson, Li & Raymond, NBER Working Paper 31161, 2023
Sources
- Generative AI at Work, Brynjolfsson, Li & Raymond, NBER Working Paper 31161, 2023. https://www.nber.org/papers/w31161
- Prompt engineering overview, Anthropic documentation 2026. https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview