Article

Prompt engineering for business: 5 techniques that work immediately

Better output from ChatGPT, Copilot or Gemini starts with a better instruction. Five concrete techniques that business owners and teams can apply immediately, plus what you should never put in a prompt under GDPR.

Short answer
  • Prompt engineering is the practice of writing a precise instruction so that an AI model delivers usable output without multiple correction rounds. With five techniques you structurally get better results from ChatGPT Plus ($20/month), Microsoft 365 Copilot ($30/user/month) or Gemini in Google Workspace (Business Standard approx. €11.50/user/month).
  • The five techniques are: assigning a role, specifying the output format, requesting step-by-step reasoning, providing examples (few-shot), and chaining prompts. Each solves a different problem: vague output, wrong formatting, faulty reasoning, lack of context, or tasks too large for a single prompt.
  • GDPR rule: do not put personal data in a prompt unless you have a Data Processing Agreement (DPA) with the provider and a lawful legal basis. This applies to names, email addresses, national ID numbers and medical data. EU AI Act Article 4 (in force 2 February 2025) requires you as an AI tool user to understand what your tool can and cannot do.

What is prompt engineering and why does it matter?

A prompt is the instruction you give to an AI model. Prompt engineering is the deliberate structuring of that instruction so that the model delivers what you need in the first or second attempt. The difference between a vague and a good prompt is not just time savings: a poorly formulated instruction leads to generic output you still need to correct manually, whereas a good prompt delivers direct value. The five techniques below are model-agnostic: they work on ChatGPT, Copilot, Gemini and comparable models.

Five techniques

The five techniques explained

Each technique solves a specific problem. Combine them for complex tasks.

Technique 1: assigning a role

Tell the model what role it takes on for this task. That gives the output a consistent tone, vocabulary and perspective. Without a role, a model responds as a generic assistant. With a role, it responds as the expert you need. Example: start your prompt with 'You are a senior B2B copywriter for a Dutch software company' before asking for a sales text. The model will then automatically adjust its register: more business-like, more concrete, focusing on customer benefits rather than product features. Applicable to ChatGPT Plus, Copilot and Gemini.

Technique 2: specifying the output format

Explicitly state in which format you want the output. Without a format instruction, the model chooses a format based on its interpretation of your question, not on your intended use. Examples: 'Give me a list of five bullet points with no introduction', 'Write this as a table with three columns: name, benefit, example', or 'Output the result as a JSON object with the keys title, description and price'. Format instructions are especially useful when you want to paste the output directly into a spreadsheet, CRM or other system.

Technique 3: requesting step-by-step reasoning

Ask the model to reason step by step before giving a conclusion. This is called chain-of-thought prompting and significantly increases accuracy for logical, mathematical or multi-step reasoning. Add to your prompt: 'Think step by step' or 'Explain your reasoning before giving an answer'. The model then walks through the intermediate steps visibly, so you can check where the reasoning is correct or goes astray. Particularly useful for analyses, calculations and decisions with multiple conditions.

Technique 4: providing examples (few-shot)

Give the model one or more examples of the desired input-output pair. By showing an example rather than only describing it, the model learns the pattern you intend. Example: 'Here are two customer reviews and the corresponding summary I want. [review 1] -> [summary 1]. [review 2] -> [summary 2]. Do the same for the following review: [review 3]'. Few-shot is especially effective for tasks with a recurring format, such as customer communications, reporting or product descriptions.

Technique 5: using prompt chains

Break complex tasks into a series of smaller prompts where the output of one becomes the input of the next. Instead of asking an AI model to write a complete market analysis in one prompt, you first ask for a list of five relevant questions, then a detailed answer per question, and finally a summarising conclusion. Prompt chains give you a checkpoint between steps, reduce hallucinations and deliver better final quality. In ChatGPT you use the chat interface; with the API or Copilot Studio you can automate this flow.

GDPR and EU AI Act

What should you never put in a prompt?

The most common mistake in business AI use is copying real customer or employee data into a prompt. Under the General Data Protection Regulation (GDPR), names, email addresses, national ID numbers, medical information and other personal data are protected. They may only be entered into an AI system if you have a Data Processing Agreement (DPA) with the provider, a lawful processing basis, and the data does not leave the EU without an adequacy decision. ChatGPT via the API (Team or Enterprise with DPA) and Microsoft 365 Copilot (EU Data Boundary) generally meet these requirements, but you remain the data controller and are liable for lawful use. When in doubt, use anonymised or synthetic data in your prompt. EU AI Act Article 4 (in force 2 February 2025) also requires you to understand what your AI tool can and cannot do: that is not just a compliance obligation but a practical necessity for evaluating output.

Frequently asked questions

Frequently asked questions about prompt engineering

What is the difference between a good and a bad prompt?
A good prompt has three elements: context (who you are, who the audience is), task (exactly what the model should do) and format (in what form the result should appear). A bad prompt is missing one or more of these elements and leads to a generic or incorrect answer. Example of a bad prompt: 'Write a text about our product.' Example of a good prompt: 'You are a B2B copywriter for a Dutch software company. Write an 80-word paragraph about our product [name] for CFOs. Name three concrete benefits and end with a call to action.'
Does prompt engineering work with all AI models?
Yes, the five techniques are model-agnostic: they work on ChatGPT Plus ($20/month), Microsoft 365 Copilot ($30/user/month), Gemini in Google Workspace (Business Standard approx. €11.50/user/month) and open-source models like LLaMA. The exact phrasing may vary slightly per model: newer models (GPT-4o, Gemini 1.5 Pro) follow format instructions more accurately than older versions.
Can I put customer data in a ChatGPT prompt?
Only if you have a Data Processing Agreement (DPA) with OpenAI and a lawful processing basis. The ChatGPT Plus subscription (consumer version) has no DPA and must not be used for processing personal data of customers or employees. ChatGPT Team or ChatGPT Enterprise including DPA are the correct business options. Microsoft 365 Copilot has a DPA via the Microsoft Product Terms and processes data within the EU Data Boundary. In all cases, you as the data controller are responsible for lawful use.
How long can a prompt be?
The maximum prompt length depends on the context window of the model. GPT-4o has a context window of 128,000 tokens (roughly 100,000 words); Gemini 1.5 Pro up to 1 million tokens. Keep prompts as short as possible and as long as necessary. A prompt of 50 to 200 words is sufficient for most business tasks. Prompt chains are better than a single long prompt when the task involves more than three or four steps.
What does the EU AI Act require for prompt engineering?
EU AI Act Article 4 (in force 2 February 2025) requires every organisation that uses AI tools commercially to train employees role-specifically: they must know how the model works, what it cannot do and when human oversight is required. For prompt engineering, this means employees must understand that AI output always needs to be checked for factual accuracy, bias and GDPR compliance. An AI register documenting the tools used is the minimum required documentation.

Want to use AI more effectively in your organisation?

Book a free call. We look together at the AI tools you already use and how prompt engineering and automation can generate more value.

Book a free call →

Free and without obligation · Reply within one business day · Fixed price up front, no surprises