Many teams already use ChatGPT, Claude or Gemini at work. Admin staff draft customer replies, finance staff summarise reports, marketing writes first drafts of captions. The results are often uneven. One person gets a clean answer they can use straight away, while a colleague on the same team gets something generic that has to be rewritten.
The difference usually comes down to how the request is written. Prompt engineering for business teams is mostly the habit of giving the AI enough to work with: who the reader is, what the goal is, what a good result looks like and what format it should take. This article covers writing good prompts, checking the output, and keeping proven prompts as shared team templates.
Context in prompt engineering
The AI only knows what you tell it about your business. "Write a sales email" gets you a generic sales email. The same request becomes far more useful once you add who is sending it, who is receiving it, what is being offered and what the relationship already looks like.
Take a medical equipment distributor in Semarang that wants to introduce a new product to the clinics it already supplies. A better prompt would read something like this: "You are helping the sales team of a medical equipment distributor. Write a short email to a doctor who owns a clinic and has been our customer for two years. We want to introduce a new model of digital blood pressure monitor. Keep the tone polite and low-pressure. Offer a demo visit next week."
A few things are worth considering every time you write a prompt:
- Role: who the AI is helping, and in what capacity.
- Reader: who will receive the output, and how much they already know about the topic.
- Goal: what you want the reader to do or understand afterwards.
- Constraints: things not to mention, maximum length, and terms that must be used.
- Source material: the data, meeting notes or documents the answer should be based on.
Show an example of what you want
The quickest way to get the AI to write in your company's style is to show it. If you have three customer replies you think are just right, paste them into the prompt and ask the AI to follow the same style. This is often called few-shot prompting.
Examples work better than long descriptions of tone. "Friendly but professional" means different things to different people. A real example shows how formal the greeting is and how the message closes.
Keep in mind that the AI tends to copy examples closely. If every example you give is about a late delivery, replies to other complaints may start to sound like delivery apologies as well. Pick examples that cover a few different situations.
Decide the format up front
Output is easier to use when the format is clear. Say whether you need paragraphs, bullet points, a table or text that is ready to paste into WhatsApp.
When the output will be processed further, format matters even more. An admin at a cleaning services company who summarises daily reports from site supervisors could ask for a table with columns for location, issue and follow-up. It can go straight into a spreadsheet.
When you ask the AI to read a document, ask it to keep what the document says separate from its own opinion. A simple line such as "if the information is not in the document, write not found" is enough to reduce the risk of the AI filling gaps with guesses.
Check the output before using it
AI can write very confidently about things that are wrong. Numbers, names, dates and references to rules are the parts that most need checking. Get the team used to rereading AI output with these questions:
- Do all the numbers and names match the original source?
- Does it make any claim or promise you never made, such as a discount or a delivery date?
- Is the tone right for this particular recipient?
- Does it include customer or internal data that does not need to be there?
- Would you be willing to sign this on behalf of the company?
The last question tends to be the most useful. AI helps with the draft, but responsibility for the content stays with the person who sends it.
On data, set a simple rule about what can be pasted into an AI tool. Personal customer data such as national ID numbers (NIK) or bank account numbers should not go into public AI services. This follows the principles of Indonesia's Personal Data Protection Law (UU PDP).
Shared prompt templates for the team
A good prompt is wasted if it only lives in one person's chat history. Collect proven prompts into templates the whole team can use. The format can be simple, such as a shared document grouped by type of work: complaint replies, meeting summaries, draft quotations and weekly reports.
A good template has a fixed part and a part to fill in. The fixed part holds the role, writing style, format and constraints. The fill-in part holds the specific material, such as the text of a complaint or the notes from a meeting. New staff can then produce solid drafts from their first week.
Assign one person in each team to look after the templates. When a new product launches, a policy changes or the output starts to drift, this person updates the template and lets the team know. Record the date of each change so it is easy to trace why results have shifted.
Next steps
Good prompts are enough for a lot of everyday work. Once a team is repeating the same task hundreds of times, or the AI needs to read data from company systems, it usually makes more sense to build the AI into the application itself, with clear rules and proper testing. Averoa helps companies take that step through its AI integration service.
