Business Prompts by Department: Templates and How to Adapt Them

Business Prompts by Department: Templates and How to Adapt Them

Good business prompts are not clever phrases. They are instructions with a role, context, task, limits and output format, written for one specific process. This guide gives you business prompts by department (sales, finance, HR and operations) and the logic for swapping each field for the real context of your company.

Key takeaways

• A business prompt is a work specification: who does the job, with which data, how far they can go and how they deliver.
• The structure is the same in every department. What changes is the context, the limits and the quality standard.
• Every template needs a stop rule: when the AI does not know, it says so and hands off to a person.
• Personal data and decisions that affect people call for extra care, with human review and a record.
• A prompt is only ready after it has been tested on real cases and compared with what your team produces today.

What makes a business prompt different from a personal prompt?

A personal prompt is usually disposable: you ask, read the answer, adjust and move on. A business prompt will be reused by several people, on different days, with different data. That means it has to produce a predictable result, not just a good answer on the first try.

This changes where you put your effort. Instead of polishing the wording, you define what goes in (which data), what comes out (which format), what is forbidden (promises, prices, decisions) and who reviews. Anthropic describes this kind of work as context engineering: choosing deliberately what the model sees for each task instead of dumping everything into the conversation. It is worth reading their material on context engineering for agents to understand the logic behind it.

Internal policy matters too. If every person pastes customer data into a different tool, the best prompt in the world does not remove the risk. Before you distribute templates, define what may and may not go in, as we describe in company policy for employee ChatGPT use.

What is the base structure that works for any department?

Every template below follows six fields. Learn this order and you can build a prompt for any process.

• Role: who the AI represents and who it works for.
• Context: what it needs to know about the company, the product and the situation.
• Task: one clear action, with the right verb (summarize, classify, draft, compare).
• Input data: what will be pasted or attached, and what must not be used.
• Limits: what it must not do, claim or decide, and what to do when information is missing.
• Output format: a fixed structure, so the team can read it quickly and another system can process it.

The limits field is what most separates an amateur prompt from a professional one. Without it, the model fills gaps with plausible assumptions, and that is where invented deadlines, discounts nobody approved and facts that do not exist come from.

Sales prompt template

Sales is the area where AI helps most with preparing and recording, and least with deciding. Good uses include preparing for a meeting, summarizing an account's history and drafting a follow-up. The template below prepares a sales contact.

• Role: you are a sales preparation assistant at (company), which sells (product or service) to (type of customer).
• Context: our process has these stages: (funnel stages). This customer is at stage (stage).
• Task: read the notes below and deliver a summary of the situation, the customer's likely questions and a suggested next step.
• Input data: (CRM notes, latest messages). Do not use information that is not in this text.
• Limits: do not mention price, deadline, discount or commercial terms. Do not invent facts about the customer's company. If information is missing, list what needs to be asked.
• Output format: situation in three lines, likely questions as a list, next step in one sentence, short WhatsApp message draft.

To adapt it, replace the funnel stages with yours and add your real qualification criteria. If the team already logs everything in the CRM, the natural next step is connecting the prompt to that data, which is the subject of how to integrate AI with your CRM.

Finance prompt template

In finance, mistakes cost more and tolerance is lower. The safest use is therefore support work: checking, explaining and drafting, never paying, approving or changing entries. The template below analyzes discrepancies in a set of entries.

• Role: you are a support analyst in the finance team at (company). You do not approve or change anything.
• Context: we follow this rule: (reconciliation rule or expense policy). Treat as a discrepancy: (company criterion).
• Task: compare the two sets of entries below and list the differences.
• Input data: (statement and entries, without unnecessary personal data).
• Limits: do not recalculate amounts on your own without showing the calculation. If an item is ambiguous, mark it as pending review. Do not suggest payment.
• Output format: a text table with item, discrepancy found, probable reason and confidence level (high, medium, low).

The adaptation here is in the rule: the prompt only works if your company policy is written clearly. If the rule lives in one person's head, the first job is to document it. When you connect your management system, start with read-only access, as we explain in integrating AI with Brazilian ERPs such as Bling, Omie and Tiny.

HR prompt template

HR deals with people and personal data, so the prompt needs to be more restrictive. AI can help draft job descriptions, organize answers from internal surveys and summarize policy documents. For decisions about candidates or employees, human review is mandatory and the design needs attention to bias and data protection law, as we cover in AI resume screening: bias and compliance for HR teams.

The template below summarizes resumes against objective requirements, without ranking people.

• Role: you are a recruiting support assistant at (company). The person responsible for the role makes the decision.
• Context: the role requires (objective, verifiable requirements). The nice-to-have requirements are (list).
• Task: for each resume, indicate which requirements are supported by the text and which are not.
• Input data: (resumes with identification reduced to what is necessary).
• Limits: do not infer age, gender, origin, marital status, health or any personal characteristic. Do not assign scores or rankings. Do not recommend hiring or rejecting.
• Output format: a list of requirements with status (supported, not supported, not stated) and an excerpt from the resume backing each item.

Adapting this means rewriting the requirements so they are observable. A requirement like proactive profile cannot be verified in the text, and the AI will invent a criterion. Experience with (tool) for at least (period) can be verified. For data handling, see also the guidance from the Brazilian data protection authority (ANPD) and our guide to LGPD and artificial intelligence for companies.

Operations prompt template

Operations is where there is the most repetition and therefore the most to gain. Good uses include turning procedures into checklists, summarizing the day's incidents, classifying tickets and preparing shift handovers. The template below classifies and summarizes incidents.

• Role: you are an operations assistant for (unit or area) at (company).
• Context: the incident categories are (closed list). The urgency levels are (definition of each level).
• Task: read the incidents below, classify each one by category and urgency level, and summarize what is still pending.
• Input data: (shift records).
• Limits: use only the categories on the list. If none fits, answer with no category and explain why. Do not decide the priority of safety-related incidents, send them straight to the person in charge.
• Output format: one line per incident with category, urgency and pending item, followed by a shift handover summary of up to five lines.

Adaptation is easier when the category list already exists. If each person names the problem differently, standardize the names first. In specific sectors the logic changes little, as the articles on how to start with AI in manufacturing, beginning with maintenance and quality and on AI for restaurants show.

How do you adapt a template to your company's context?

Copying the template is only the start. Adaptation follows a short sequence that also works as a review checklist.

• Step 1: describe the current process in a few lines: who does it, how long it takes and where it tends to go wrong. This becomes your baseline.
• Step 2: replace the placeholders in parentheses with real company terms: stages, categories, policies, system names.
• Step 3: reduce the input data to the minimum necessary. If a field does not change the answer, remove it.
• Step 4: write the limits from mistakes that have already happened or that would be serious. Each limit should have an alternative action (hand off, ask, mark as pending).
• Step 5: build a set of test cases with common, difficult and forbidden situations, and run the prompt against them.
• Step 6: compare the outputs with what the team would produce on its own, and adjust the prompt, not the test case.
• Step 7: define who reviews the output and what happens when the reviewer rejects the result.

Illustrative scenario: a services company with a small sales team adapts the sales template. In the first test, the message draft promised a visit date that nobody had confirmed. The fix was not changing the AI model. It was adding a limit to never mention dates or deadlines, and adding an output field with the questions the salesperson must answer before sending. The second test passed the difficult cases, and only then was the prompt released to the team.

How do you test, protect and maintain these prompts?

A prompt is a company asset and needs an owner, a version and tests, like any other procedure. Record who created it, when it changed and why. Keep the test cases alongside it, because any change of model or wording can alter behavior. OpenAI's guide to evals shows the general idea of testing outputs systematically, and it is a useful reference even if you use a different tool.

There is also a security risk. If the prompt processes text from outside, such as emails, customer messages or resumes, that text may contain hidden instructions trying to change the AI's behavior. The OWASP cheat sheet on prompt injection describes practical defenses, and the main one is simple: clearly separate instructions from data, and do not give the AI permissions it does not need.

In the way we work, the method has four steps that apply directly here. Map the process and the baseline. Limit the data and actions of the prompt. Pilot one real flow, recording the outputs and agreeing on a success metric. Hand the operation over to the team, with documentation and a decision to expand, adjust or stop. To measure the pilot, see how to measure an AI automation pilot.

Frequently asked questions

Do I need a different prompt for each person on the team? No. The ideal is one template per process, maintained by one owner and used by everyone. Individual variations make results less predictable and reviews harder.

Can I put customer data into a prompt? It depends on the tool, the contract and the legal basis for processing. Send only what is necessary, avoid identifiers when they are not essential and define in internal policy what is allowed. When in doubt, treat it as a data protection question and involve whoever handles privacy at your company.

Does the same prompt work in any AI model? Almost never with the same result. The structure holds, but tone, adherence to limits and output format can vary. When you switch models, rerun your test cases before releasing.

When is a prompt no longer enough and I need an agent? When the process requires querying systems, taking actions or keeping context across several steps, with logging and human approval. As long as the task is reading, summarizing and drafting, a well-tested prompt is usually enough. If volume or risk grows, map the flow before you decide.

Sources

• Anthropic: Effective context engineering for AI agents
• OWASP: prompt injection prevention
• OpenAI: evals guide
• ANPD: Autoridade Nacional de Proteção de Dados