How to brief ChatGPT for business work: context, output and checks

“Analyse our customer feedback” gives the model considerable freedom. It could summarise messages, group topics or propose product improvements. Each result might sound reasonable. Yet a manager may need something more specific: choose one problem for the next release and identify the messages that support it.

Before writing the prompt, finish this sentence: “After reading the answer, I will be able to…” Approve a form change, question a supplier, compare two proposals or locate contradictions in a procedure. This clarifies both the required output and the information that actually matters.

The model can handle a bounded part of the work, such as a first pass at labelling feedback. Priorities remain with people who understand the product, delivery constraints and customer commitments. State that division of responsibility in the task instead of asking the model to make decisions without the necessary context.

In an illustrative booking service, a customer says, “I can’t reschedule.” Without context, this looks like a feature request. If rescheduling already exists but the link is buried in an email, the problem may concern discoverability. Different starting facts lead to different proposals.

Include the process, relevant constraints and the definitions you use. Explain what counts as an active customer, which period the feedback covers and which channels appear in the sample. Where information is missing, identify the gap. Do not invite the model to silently turn an unknown into a fact.

Stable preferences such as language, tone and response structure can live in reusable instructions. Project-specific facts, current decisions and documents should accompany the task. Before sharing material, remove unnecessary personal or confidential information and check whether the chosen tool and use are permitted under your organisation’s policies.

A feedback brief might say: “Find recurring obstacles to booking in the attached anonymised messages. For each obstacle, include its topic, message identifiers and a short supporting extract. Separate what the customer reported from your hypothesis about the cause. Do not infer frequencies beyond this sample.”

You can then add: “Suggest three changes the team could test. Connect each to an observation, an expected effect and a verification method. List unanswered questions where evidence is insufficient. Do not promise revenue growth or prioritise a complaint solely because it uses emotional language.”

The useful part is the relationship between observation and evidence. A proposal is labelled as a hypothesis, and its next step is testable. A lengthy generic expert persona does not supply those conditions. One well-chosen example of the desired classification may clarify the job better than several paragraphs of abstract instructions.

Asking for a finished strategy immediately makes it difficult to locate weak assumptions. First inspect what information was extracted and how it was grouped. Then discuss interpretations, followed by possible actions. At each stage, you can correct the foundation without discarding work that remains useful.

For criticism, ask about contradictions, missing cases and unsupported assumptions in a specific proposal. A second answer from the same model is not independent confirmation. Check important sources and calculations separately, especially when the result affects money, contractual commitments or customer data.

OpenAI’s guidance on ChatGPT accuracy notes that confident answers may contain errors and invented references. Check whether a cited source exists and whether it supports the relevant conclusion. Confidence in the wording is not a quality measure.

After several attempts, keep a template with fields for the goal, input material, constraints, output and checks. Save one useful example and one failure alongside it. A new colleague then gets a repeatable way of working, with enough context to recognise when the result needs correction.

When the same task recurs frequently, ask whether it needs a connection to its data source, a result history and a handoff to a person. In Saleon, company knowledge and conversation scenarios connect to an operational workflow and manager takeover. That is a product-design question extending beyond prompt wording.

The accompanying Russian-language ChatGPT video offers further examples. For your own team, begin with one recurring task and an inspectable result. If the next step is integration into everyday operations, we can discuss the data sources and boundaries of that automation.