Which business process
should AI support?

Which task is a good starting point?

Look for recurring work with a recognisable output: collecting details for an offer, preparing a summary or assembling documents. Describe today’s steps and identify where delays occur. Using AI is not itself a business objective.

Illustrative example: a colleague gathers information from email and CRM, then prepares an offer. A pilot can cover only a draft for review. Sending it to a customer, changing a price and writing to a system can stay out of scope.

Is the information you need available?

Identify sources and their owners. Check that documents are current, data is complete and it may be used in the chosen environment. Define access to specific resources; connecting all email or every drive should not be the default starting point.

RAG combines information retrieval with a model’s generation of an answer. It can help use company knowledge, but does not guarantee correct answers. Sources, permissions and outputs still need checking, especially when information is missing or contradictory.

What may AI do independently?

A chatbot handles a conversation; RAG supplies retrieved information. An agent can use tools and perform agreed steps. A harness connects models, knowledge and tools and coordinates their work. These names describe mechanisms, not a ready-made set of permissions.

Agree which actions require approval and what happens when something fails. Preparing a draft can be automatic, while sending a message or changing data may need a person’s decision. Also plan how to stop the process and return to manual work.

How will you know the pilot is worthwhile?

Prepare a fixed set of examples: routine tasks, missing data and unusual situations. Record the expected result and failures you will not accept. Assess completion of the whole task, consistency with source data and correct tool use, not just fluent answers.

Compare working time, review time and costs with the current process. Include corrections and cases needing intervention. The results can then inform expansion, a scope change or stopping the pilot. There is no universal savings percentage that can be promised before measurement.

• What result counts as correct?
• Which failures stop the process?
• Who approves the output and further development?

Sources and next step

These sources describe mechanisms and good practices. They do not confirm your company’s configuration or compliance. The examples in this guide are illustrative.

• Microsoft Learn: Retrieval-augmented generation (RAG)
• Microsoft Foundry: Run evaluations

To apply these questions to your business, explore the related service. A description of your situation is enough for an initial discussion; do not send passwords or confidential data through the form.