Choosing a useful first AI project

Begin with a task, not a general ambition.

“Use more AI” does not explain which part of the work should improve. A stronger starting point is a recurring information task with a clear purpose. Think about questions the team regularly needs to answer, information people repeatedly search for, or a document process that involves the same preparation steps.

Describe the task in one sentence. Include who does it, what information they use, and what they need at the end. If that sentence is difficult to write, spend more time understanding the work before choosing an assistant or changing the process.

Look for a focused, reviewable result.

A useful first project has an output that someone can examine. For example, a team might explore an assistant that helps locate information in an agreed knowledge set, organizes material for a document, or prepares a summary for a person to review. These are examples to consider, not a requirement to introduce all of them.

Keep the first scope narrow. Decide which part of the task the assistant would support and which parts remain with people. A clear boundary makes it easier to discuss whether the result is useful and what still needs attention.

Understand the information the task needs.

An assistant's role depends on the information available to it. List the sources a person uses today and consider whether they are organized well enough for the proposed workflow. Notice where information is missing, inconsistent, or difficult to locate.

Before outlining the setup, answer a few practical questions:

• Which information is needed for this particular task?
• Where does that information currently live?
• Who is responsible for maintaining it?
• What should happen when the needed information is absent?
• Who can decide whether the resulting output is suitable?

Treat these questions as part of the project itself. Structuring the information can be as important to the workflow as choosing the assistant's role.

Put human review into the process.

Do not leave review as an undefined final step. Identify the person or role responsible for checking the output and describe what they are checking. They may need to confirm that the relevant information was used, that the result addresses the task, or that a document is ready for its next stage.

Also decide what happens when the output is incomplete or unhelpful. The workflow should give people a clear way to return to the source information, correct the result, or continue the work themselves. The person responsible for the decision should remain easy to identify.

Define a practical first conversation.

Bring one example of the task, the information used to complete it, and a description of a useful result. Explain where the work is repetitive and where judgment is required. Include an example that is straightforward and one that usually needs extra attention.

That material gives a focused AI discussion a foundation. It connects the proposed assistant to real work and helps define a scope around useful information, understandable output, and human control.

AI assistant or workflow automation?

Use a defined workflow when the work follows explicit rules, such as routing a complete request to an approver. Consider an AI-assisted step when the task involves interpreting or drafting language and a person can evaluate the result. Some projects need both: a controlled workflow around a reviewed AI draft.

Write a pilot brief before selecting a product

• Task: describe one output and the person who will use it.
• Sources: identify approved information and its owner.
• Boundaries: specify what the tool must not access or decide.
• Review: name the person who approves the output before use.
• Evidence: define representative test cases and acceptance criteria.
• Exit: decide when to stop, revise, or expand the pilot.

Check quality with a repeatable test set

Include routine requests, incomplete information, contradictory sources, and questions outside the approved source material. Review factual accuracy, source traceability, permission boundaries, and correction time. A polished answer should not pass when its supporting information is missing.

Keep the input, expected result, reviewer notes, and observed output for each case. When a tool or source library changes, repeat the relevant tests rather than assuming earlier results still hold.

Budget for the full operating process

Include source preparation, access configuration, integration, testing, training, and recurring review alongside subscription costs. Time saved during drafting can be offset by heavy correction or verification work. Measure the complete task before deciding whether to expand.

Explore applied AI services, learn how workflow automation can support controlled handoffs, or scope a focused pilot with FortLyn.