AI tools are now easy to add to almost any website, service, or application. That makes it tempting to start with the question “Where can we use AI?” A better question is “What do our users need, and is AI the best way to meet that need?” During Discovery, your job is to find out whether AI would help your users, who it would affect, and what could go wrong. Answering these questions early saves time and money, and it protects the trust people place in Yale’s digital services.
Before adding AI to a website or tool, confirm it solves a real user problem and understand how it will affect the people who use it.
Good AI starts with a real user need, not with the technology.
AI is a tool, not a goal. Before you decide to use it, be clear about the user need you are addressing and whether AI is the right way to meet it.
State the problem without mentioning AI. For example: “Students can’t find application deadlines on our site,” or “Staff spend hours answering the same five questions by email.” If you can’t describe the problem without naming the technology, you may not have found a real user need yet.
Look for evidence from your users, not only from stakeholders. Support tickets, search logs, analytics and interviews all show where people struggle today.
AI tends to work well when:
- People need to search, summarize, or make sense of a large amount of content.
- Answers depend on context and can’t be handled by a simple rule or form.
- Users benefit from a first draft or a suggestion they can review and change.
- The same kinds of questions come up again and again.
AI is often a poor fit when:
- Users need the same, predictable result every time
- Mistakes would be costly, such as decisions about money, health, grades, or eligibility.
- The real cause is confusing navigation, outdated content, or unclear labels. Fixing the content or structure may solve the problem faster and more reliably.
- Users value doing the task themselves or would not trust an automated answer.
Assisting means AI helps a person do the task: it suggests, drafts, or summarizes, and the person decides. Automating means AI completes the task on its own.
Many users want help with the parts of a task they find tedious, and want to stay in control of the parts that matter to them. Ask users which parts they would happily hand off and which they want to keep. In higher education, where decisions often affect people’s academic or working lives, assisting is usually the safer place to start.
Ask Yourself: Can I describe the user problem without mentioning AI?
This helps you confirm that you are solving a real need, not adding technology for its own sake.
Look beyond the primary user
List the groups the feature will touch. This includes the people who use it directly and the people affected by its output (for example, applicants whose questions are answered by a chatbot), and the staff who will maintain it, handle escalations, or correct its mistakes. Use your existing user segments and archetypes as a starting point.
Attitudes toward AI vary widely
Some people are eager to use AI. Others are skeptical, have had bad experiences, or worry about privacy. Ask users how they feel about AI in this specific situation, what they would expect it to do, and what would make them trust or distrust the result. Their expectations will shape how you explain the feature later.
AI can remove barriers or create new ones
AI can make information easier to find and understand for many people, including people with disabilities. It can also create new barriers. Chat interfaces can be hard to use with a screen reader or keyboard. Generated text can be long or complex. Outputs can reflect bias and work less well for some groups than others.
Include people with disabilities and people from different backgrounds in your research, and ask how the feature would work for them. Learn more in the AI and Accessibility recording.
Know what information the AI will see
Identify what information users might enter and what data the AI would draw on. Yale’s AI Guidelines for Staff recommend that moderate or high-risk data should only go into AI tools approved for that use. Check each tool’s data classification on the Yale AI Tools and Resources page. Also, ask users what they would be comfortable sharing, since that can differ from what is technically allowed.
Ask Yourself: Can I describe the user problem without mentioning AI?
This helps you confirm you are solving a real need, not adding technology for its own sake.
Connect the AI feature to what users are trying to do
Use the approach in Define the User Goal to describe what users need. For example: “As a prospective graduate student, I want to quickly confirm application deadlines and requirements so I can plan my application with confidence.” A clear goal keeps the team focused on outcomes instead of features.
Plan for what happens when AI gets it wrong
AI can give answers that sound confident but are incorrect. Ask what would happen if the AI gave a wrong, incomplete or outdated answer. A wrong suggestion for a lunch spot is harmless. A wrong application deadline or policy answer could seriously affect someone. The higher the stakes, the more you need human review, clear sources, and an easy way to reach a person.
Decide how you will know it is working
Agree on measures before you build, and collect a baseline now so you can compare later. Useful measures include:
- Task success: can users find the right answer or finish the task?
- Accuracy: how often are AI answers correct and complete?
- Time on task compared with the current experience
- User trust and satisfaction
- How often users switch to a non-AI path or contact a person
- Effect on staff workload, such as fewer repeated support requests
Use approved tools and processes
Review the Yale AI Guidelines for Staff and the list of Yale AI tools. Staff working on AI-driven projects should follow the AI Request process so the right safeguards are in place. Starting this early prevents delays later.
Ask yourself: What would happen if the AI gave a user a wrong answer?
Understanding the cost of errors tells you how much oversight and transparency the design needs.
Research methods for AI projects
You can use the same research methods as in any other project, with a few adjustments.
- Look for repeated questions and tasks. These often show where AI could help or where content needs fixing.
- Add questions about AI, such as “Have you used AI tools for tasks like this? What was that like?”, “What parts of this task would you be happy to hand off?” and “What would make you trust or not trust an automated answer?”
- Review site search terms and drop-off points to find where people get stuck.
- Before building anything, have a team member play the role of the AI behind the scenes and respond to users’ questions. This shows what people actually ask and expect, at very low cost.
Evaluating Existing Experience
- Understand how well the current experience works so you have something to compare against.
Summary
The best AI features start with a clear understanding of users, not with the technology. By naming the user need, checking whether AI is the right fit, understanding who will be affected, and deciding up front what success and failure look like, you set your project up to deliver something people will trust and use. Sometimes the right outcome of this work is deciding not to use AI at all, and that is a success too.
Key Takeaways
- Describe the user need before choosing AI as the solution.
- Check whether better content, navigation, or a simple rule would meet the need instead.
- Decide whether AI should assist users or automate the task, based on what users want.
- Consider everyone the feature affects, including people with disabilities and the staff who support the technology.
- Weigh the cost of wrong answers and set success measures before you build.
- Use Yale-approved tools and follow the AI Request process.