Human-Centered AI: Design Work Around People

Human-centered AI means designing and using AI around people’s needs, abilities, and choices. You involve the people affected, give someone responsibility for the outcome, and use feedback to improve how the work happens. The test is whether people can get a useful result and retain meaningful control.

A team can use fenbs to keep the human decisions and lessons behind those improvements available to its connected assistants. That record helps preserve why a process changed when another person or AI app picks it up.

What does putting people first actually change?

You start with a problem a person has, such as finding the right support instructions, rather than a feature you want to demonstrate. Ask what a good result looks like for that person, what effort the process requires, and what could make them feel stuck or excluded.

Stanford HAI describes its approach↗ (opens in a new tab) in terms of human impact and augmenting people. Applied at work, that means evaluating the whole experience: the AI’s answer, the human’s understanding, and what happens when the answer is wrong.

For example, a draft-writing assistant may produce polished responses quickly but leave a new employee unable to explain them. A better design might show the supporting policy and invite edits before sending. The wider AI governance library covers the rules and responsibility surrounding these choices.

Whose needs should you ask about?

Include the people doing the work and the people receiving its results. A support representative sees the effort of checking a draft. A customer sees whether their question was understood. A manager sees repeated problems, but may miss friction that employees have learned to work around.

• Ask frequent users to show the actual task, including the awkward steps they usually skip in a demonstration.
• Include occasional users and people who need accessible language, assistive technology, or another way to complete the task.
• Ask people affected by a result how they would notice an error and get it corrected.
• Include whoever must maintain the process after the initial team moves on.

Do not treat one interview as permission to speak for everyone. Note who participated and whose perspective is still missing. For the broader division of work, see AI and humans working together; the focus here is how their needs shape the design.

Offer a way to contribute without speaking in a group meeting. Someone may notice a serious problem but feel uncomfortable challenging a manager in public. A private feedback route can reveal different needs, provided someone is responsible for responding.

How do you give people meaningful control?

Control requires a usable choice at the moment it matters. A person needs enough information to accept, change, reject, or pause the output. If rejecting it creates extra work or requires a manager’s permission every time, an approval button may offer less freedom than it appears to.

Say a small design studio uses AI to summarize client feedback. The designer should be able to inspect the original comments, mark an uncertain interpretation, and correct the summary before it becomes a work instruction. The client should have a way to clarify what they meant.

Human-in-the-loop examples show specific review points. Choose the review point because it can change the outcome, then save the agreed boundary in fenbs Decisions and rules so future assistants read the same instruction.

Who is accountable when AI helps with the work?

Name the person responsible for deciding whether the process is suitable and the person responsible for checking a particular result. They may be the same person on a small team. Either way, make sure they have time, relevant knowledge, and authority to stop the work.

Responsibility also includes listening after release. Decide who receives complaints, who can correct an affected result, and who approves a change to the process. A shared inbox without a named responder can leave an obvious problem unresolved because everyone assumes someone else is handling it.

The GAO AI accountability framework↗ (opens in a new tab) connects governance with data, performance, and monitoring. That is a useful reminder to keep ownership connected to evidence over time. For assistants that take actions, human-in-the-loop AI agents examines the review boundary in more detail.

How can feedback improve the process?

Make feedback specific enough to investigate. Ask what the person was trying to do, where the result failed, and what they did next. Separate a factual error from an unclear instruction or an inaccessible interface. Each problem may need a different change.

Keep a small set of realistic tasks and repeat them after a change. Include a difficult case and a case where the correct action is to ask for help. Compare the result with the original problem rather than measuring only whether the AI produces more text or fewer clicks.

NIST’s AI RMF Core↗ (opens in a new tab) includes gathering and incorporating feedback from affected people. In practice, close the loop: explain what changed, what did not, and why. Keep that lesson in fenbs alongside the shared work record, without copying private customer conversations into it.

What mistakes make a process less human-centered?

One mistake is asking whether people like the idea instead of watching whether they can complete the task. Another is counting agreement as success when reviewers feel unable to disagree. Invite people to demonstrate a failure and reward useful corrections, including the decision to avoid AI for a particular step.

Also check whether convenience for one group moves effort to another. A summary may save a manager time while forcing employees to correct the same misunderstanding repeatedly. Count that correction work and preserve a manual route while you investigate. Consider whose time disappears from your measurements.

Finally, keep decisions understandable when sessions change. A successor needs the feedback and reasoning, not just the latest instruction. The guide to context switching between humans and AI assistants explains why a precise work record matters at that transition.

How to do this with fenbs

1. Create a free account, then add a project for one AI-assisted process and describe the people it should help.
2. Create tasks for interviews, a small trial, and the checks still needed. Write which perspectives are missing and who will follow up.
3. Record approved boundaries in Decisions and rules: who reviews, what they must see, and when work pauses for a person.
4. Connect a supported assistant and ask it to read the project context before working. People decide and sign off; the assistant can record their choices.
5. After the trial, add a lesson describing what happened and what to do differently. Save progress with the next question or task to resolve.

fenbs is free for one person, no credit card. It is useful for retaining the work behind a decision; it does not conduct user interviews, test accessibility, or prove that a process is fair. Keep sensitive research material in an appropriate system and save only the permitted summary and next steps.