AI4RA Tip of the Week #28

Human-in-the-Loop: What “Automated” Really Means

When a task is “automated,” is there still a human in the loop? And if a task can’t be fully automated, can AI still help?

The short answers are yes and yes. In research administration, “automated” almost never means “unsupervised.” It means AI handles the repetitive, time-consuming parts of a process while a person stays responsible for the outcome. This is why AI augments RAs, not replaces them.

What Is Human-in-the-Loop?

“Human-in-the-loop” describes any AI process where a person reviews, guides, or approves the AI’s work at key points. It helps to think of automation as a spectrum rather than a switch:

  • AI assists, you decide. AI prepares a draft, summary, or list of flagged items, and you make every decision. Example: AI highlights non-standard terms in an award, and you negotiate them with the sponsor.
  • AI does, you review. AI completes the task, and you check the output before it’s used. Example: A workflow extracts reporting deadlines, and you verify them against the award before they go on the calendar.
  • AI runs, you audit. For well-tested, low-risk tasks, AI runs routinely and a person spot-checks results on a regular schedule. Example: Sorting incoming funding announcements by agency, with a weekly review of a sample.

Most RA work belongs in the first two categories. The more a task involves compliance, money, or a sponsor relationship, the closer the human should stay.

Where Does the Human Fit?

People play a role at every stage, not just the end:

  • Before: Defining the task, choosing the right tool, and selecting trustworthy source documents.
  • During: Reviewing checkpoints in multi-step processes, so an error in step one doesn’t carry through to step five.
  • After: Verifying outputs, correcting mistakes, and signing off on anything that leaves your office.

Why This Matters for Research Administrators

Research administration is built on accountability. When a budget goes to a sponsor or a deadline goes on a calendar, someone at your institution is responsible for its accuracy. “The AI did it” isn’t an acceptable answer to an auditor, a PI, or a sponsor.

However, tasks that can’t be fully automated, such as those requiring judgment, negotiation, or relationship management, can still benefit from AI assistance. For example, you can use AI to draft talking points with a PI, but an RA still needs to have the conversation. The judgment stays with you, and the busywork doesn’t.

How to Use This Resource

Before automating any part of your work, ask yourself these four questions:

  1. What’s the cost of an error?
  2. Where are the checkpoints?
  3. Who owns the outcome?
  4. How will you document the review?

Remember: AI augments research administrators; it doesn’t replace them. The goal of automation isn’t to remove people from the process. It’s to move their time from repetitive tasks to the judgment calls that only they can make.