4 Ways Agentic AI Will Reshape Higher Ed Work

Gerry Bayne: Welcome to EDUCAUSE Exchange, where we focus on a single topic from the higher ed IT community and hear insights, perspectives, best practices, and more. Generative AI entered higher education largely through tools that answered questions, generated content, and responded to prompts. Agentic AI raises a different set of possibilities. These systems can carry out tasks, connect multiple steps, use tools, and act across workflows with varying levels of human oversight. As colleges and universities begin to experiment with these capabilities, the questions extend beyond what AI can produce. They include how institutional work may change, who is accountable when agents take action, and what happens to human expertise as more execution is delegated. In recent conversations I've had, four themes have emerged around how agentic AI could reshape higher education work. The first theme on our list is that workflows become less linear and more autonomous.

Mark Daley: So past automations were brittle. You can automate an exact process, you can have a little bit of decision-making capability, but it came down to pretty simple classifiers and usually pretty simple rules. So our classic robotic process automation, take a business process, make it robotic, and robotic's the right word because it sort of feels stilted and very brittle. What we have now is general purpose machine intelligence that can reason kind of like a human assistant would in that it can make wise decisions based on a complex context, which is not something you get with robotic process automation. It's either A or B, something like Claude Fable running an agent or Anthropics Fable can really take in a complex set of inputs and make a best decision. It may not be the decision you would make, but that could be true of a human colleague too.

Szymon Machajewski: An agent can provide scale. So let's say that you do something once and you want this repeated multiple times. Agents then are very good at doing that. Agents also allow you to respond to events without you have to be awake. So let's say that in your class, you want to monitor your discussion forum and you want to make sure that if someone says something that is tricky in some way and you would set the rules for it maybe using words that don't belong, agents can also chain multiple tasks. In the past, especially in 2022, we asked a question and LLM told us what it knew. Today that doesn't happen. You ask a question and it bounces like in a kind of a game machine, bounces the ball all over, multiple questions, criticisms, and then something comes out so that most pro accounts at this higher level will say, "We'll notify you when we're done."

Mark Daley: The place I'm seeing it most prevalently in the academy right now is in research. And of course not across the research endeavor, but agentic AI has been around in a usable form long enough now that our top researchers who are also early adopters are really starting to engage with this technology and use it in the research enterprise. And it's everything from the beginner baby's first steps in agentic AI is tools like ClaudeCode and OpenAI's Codex where you're using agentic software engineering tools all the way through to colleagues who have created an entire ecosystem of agents. So every morning this agent goes and searches the literature and brings me a summary and then passes it off to another agent. And so there really is a broad range. But to be clear, this is probably like 5% of the researchers in the institution at this point. But I bet next month it'll be 10%.

Gerry Bayne: Our second theme is that accountability becomes an ownership and design problem.

Mark Daley: It's looking like we're going to be going into a world where there isn't a handful of institutional agents. Each individual employee is going to have their own little army of agents. And I mean that's already happening in the shadow IT, shadow AI world. These researchers I'm talking about and probably some staff colleagues who are afraid to say it, but are already using agentic AI, things like claude co-work. So how do we get our arms around all of that? And I think the right mechanism for accountability is if it's your agent, you are accountable for what that agent does. Full stop.

Szymon Machajewski: Now we're trying to find the guilty party between maybe the end users. Is it a student, is it the instructor, is it an institution? Or maybe it's the vendor. And I think we need to broaden our views in that at the end of the day, I believe that it is the designer. So institution typically will be the one that picks the tools or the institution, especially in the academic environment, is the one that creates the policies. These need to carry most responsibility. Now, when we talk about students, we invite them to teach them, to change their mind. In fact, we invite them to fail, and it's our responsibility to create safe failure for them. So if they activate certain tools and if we do not have productions against them, I believe that the fact that students are failing needs to be very much a shared responsibility.

Gerry Bayne: The third theme on our list is that human expertise shifts from execution towards judgment.

Mark Daley: Everyone's going to be a manager. No one's going to be an independent contributor anymore. Domains where technical expertise was the absolute requisite. So to be an accountant, there's a lot of technical expertise you have to have to be an effective accountant. More and more of that technical expertise is going to be outsourced to machines. You're still going to have to have that expertise so that you can oversee the agent and know that what it's doing is right. But the more you outsource, the restier your own skills are going to get. So balancing that is going to be really, really tricky. So everyone's a manager. Everyone has to watch that their technical skills don't atrophy. And along with that, we have to be really careful about agency laundering. So it's very easy to say human in the loop, so everything's fine. But as a human in a loop, if the agent comes back to me and says, "Hey, is this okay?" And I look at it and it's okay. And then after seven times, this agent's really smart. It's always right. I'm just going to start clicking approve without really paying too much attention. Not because I'm a bad human, but because my attention span in a day is limited and the agents isn't. And the agent will keep coming back to me, approve this, approve this, approve it, and I'm just going to start clicking approve. And from a governance standpoint, I can go back to my board and say, there is a human in the loop, but actually practically there isn't.

Szymon Machajewski: The cognitive triage will be the basic skill that we have to teach and we have to adapt ourselves. Stop thinking about AI as a cheating machine. And once we triage these thoughts, that's the first level. But at the end of the day, when it comes to the results, are students cheating with AI or are they actually learning? That's a design question. So we should not expect that students are going to make different decisions that business people do. So if in a business office someone is going to generate an email or an essay, whatever it might be, and it is exactly what their boss is looking for, we would expect students to basically what we are asking a student to do is not sustainable because today in the business world, working without AI, so a mind without a co-thinking synthetic mind does not exist in the wild, or the ones that do are going to see their businesses go backwards. So if this does not exist in the wild, why would we be teaching it?

Gerry Bayne: And the final theme is that save time has to go somewhere.

Szymon Machajewski: When in student support, so in advising or in tutoring, when in student support, we are able to utilize agents so that these advisors who now are really, really on the burnout edge most of the time, when we see fewer cases being needed to handle by them, when they have a little bit extra time maybe to spend on their own mental health or the mental health of the students, are we going to add more cases to their load? Or are we going to allow this technology to have them focus on the more important things? Which to me, enrollment, all those things are important. The mental health of students is something that we have not had time for. We have not had enough professionals. We did not give them enough attention towards.

Mark Daley: This is an old story in organizational economics and psychology. Technology that's supposed to free up time inevitably creates new demands because you are always competing against the person who will take the advantage and not say, now I'm going to take Fridays off. Now I'm going to sell even harder on Fridays and I'm going to out-compete everyone else. So that economic ratchet is there. I don't think my overall workload has changed. I think where I spend time in the workload has changed. And so if I measure workload the way I used to in podcasts, in academic papers as output, then it looks like I'm working harder. It looks like I'm working more hours, but that's because I've got a 2010s mindset about what goes into creating those products. If I actually look at the number of hours I'm working, it's about the same. I'm just doing different stuff than I did before, and I'm more productive with my old KPIs.

Gerry Bayne: Four ways agentic AI will reshape higher ed. Workflows become less linear and more autonomous. Accountability becomes an ownership and design problem. Human expertise shifts from execution towards judgment, and time saved with AI has to go somewhere.