Emily Gracke, EG, Host [00:00]: Hi everyone, welcome to The Big AI Debate, where we take on some of the biggest questions facing the people working at the forefront of technology.

I am Emily Goracke, Senior Conference Producer for The AI Summit Series. Today, we are asking what happens when AI stops helping us make decisions and starts making them for us.

Traditionally, AI has been a tool. It is given information, it returns an answer, and the human decides what happens next. But that conversation is changing. AI can now assess a situation, decide what should happen next, and increasingly take action on its own.

So the big question today is this: when AI makes the decision, who is responsible for what happens next?

My guests today are Sidharth Thapar, AI Manager and Senior Applied Scientist at Staples, and Daniel Beecham, CEO and founder and an adjunct professor teaching AI for business at Georgia Tech.

Guest introductions

EG [01:15]: I would love for each of you to introduce yourselves, share a little about who you are, and explain how you got into AI. Sidharth, let me start with you.

Sidharth Thapar, ST [01:25]:  Thank you, Emily. I am excited to be here.

I have about ten years of applied AI experience across several industries, including security, manufacturing, and retail. My path into AI started when web development was one of the hottest areas in tech. Around that time, I came across professors who had been working in deep learning and machine learning for years.

If you go back to that period, GPUs and CPUs were becoming more affordable. The capabilities had existed for a long time, but they were not yet practical for everyday use. Once they became feasible, the field became incredibly interesting to me. I could see that something big was coming.

I started building adaptive systems and gradually moved into regression and classification. After my undergraduate degree, I realized I needed to specialize if I wanted to become truly strong in the field. I pursued a master’s in computer science with a specialization in AI. At that time, dedicated AI degrees were still uncommon, so the route was more general. I studied at Northeastern and also took advantage of courses at MIT and Harvard. That is how I entered the field, and I am still enjoying the ride.

I never imagined the landscape would look like this ten or fifteen years later, but it is a very exciting space to be in.

EG [03:01]: It sounds like you have seen several major shifts in the field. Daniel, tell me a bit about yourself.

Daniel Beecham, DB [03:09]: I am also an adjunct professor teaching AI for business at Georgia Tech.

After college, I joined a small startup in Seattle that focused on AI for marketing technology. That was roughly ten years ago, which was still fairly early in the modern AI cycle. The company failed badly, so I moved into consulting. Even there, I kept ending up around AI products.

One of the first was a predictive analytics platform, which today would simply fall under the broader category of AI. Over time, I worked on many digital transformation and data governance initiatives. Eventually, I saw enough maturity in multi agent architectures to feel it was time to start something of my own. I found a strong co founder and moved into the procurement space.

EG [04:10]: You both bring strong and very different backgrounds to this discussion. Let us jump into the first question.

Who owns AI decisions

I want to start with the basics. AI is now built into almost every platform and industry we touch. But the real issue is what I said at the beginning: AI can assess, decide, and act.

So is it still just a tool?

A lot of enterprise leaders want to understand who owns AI now and who is actually making the decisions. Daniel, where should accountability sit when multiple people are working on the same agent or system? And how involved are organizations today?

DB [05:28]: On a practical level, the person using the tool should be accountable for the outcome, assuming the result is actually used for a decision or delivered as part of the work.

There are other layers of accountability when you look at frontier labs and the people building the foundation models themselves. That is a separate and important discussion. But for businesses, the key principle is straightforward: the person doing the work has to own the result.

EG [06:15]: Sidharth, from a technical perspective, is there a way to isolate where a decision went wrong inside an AI system? Or is this simply something we need to accept as a future challenge?

ST [06:36]: That is exactly the right question.

I agree with Daniel that organizations should clearly decide who is responsible for each outcome. But there are technical ways to make those decisions more secure and more traceable.

The real challenge is how to get the most out of the autonomy we give these agents. I think we need a selective autonomy framework. A system might retrieve information, rank it, reason over it, and then produce an action. In the beginning, you should always start with a human in the loop.

Even if the first version of a model performs extremely well, you should never depend on it completely. Once you see that it is reliable in the vast majority of cases, then you can introduce logging systems, tracking systems, and automatic escalations that make it easier to detect when something is going wrong, even if the agent is making the final choice.

One more thing is extremely important from an engineering standpoint: rollback capability. A system may reason well and still make a bad decision. Humans do the same thing. No one is perfect.

In production systems, we always have response timelines and escalation structures. There should never be only one person responsible. You need level one, level two, and level three escalation paths. Those should be tracked automatically, and they need KPIs and metrics so you can evaluate whether the action was actually correct.

What human in the loop really means

EG [08:41]: That connects directly to a bigger concern for many enterprises, especially as cybersecurity becomes more important across every organization. Companies want to be prepared if something goes wrong.

That brings us back to the human in the loop discussion and how we train people to work with AI. Daniel, do you think we have to retrain the human brain to understand how to give AI the right objectives? Does AI truly understand what we are asking it to do?

DB [09:40]: I have a slightly provocative view on this. I do not believe the average person needs to know much about AI at all.

We live in a bit of an AI bubble. If you walk into a coffee shop or grocery store, most people are just living their lives. AI is present in the background, but they are not thinking about it in technical terms.

When I think about the future of AI products, I think about invisible experiences. Those will be the most successful. We are already starting to see some of that in voice interfaces and in products where retrieval is improving steadily.

It is not perfect yet, but it is getting better. At some point, retrieval will become accurate enough for everyday use and for many business functions. So when you ask what people need to know, I think the better question is what is likely to happen next. The average person does not need to understand the mechanics. We do, and that is why we are here.

EG [11:22]: That takes us back to the human in the loop conversation, which has become central in the last several months. People define it in many different ways. Some define it technically. Others talk about it from the perspective of what is good for humanity.

Sidharth, what does a human actually need to know or oversee when working with AI so that the action truly matters?

ST [12:23]: We should think of these agents as superpowers that assist us, not as tools that take away our critical thinking.

They can make us far more productive. I often think of it this way: almost everyone can become dramatically more effective if they use AI the right way.

At the same time, adoption at the enterprise level comes with skepticism. People worry about environmental impact and data centers, and those are legitimate tradeoffs. We need to be honest about the cost and the impact, while also recognizing the productivity gains.

The core issue is how to improve productivity without losing quality. Should a human touch every decision? Probably not. But we can never skip auditability. Even if a human is not reviewing everything live, there should still be an auditing system.

At this stage, generative AI is quite good at handling repeated tasks, even some that require deeper reasoning, as long as the objective is framed clearly in natural language.

EG [14:14]: Many enterprises now feel that AI can do the work of fifteen, twenty, or even a hundred people because the output is so fast. Do humans still have the ability to keep up with that pace? Do we still need a human touchpoint to make sure AI is making the right decisions?

ST [14:50]: It comes down to management. You delegate tasks that can be delegated, and once you delegate them, you do not review every single decision in detail.

If you are trying to manage several use cases at once, you cannot touch every action. You have to think about it the same way you would think about managing any other system or team.

AI, education, and the next generation

EG [15:09]:

Daniel, speaking as a professor, how do you think this affects the next generation entering the workforce? They are growing up with AI in every platform they use. Are they behind because they may not understand older processes, or are they ahead because they have daily experience with these systems?

DB [15:51]: They are absolutely the first real generation to grow up and learn with generative AI behind them, so they do have an advantage.

What drives me crazy is that many enterprises have reduced junior hiring. That makes little sense to me. These young workers may not do the exact same work agents are doing, but they could become better managers of agents than many current managers are at managing people.

The level of responsibility is rising. Entry level work is shifting upward. You can see that across many functions. AI has raised the baseline. Suddenly, everyone can produce a decent piece of writing. That means the average has improved, but so has the ceiling.

The real question becomes: how do you go beyond the new average? That is where actual skill, judgment, and the ability to use AI well become important.

Another interesting point in education is that there is research moving in both directions. Some argue AI can lower intelligence because people outsource too much thinking. Others argue that societies that value education will now dramatically outperform those that do not. That gap may become much more significant.

EG [17:55]:

That is a difficult issue for many large organizations. They are asking what it means to hire younger talent that may know a lot about AI but still be new to the workplace.

At the same time, many professionals feel AI creeping up on them. They want to know whether they are educated enough to stay competitive in the market. It feels like everyone is asking what the next step should be in their career and education.

DB [19:05]: I tell my students that one of the three keys to success is adaptability.

On my way here, I saw a quote that said when a new technology takes hold, you can choose to be part of the steamroller or you can choose to be the street. That captures it perfectly.

I can either adapt to this technology or pretend it is not happening. Personally, I would much rather stay on the leading edge than spend my time catching up or let other people outperform me. I am competitive, so I know which option I prefer.

Lightning round

EG [21:45]: We are introducing a lightning round to the show. I want to know what you would let AI do and what you would not let AI do.

Daniel, would you let AI negotiate a major supplier or customer contract?

DB [22:04]: Not today. Give it a few years.

EG [22:10]: Sidharth, would you let AI access and act on sensitive customer data?

ST [22:19]: No

EG [22:20]: Daniel, would you let AI hire or make an employment decision for a person?

DB [22:35]: Honestly, it might be less biased than I am, and that makes the question harder. But I still would not want AI making that decision for me. I would rather have a person who is truly good at hiring handle it.

EG [22:59]: Sidharth, would you allow AI to hire another AI agent to work beside it?

ST [23:11]: Sure.

Governance, regulation, and accountability

EG [23:12]:

That takes us into a much bigger and more sensitive topic: regulation.

People talk about sustainability, sovereignty, and safety, but one question often gets skipped. What are we actually trying to regulate? Are we regulating AI itself, its capabilities, how it is used, or the decisions it makes?

Sidharth, from a technical perspective, what are the fears and challenges if regulation comes in after companies have already invested heavily in building these systems?

ST [24:38]: I am a strong believer in regulation.

I do not think we only need enterprise governance. We also need government level regulation. Until we have real governance, we are going to struggle with bias, safety, and misuse.

People often focus on whether the models are capable. I do not think anyone doubts that they are. The harder question is how we remove bias and establish safe frameworks. At this point, the engineering challenge is less about making the models stronger and more about making the systems around them more reliable.

EG [25:31]: If each of you could write one rule for AI today, what would you focus on? Daniel, let me start with you.

DB [25:56]: I want to take a step back first.

I think regulation can become a race to the bottom. How much is enough? Who gets regulated? A lot of regulation happens because it serves someone’s interest. So if we are slowing things down in service of someone else’s interest, are we really doing what is best for society?

That is why regulation is hard to discuss precisely. Who benefits? What does slowdown actually look like? Can anyone measure it? Unless there is some clear and illegal coordination, companies are still competing with one another. So what is the point of reducing speed if the average consumer is not the one gaining from that decision?

EG [27:10]: Sidharth, what is your response to that?

ST [27:17]: I agree that it is a difficult problem. Governments are not always made up of people who are experts in large language models, and that is exactly why we need stronger technical participation in governance.

We have already seen cases where features were launched too quickly without enough thought behind them. That is exactly my point. We can build and release these capabilities very fast, but that is where AI engineering and AI operations need to step up.

Even in traditional software systems, we rely on fallback mechanisms, logging systems, and fast rollbacks. We need the same discipline here. We should distinguish between work that can be automated safely and work that should remain with humans.

This is not a new question that only agent systems created. We have been dealing with this since machine learning became effective. But now the stakes are higher because the capabilities are broader.

Governance is essential. It can be as simple as protecting yourself from burning through an annual token budget in a single day, or as serious as preventing a harmful action from reaching the public. We are launching too many products too quickly.

Maybe the right answer is for the broader technical community, including open source communities and industry experts, to help create frameworks for different classes of models and systems.

EG [29:23]: Daniel, you have been emphasizing accountability throughout. How do you see that fitting into this discussion?

DB [29:25]:  Accountability matters much more to me than governance.

The internet is a great comparison. It has done tremendous good, but it also enabled bad actors. The same thing happened with email, telephones, and every other communications tool. Scammers once went door to door. Now they send messages instead.

Bad actors exist everywhere. The key is that people must be accountable for what they do. That is more important than trying to regulate the technology itself. It is about who takes the action, not the tool they used.

EG [30:15]: I agree that this is a major issue for society to understand. One of the bigger questions may be whether we should spend less time defining what AI is and more time defining what AI is allowed to do, who is allowed to make those decisions, and who the right people are to guide it.

DB [30:50]: Or maybe we pick an AI overlord to assign the rest of it.

More seriously, I think we sometimes talk as if there is some magical group of superhuman decision makers who can make these choices better than everyone else. Who are we to assume that? And then, on top of that, to imagine governments are ready for it when we already know the level of AI expertise there is often lacking.

EG [31:31]: That is why this remains such a massive and ongoing conversation. We need to keep bringing it up not only in enterprises and infrastructure discussions, but across society more broadly. Who will make these decisions? What are the guardrails? Who are the best minds to put together to shape them?

DB [32:04]: A lot of those decisions will be driven by consumers.

When companies stop aligning with people’s values and needs, they fail. We have seen that repeatedly in business history. I think people will decide what kind of future they want with AI, and then work backward from that by choosing which companies deserve their money and trust.

EG [32:46]: Sidharth, any response to that?

ST [32:51]: I agree that consumer expectations need to evolve too. Governments and enterprises both have limits.

If companies release a feature that people continue using, they are not going to stop unless someone makes them stop. Their systems will always prioritize revenue and margin. That is the reality for most companies.

That is why I think there has to be collaboration. Enterprises are in a race. If they do not keep moving, someone else will. But we still cannot avoid governance. It has to exist at a central level. If governments do not have enough expertise, then industry experts may need to be part of building those frameworks.

What often happens is that we launch things first and then roll them back later. But sometimes the damage is already done. We should absolutely be willing to roll back when risk becomes clear. The problem is that sometimes it is too late.

Think about recent data leaks. Once information is out, you cannot really take it back.

EG [34:16]: Exactly. Several major companies have recently had public trust issues for the very reasons you are describing.

ST [34:35]: And that is why I look at data engineering as a more mature discipline right now than machine learning engineering or AI engineering. Data engineering has been around longer. Data governance was not a major topic ten years ago, but now no enterprise application exists without it because we have seen the downside.

AI is a bigger challenge, and it is not solved. But even if progress is slower, it has to be controlled. I do not think there is another way.

EG [35:20]: Daniel, I can see you have thoughts.

DB [35:24]: I respectfully disagree with some of that.

Consumers do have power, as long as they have real alternatives and we do not end up with monopolies. If a company loses trust because it failed in data governance or in the way it governed its models and agent systems, then it loses customers and revenue. It either evolves to meet customer expectations or it fails.

To me, that is a more effective governance mechanism than gathering a room full of people we have decided are experts and asking them to make choices for everyone else.

ST [36:16]: But sometimes the damage is already done.

DB [36:17]: What damage, though? Sometimes innovation feels like ripping off a bandage. Think about farmers when automation first changed agriculture. That transition was hard, but it ultimately improved output and living standards.

If AI is doing something similar, are we just trying to delay the inevitable?

ST [36:42]: I think data leakage is a strong example. Once that happens, it cannot really be undone.

DB [36:53]: Data leaks existed long before large language models.

ST [36:57]: Exactly, and that is my point. Data engineering is more evolved because it has been around longer. We learned through experience that without governance, we cannot rely entirely on consumers or companies to step in. People will continue using useful new features. But some forms of damage are irreversible.

EG [37:26]: As the moderator, I am going to say that I agree with both of you.

That is why these conversations matter. You both make excellent points, and I think if we asked a broader group of people working through these issues, we would see a very even split. These discussions are exactly what help move the field forward.

Governance matters, and so does public response. If enough people reject something, a company will fail. So both of those forces matter.

Closing reflections

EG [39:00]: We are almost out of time, so let me take us to the final question.

We have talked about decision making, accountability, objectives, human oversight, and regulation. So moving forward, should we give AI more decision making power, or should we slow down and put some limits around it?

DB [39:31]: I think a lot of judgment calls will inevitably go away.

There are simpler examples, like choosing a coding language for a technical problem. People may stop spending much time on decisions like that. But there are also obvious boundaries. Should an AI agent decide whether I should marry someone? No, I do not think so. Some decisions will stay personal and human.

Other decisions are easier to delegate. If AI wants to suggest where I should eat dinner tonight based on good reviews and my preferences, that seems fine.

EG [40:21]: The easy decisions are nice. The bigger ones are where it gets difficult. Sidharth, where do you draw the line?

ST [40:33]: I agree with Daniel. Easier decisions are the natural ones to automate, and we should absolutely use AI there because it will increase productivity.

But more difficult decisions need to be monitored closely, even if we start delegating parts of the process. As systems improve, we can scale up. Still, adoption is unavoidable. We are in a world where if you do not adopt these tools, you are going to fall behind.

EG [40:59]: Sidharth and Daniel, thank you both for joining the very first episode of The Big AI Debate. This has been a wonderful conversation and an important one as we continue figuring out how to approach AI and what outcomes we want from it.

As always, the views expressed here are those of the speakers and not necessarily those of their companies, organizations, or institutions.

I think the question we leave the audience with is this: as AI becomes more capable of making decisions for us, how much are we willing to give up? Where is the ultimate limit?

To everyone listening, thank you for joining the conversation. The future of AI will not be shaped by technology alone. It will be shaped by the decisions we make, the permissions we give, the control we keep, and the line we draw.

Until next time, stay curious, stay inspired, and keep exploring what is next in AI. I am Emily Goracke. Thank you for listening.