← Harvard Data Science Review Podcast
Harvard Data Science Review Podcast
Masterminds and Mindware for Agentic AI: Contextualized and Applied
00:27:28
Loading summary
Transcript29 lines
- [00:01]
A
Hello and welcome to the Harvard Data Science Review Podcast. I'm Liberty Vittert Capito, the feature editor of the Harvard Data Science Review and I'm joined by my co host and editor in chief, Xiao Li Ne. Agenic AI is moving fast from systems that assist us to systems that develop ideas and make decisions. But what does this mean for our work and where is it taking us next? Today we're diving into Agenic AI with two people who are helping define the field. Dirk Hoffman and Ella Cruz. They are the co founders and co CEOs of Dean Studios. They have written extensively on the topic for the Harvard Data Science Review and together with the Harvard Data Science Initiative, they've created a two and a half week online course, Agenic AI contextualized and Applied. In this conversation we'll unpack what makes AI truly agenic and how to apply it responsibly, where it adds real value and how to work effectively alongside these systems. Whether you're building these systems, deploying them in your own organization, or using them day to day, this episode is for you.
- [01:12]
B
Well, thank you Ula and Dirk for joining us. So I'd like to start by asking first tell us a little bit about your company, Dan, and particularly explaining the name because I always find the names fascinating and just give our audience, which is mostly data scientists, a broad sense of the business you're in.
- [01:34]
C
Well, thank you Charlie, it's great to be here. We got the idea about Dane and Dane stands for data, AI and insights at some point in late 2015. Initially it's three founders, so we are two Finns and one German and we got to know each other at Nokia, the mobile phone Nokia company back in the day, which was now in hindsight actually doing very, very interesting things in AI and really truly big data, with maps and with supply chain and all these services applications that we used. So that was the start and we were looking for really making an impact on how companies transform themselves using data and AI. And that's how we got together and how we got started. And over these 10 years we have now grown. So we have a team of data scientists, data engineers, strategists, software engineers, BI developers, and basically we try to think of ourselves as an end to end consultancy where we offer both the strategic help and then also do implementations as well.
- [02:50]
B
Back in 2020 you threw a published article in Harvard Data Science Review titled how to define and execute your data and AI strategy. Now, five years may not seem long, but in the space of AI that's like ancient and so the question I have for you is what were the core arguments you were making then and what are the concerns, opportunities that felt most urgent back then?
- [03:17]
D
At the time, what we wanted to highlight, that you need this structure to really translate your business strategy into a systematic approach. That it's understood that it's not only driven by the technology and advancement of the technology, this is also why the structure still holds, because it's timeless. And at the time you could see if you talk about an AI strategy, and I think while companies were very used to do business strategies, it was not clear that you also need to do that for AI and data. And when we now look at companies of five years later, you see some of the companies, they have done almost every year an update of their strategy. So as you have your finance sector, your marketing sector. So now AI and data science is an incremental part of the DNA of a company. And I think five years ago this was still something new and not that obvious. Now of course, many companies went that route. But also, as just said, I think it's still important to see that this is not a one off exercise, it's an ongoing exercise.
- [04:24]
A
I wanted to dive a little bit into this new article that's coming up for you all in the Harvard Data Science Review for this January issue. One thing that we have seen is that, you know, hindsight is 2020 and when people look back and they go, oh well, we knew this or we didn't know this, you know, looking back over the last five years, since you all wrote your article in 2020, but back to, you know, 10 years ago when you all started your company, what assumptions have you all gotten right and what have you gotten wrong?
- [04:54]
D
This is a very good question and of course always good to reflect. I think, of course in some areas we anticipated that companies will, will accelerate faster in that sense, kind of more taking advantage of it. And of course you could say maybe learning and maybe underestimating how hard it is for companies to change the way how they do things. We both are more on the optimistic side, so we always see more opportunities. But of course that might not be always the case for companies. For us, it was clear that the future, and that's also you could see a driver of the article now is you will not be having a sustainable business if you're not leveraging the potential of data and AI. The most successful companies will have a hard time to tell you how many data or AI people they have in their organization because it's an obvious skill across the whole team. And so on. And this is something for them is a prerequisite.
- [05:51]
C
Yeah, the world is going to the direction where the lines between say business and technology are blurring. So if it used to be that business was the brain and the IT people were the legs that were just implementing, that's changing now because it's so much easier also with Vibe coding and with all the apps for business people, if they have a vision, if they know where they're going to use the tools and express themselves and the vision. And then of course it needs to be scaled. It isn't enterprise ready if we here are doing our own work. But you can get done so much more than before, which will be a fundamental change in companies like still 2020. We thought more that for example data scientists, that it makes sense to have a central unit where you're to some extent outsourcing these skills. But now I think that these lines are getting blurry and blurrier. So we cannot separate technology, data, AI from business anymore. So it's evolving everywhere. And that's what we're saying with the agents and the agentic AI, that we will all be really working side by side with AI agents in our everyday work.
- [07:15]
B
How do you define agentic AI in practical terms? Because obviously there are many takes on that. And fundamentally why do you see that as using Dirk, your term its transformative shift rather than just another incremental advancing.
- [07:31]
D
Automation for for example, from a definition, this is of course I think an interesting one. And maybe when we talk about agentic AI, first of all, I think we refer to workflows and processes and so on. So it's not a single function and task, but it's the combination of agents. And I think for example, in the discussion with leadership teams is think about kind of what makes a good co worker for you. A good coworker is somebody with a lot of experience. Experience refers to memory. That means you have the access to a lot of information, learnings from the past, which you can leverage. Same time experience requires also the skills. That means skills comes from your education, your ability to, for example, do calculations, summarizing things and so on. Third element, and that is especially from an agentic perspective, is that for you as a colleague, it's clear what you are expected to provide as an input and it's clear what you also expected to deliver to others. So you are embedded into the workflow. And then the last element is you could say in technical terms, we talk about guardrails. As a coworker you would talk about there are Certain rules, how things are done, there are certain rules, how you behave. So agentic AI means it's embedded in the operating system of companies, so there will be agents and engaging with humans and vice versa. And I think that that also means we need to learn, and that's the upskilling part on the human side, how do we interact with such solutions? How do we create the context needed, how do we work in collaboration? At the same time, of course, it also provides very strong requirements on how you need to design the agents in such an environment. So how you need to provide the information that we as human can take the right decisions when decisions are needed or expected from us along the process.
- [09:30]
A
You know, when I teach some of my classes or I talk to some of the professionals that I teach, you know, I say, you know, how many of you all use AI or whose company you know? Someone from up top said you need to use AI? Pretty much all of them raise their hands. And then I say, how many of you guys is it useful for? And it's rare to have one hand go up because, you know, it's just, it's so hard for companies to really implement this. So where is it that you see agenda get AI making sort of a real tangible difference in organizations where everybody really understands and believes in its adoption, whether it's through your own work or sort of more broadly. And if you could also address what this issue is, where expectations are sort of outpacing what's actually happening on the ground.
- [10:17]
D
Yeah, I think we very much one of the most important things, and maybe it's partially reason for excitement, but also then the disappointment is that at the end it is about that you clear what, what makes the difference from an outcome. What I mean with that, for example, company got very excited initially about I can generate now many, many copy texts for a campaign. I can now create hundreds social posts within minutes so that the effort is very low. So that's initially very exciting. But then the disappointment comes because actually it's not about if you can now generate hundred posts, your aim is to convert, for example, prospects, convince people about your products and so on. So that means you need to be very clear about the outcome. So this is also the starting point before going into what is the solution you want to use is are you clear what counts for you? And these tools are now very much fascinating because they lure you in trying it out. You get the result first, result looks very promising. So that's why everybody has tried it out. And accessibility was never that low as today. But After a while it wears out a bit because then people realize actually nothing has changed in my daily business. I still do at the end the same things as before. And therefore it is so important that you really go one step back, think about the outcome. And that's also what we, what we emphasize and highlight in the in the course is that it's not output, it's outcome. So you really need to think of what makes you and your business or you as a person, what makes you successful, what is what you want to reach, what's the bottom line at the end and then think about how you get there. And this is the nice opportunity that now those tools, they provide new ways of doing things. I think coding is a perfect example that where before it has taken weeks to get the first version of a product or kind of a landing page. Now you can do that in minutes or in hours. But it only will be impactful if you know why you're doing a landing page. Otherwise you will have a lot of landing pages but nothing will change. And then you go back to this frustration. I think that's also why I'm always get excited when jolie when you talk about this mindware because for me the current phase is the most intellectually the most exciting phase in my, in my life in that sense because it forced you really to rethink how you do things. I would say it's very demanding in some because we are so used to do things as we have done over the years and of course has also been proven in our career that we have done certain things right. So very demanding to figuring out okay, where do I focus on what really makes it relevant for me.
- [12:59]
B
Well, I certainly share that sentiment. I, you know, have done many things over the years. I do find I'm expanding my own mindware and I particularly want to talk a little bit about this course we have been offering together for HDSR itself is really something I certainly would not have anticipated even a year ago that we'll be teaching a course on Agentic AI. Of course a year ago we don't probably even know the term that much yet. I'm looking at the title from the Forbes magazine, right? Had this title called Three Courses to Master AI Agent and Boost your Salary in 2026. And our course is listed as number one. The title is Agentic AI Contextualized and Applied. Now the last thing we want to do is any hype as we all understand. So there got me something real here. So can you share what was your design principle? How do you make sure the course is accessible, but without oversimplifying, such a fast moving and frankly very technical topic.
- [14:04]
C
I think it's a combination. The course teaches in just two and a half weeks, as you say, some good frameworks, like this agent framework, as we call it. So you're immediately from the beginning starting to think about your own use cases, your own workflows, what you could do better. So it isn't only that you first listen to a lot of lectures and then you start doing something yourself, but you start the journey right from the beginning. And the course itself uses AI. It's very AI based. So there is a learning platform called Passkey which was or is developed by ngl, which personalizes the learning experience for every participant. So it helps you, it guides you along your way when you're doing your exercises, when you, your workflows, it asks just the right questions. You can discuss with it like a friend. So it's a very, very different and new and effective learning experience. And then I believe these live lectures that we're having, so which with all of us and other faculty are then adding interest to the topics. And there are some case studies, participants can ask questions and the community aspect is also important. So are people very high profile, busy, sea level people from large enterprises. And they have a chance to interact with each other and exchange ideas. You have very strong peer support in that course. And in fact many have expressed a wish that they can continue, which is also now becoming available after the course, continue this interaction with both the course, the tool, the personalized tool, Passkey as well as with each other.
- [16:04]
A
I think my, my biggest question and what I know that so many educators are struggling with right now is how to keep their courses current because things are changing so quickly. It's, you know, I almost sometimes feel like I'm learning right along with my students with how quick things are happening. You know, I remember when ChatGPT came out, you know, I was, I was the weekend before my classes, I was trying to figure out how everything was working and moving so that I could teach it on Monday. How do you all keep up with that? How do you keep everything that you're doing current? What's your sort of, what's your mode of doing that?
- [16:43]
D
Yeah, you go to sleep and in the morning there's another solution in the market. But a bit along the line, what Ola already said and also what we highlight in the course, that you look a bit beyond what is the technology? So not about the features, but what is the function? And so on for example, what it makes and what is required to use it in the best way. And that's a bit with the, with the framework like agent and the systematic. So I would, I would say one thing is having a clear systematic how you, how you structure, you know, the problem, you want to solve this how you structure to identify what is relevant. Because having a structure in mind, it helps you put things into this is something relevant and it should change what you should learn and what you should know. Or is it something which, which is in the same bucket than maybe five other news you have heard before? Like feeling, feeling paranoid on the one hand side that something new is changing and you need to catch up at the same time also being a bit, you know, Zen feeling like hey, yes, this has changed and so on. But bottom line, we still talk about the same. And I think more than, more than ever it's so important to have academia and then the applied, for example, applied practice is what we do together because that also helps you to keep identifying the right signals in all the noise because we have so much noise in what we hear every day. So being able to use academia and the structures, the methods behind to filter out the right signals and filter out what really is relevant. So that's also where I see more than ever it's so important to bring the different disciplines together and have this exchange and reflection.
- [18:25]
B
So far we have been talking about agentic AI as this kind of human empowering tools.
- [18:32]
D
Right.
- [18:32]
B
But as we said, we want to make sure that there's no hype here. There's obvious concerns of using agentic AIs. One of the things you will hear people talk about is, well, is there a real risk that we're actually designing system that quietly shift the decision making authority away from people? And how does that affect our own human's decision making process? We understand the mechanics, we understand the architecture, but we still don't quite fully understand, even for those of doing data science is how does it become so powerful sometimes just hallucinating for no reason. Right. Do you see any of the dangers of those things? And where do you draw a line from really a practical point perspective, as two of you have been advising lots of companies. Right. From practical perspective, how can we being powered by these tools, but not kind of enslaved, so to speak, by them? Right. And maintain our humans autonomy or decision making or thinking? Right?
- [19:40]
C
Yeah, I guess in the first place.
- [19:43]
D
Always.
- [19:46]
C
Ensure that the tools don't take decisions autonomously. I mean they're able to make recommendations and they're Able to reason and so forth, that's of course their power. But in the end if we think of like critical areas such as healthcare or finances and so forth, then to ensure that in the end somebody is checking and then you have those guardrails in place. To the extent it's possible. You have an AI governance model around it which you have defined where the policies and the regulations, the level of autonomy that you allow for the agents that may change from a company to company and also the use case. So if you do marketing, marketing isn't as critical if it goes wrong, if it isn't fully targeted versus decision about patients health. So you do need to think critically. And also like in some use cases that we have done, even if they are not super critical industries. But we have decided not to allow AI to do its own coding. So we have first used AI more as a rule base. So like an automation, okay, go do these things that we have defined. Because if we're not entirely sure that it isn't going to invent something on its own. So you're building also things stepwise and constantly testing and checking that you're still in control. But in the end it comes a bit back to also your questions of liberty about where it makes sense to use agents. And good ways to start in my opinion are cases where there is a lot of manual work and the automation would really bring efficiency, would make everybody happier, would give people people more availability to do their job better and focus on where people are needed and let in the way the machines do the machine's job.
- [21:55]
D
Maybe one thing to add is also is related to if you see it kind of agent AI or a. It's more like, you know, a bit like the calculator at school and so on. Ah, great, now I don't need to know math. So if you see that as an easy way out to get lazy, I think then you will run increasingly into problems because then you will lose even more the oversight and the control and so on. And I think more than ever kind of skills like system thinking, critical thinking will be fundamental. So actually you could say while maybe some of your muscles can loosen up a bit, but you need to strengthen other ones in your body that will be fundamental. I will see a software developer who then sees hey, this is great, I don't actually need to know any more coding and the basic practical it will be maybe boosting a bit in the beginning, but it will not be sustainable because you will be eaten up by the complexity and then you will not be able to orientate yourself and getting those things solved. And that goes also back to what I said before, is being clear what you want to achieve. And that's one of the core skills more than ever is know what you want, know what you need. And now the tools are there which help you to reach that in a better way. But in a way, I would say it's rather more intense than less intense than before.
- [23:15]
C
And then building those governance capabilities. So think about who is accountable and what risks are acceptable and who makes decisions. And then there's also all things related to data and security topics. And so it needs to be, in many ways, almost like governance by design. So when you do AI work, you're thinking about not as an afterthought. So how do I govern this, but actually from the beginning, how do I build AI governance into any of my solutions?
- [23:50]
A
So we're going to end on what we always do, which is our magic wand question. This is a little bit of a weird one, so it may take you a second. It took me a second. If you could wave your magic wand and if you could change one word in how people currently think or talk about agenic AI, what would that word be? And if you can't think of a word, I'm going to edit it a little bit to say it could be a sentence.
- [24:19]
C
I know immediately it's the word automation. People think of AI agents as simply automation, like rpa. So we just automate this, we automate that. And that's of course true. But it also falls a bit short because AI agents really, truly can work autonomously. They can make decisions on their own, they can reason, you can build entire teams, you can do many complex solutions with them. And that's something I would like to change.
- [24:55]
D
Yeah, I can second that one. And maybe my word would be, for me, it feels more. We talk about, you know, I would change the word artificial actually to augmented in that sense, because also I think what we highlighted before is the, the superpower comes if you combine, you know, our human intuition, our human contextuality, our human impreciseness, and combine it with the intelligent power of those models and algorithms. I, I still believe this is unbeatable because I think this together brings so much more dimension into the equation. The combination is super powerful. And I think, as ULA said, and then that's also why it's far beyond an automation. It's so much more.
- [25:37]
B
Well, thank you to both of you. I really can't agree more. What you just summarized, and I always tell people that at least the current artificial intelligence. There's nothing artificial whatsoever. They're all created by humans. They're trained on human data. And in the future we don't know. It's hard to predict the future, but so far I can tell it's really not artificial. But augmentation is great. Humans are always good at creating tools to do things we cannot do. The computer self is a shining example, can calculate things far faster than anybody can. If you want to learn more, please read the article by Ula and Dirk coming out of the next issue, which is January issue of hdsr and the pre print is already online with the title the Agent eccentric enterprise. Why 2 to 10 times productivity gains Demands Radical Workflow redesign.
- [26:40]
A
Thank you for listening to this month's episode of the Harvard Data Science Review podcast. Check out our show notes for links to Dirk Ala's HTSR Journal articles and the registration for their course agenic AI contextualized and applied. The next two and a half week session starts on February 17th. To stay updated with all things HDSR, you can visit our website at HDSR, MITPress, MIT.edu or follow us on Twitter and Instagramdsr. A very special thanks to our executive producer Rebecca McLeod and producers Tina Toby Mack and Aaron Keeswetter. If you liked this episode, please leave us a review on Spotify, Apple or wherever you get your podcasts. This has been the Harvard Data Science Review. Everything Data Science and Data Science for Everyone.