← Intelligence Squared
Loading summary
Transcript48 lines
- [00:01]
A
I think most enterprises aren't pushing AI enough and they are happy with too limited of gains from AI models. And they could actually be going way bigger with AI. They could actually be pushing the model to do far more for them.
- [00:16]
B
Hello and welcome to Intelligence Squared, where great minds meet. This episode is sponsored by Box. I'm Kamal Ahmed, journalist and author and today we with taking you into the world of AI agents. The reality of AI is both complex and fascinating. Across the uk, AI agents are quietly transforming how businesses work. But the real question is how do you turn promise into measurable impact? And what does the next wave of intelligent autonomous AI agents mean for the way we live and work? Now, we've heard all the fear mongering, the bold claims and a lot from many fortune tellers, but how can we possibly separate fact from sci fi, particularly when it comes to our businesses, jobs and products? Well, to help me figure all that out and more, I'm joined by Aaron Levy. Aaron is the co founder and chief Executive officer of Box, the leading intelligent content management platform that he founded in 2005. Aaron has been instrumental in the company's growth, seeing Box evolve into a leader in content and AI that helps its customers across industries reimagine how work gets done in their organizations with AI. So he's the perfect person to help us answer some of those vital questions that I'm sure many of us have about AI in the workplace. Welcome Aaron.
- [01:43]
A
Thanks for for having me.
- [01:45]
B
Now, I'd like to kick us off right away with a bit of a rundown from you. Where do you feel the state of AI adoption is in the UK right now?
- [01:53]
A
I actually don't think UK is that dissimilar from, let's say the US which is where we're based. We have a large presence in the UK and throughout Europe, track a lot of the similar metrics in adoption. I think in general, if you look at most businesses or even the consumer space, AI is, you know, mostly at the, let's say the ChatGPT kind of point of use cases, which is we're mostly using AI to be able to ask questions of a sort of a general intelligence that both can answer from its underlying knowledge set as well as search the Internet and find information that's relevant. So we're kind of using these things as sort of information assistance where you want to look up a set of facts about, you know, a historical event or get healthcare advice quickly or be able to have something get quickly summarized. That tends to be the state of AI adoption right now. Generally speaking, really across most major regions, I think what we're starting to see in the earliest phases right now, a lot of it is positioned primarily around engineering use cases. But you're starting to see this in other areas of knowledge work is this idea of AI agents that will actually help us go and automate real tasks inside of a workflow. And those tasks are growing in size. Today we're able to complete tasks that maybe are five or ten times longer or larger than the tasks we were able to complete maybe a year ago. And that will just continue to grow more and more over time. And so if you look at maybe AI one or two years ago, you could use an AI system to maybe generate a few hundred lines of code in your coding workflow. And now we're having engineers that are often generating thousands of lines of code in a single prompt, and then the AI agent is going off and doing that work. So I'd say we're in the earliest stages of seeing what this idea of AI agents is going to look like and starting in engineering workflows. But we're going to start to see this expand into almost every domain of knowledge work. Whether you're a lawyer that needs to review an entire contract or generate one, whether you're in finance and you need to be able to analyze a large amount of financial information, if you're in marketing or sales, and you need to generate collateral or be able to figure out the exact right pitch to a client, we're going to have AI agents that actually go out and generate that information or execute those workflows for us. But I think we're still in the earliest innings. We are quite literally 1 or 2% of the way through the full transformation that we expect to see.
- [04:11]
B
Aaron, you'll know that a lot of business leaders who aren't expert in technology, although they'll have experts in technology in their business, maybe won't really understand or won't understand the technical details. And that can lead to a lot of myths swirling in the environment, lots of people really slightly misunderstanding. Can you give us a straightforward description and explanation of what an AI agent is and how that differs from other types of AI that businesses may already be using?
- [04:44]
A
Yeah. So for the most part, an AI agent is really very similar to the current AI implementations that most companies have, with usually one added capability, which is that they can continue to execute on tasks on an ongoing basis beyond just the single prompt or the single pass through the AI model that today we tend to get from AI systems. So right now, for the vast majority of AI adoption in most organizations and in our personal lives, we ask one question, that question gets passed to the AI model. The AI model then either generates a response from what the underlying information in its model, or it maybe will do again a search across the Internet and then respond back to us. That's effectively a single pass through the model to respond. An AI agent is really this idea of what if that model could actually loop through itself over and over again to be able to complete an entire task. So as an example, if you were to say, hey, I want to be able to look at these 100 due diligence documents, financial documents about a client, and I want to be able to generate a due diligence report about them, well, there's no model where you could just give all those documents to that AI model and get a response back, because it would run out of space, it would start to hallucinate information, it would sort of lose track of what it was doing, and then you get a really bad response. So what you really need is an agent to be able to say, I'm going to go read each individual document. I'm going to pull the most relevant pieces of information. I'm going to likely keep track of that information in some kind of stored memory, that is this sort of temporary memory for the agent. And then I'm going to generate a full report that might take 5, 10, 20 minutes or even longer. And you need the model to be able to again loop through multiple steps in its workflow to be able to do all of the work necessary to complete that task. That really is then the change from just asking an AI model a question to really having an agent or some kind of agentic workflow that you're executing. And the vast majority of AI is still in that former category, not yet in that latter category. But I would expect over the coming years, 1, 2, 3, 5 years, that the vast majority of AI usage in the enterprise looks a lot more like the latter use case, which is you send a task to an agent, it goes off. That could be for a minute. We even have some examples of right now in online of people having agents that go run for tens of hours and then come back with some kind of completed result. And that will be the real jump in productivity that we expect to see.
- [07:07]
B
If you're thinking about business leadership and people using AI agents in their work, you've already spoken about the legal profession. That's one great example where you can see talk through the difference between structured and unstructured data and how business leaders can think about their unstructured data, searching those documents, looking for the trends, working out workflows through those that AI agents can really help them with. Because that seems to me to be one of the big challenges for business leaders. And in. In sort of department in many organizations is working out what they've got and what lessons they should be learning from the unstructured data in their business.
- [07:47]
A
Yeah, so if we just break out the two, generally the two types of data that you have in an enterprise, you have structured data. This is the stuff that goes into a database, your CRM information, your HR information, your ERP data, accounting information. That's all structured. And we've always been able to query that data, ask it questions, automate workflows around it. That's really been the history of modern computing, is being able to use computers to understand and calculate things that are structured inside of a database. Conversely, though, about 90% of the data in an enterprise is unstructured. So what's unstructured data? Well, it's everything that you would ever create as a user that would turn into a file, let's say. So it's a marketing asset, a research file, a contract, a financial document, a lease agreement, memos that you need to be able to send out, meeting notes. All of that is unstructured. And the key distinction is that unstructured means that the computer doesn't have in advance any knowledge of what might be inside of this type of data. It could be images, it could be videos, it could be long forms of text, it could be paragraphs, it could be pages. And so it's unstructured. Inherently, it's very messy. And we tend to only be able to get value from that information. When a human is sort of on a computer, looking at it, reading it, watching it, listening to it. That's been the history of how we get value from unstructured data. So we've got this paradox, which is it's 90% of our information, and yet it's only useful when humans can go and actually look at it. So it's more, vastly more data than our structured data set, and yet we actually get so much less out of it, because you're constrained by human time being able to understand that information. And so this is really the breakthrough of AI, because AI, the first, this current wave of AI, is built on this idea of large language models, which means that they're really, really good at understanding text and words and now multimodal experiences, which means images and videos and audio and so with large language models, we can now start to really understand the full value of all of this unstructured data that we're sitting on. So imagine being able to take the million contracts you have in your business if you're a large enterprise and being able to say something like, please generate a report about all of the contracts that are expiring in the next 30 days with this certain clause that has this certain risk level. And because you can now pull out all of the structure from that unstructured data using AI agents, you could instantly now be able to run that query and ask a question and get insights back for the first time from that large data set. Or imagine you have 10,000 research papers on a particular medical topic and you want to be able to generate a report based on the insights from all of those research papers because you're doing a new medical discovery that you're working on. Well, imagine now AI agents go and quite literally read through all of that data to find the most salient information that needs to go into that ultimate report. These are tasks you could never give computers before, but now is possible because of AI and AI agents. So with Box, for instance, we help customers and enterprises be able to manage and secure, store and govern their most important unstructured data. And now we're letting customers take AI agents and allow you to go work with all of that data at scale, using the power of AI. So the use cases are quite endless. You could do automatic client onboarding, automatic contract lifecycle management, digital asset management, where an AI agent understands all of the digital assets in a particular workflow so you can find exactly what you're looking for instantly. The ability to generate reports on any of your data or go do deep research on your information. So this is really the moment that we're in. We're finally, for the first time ever, we can flip this problem on its head, which is the more data that you have, the harder it is to find things. We can now actually deploy AI and say the more data you have, the more value you can now generate in your enterprise.
- [11:31]
B
Aaron, you used the word govern there, which I think is a really interesting one. Obviously for lots of these sectors that you are speaking about. Let's take law, it could be healthcare, it could be many different sectors. Where are we with governance and bringing in AI products into your organization? How much do you have to understand the regulatory environment that you are working in? And is one of the issues for many businesses is how do I know what I can trust? How do I know that the Guide rails will be there. And how do I know that? What I'm using is aligning with regulatory positions, which are often very fluid country to country.
- [12:12]
A
Yeah. So interestingly, most companies don't have an AI problem. They usually have a data problem. And the way to think about it is before AI, if you, if, let's just say AI never happened and you went to a company, you said, tell me where are all your contracts managed for all of your clients? The average company would say, well, Here are the 10 systems that a contract goes through. Sometimes it's in the storage repository, sometimes it's in this document management system, sometimes it's in the CRM system, sometimes it's in this E signature provider. And that would be how most companies manage their contracts. Now imagine the same thing for your digital assets or your research data or your invoices or your financial documents. You can go category by category. And most enterprises never really had to think about, well, I need one source of truth for this data. The reason was it was a sort of a low grade problem for most employees to say, okay, I think I was working on that document inside of that document management system or that contract system or that esignature tool. And they can generally keep track of that information. They have enough kind of context, they can kind of remember where they were doing their work. It was okay. It didn't lead to that many kind of existential challenges. It was painful for the user, but it was not like a company ending event. Let's think about now a world of AI agents. Let's pretend that we've got a hundred times more AI agents in the enterprise than we have people, because that's sort of the trajectory that we're on. The AI agent doesn't have the same context that a person does because a person has so much more external signal about what's going on in the business and what they did last and what a person just mentioned to them in the hallway about where to find something. The AI agent doesn't have any of that signal. So if you tell an AI agent, go find me the last contract that I sign with this client, they're going to go through all of those systems. They're going to do their best to figure out like what maybe they should be finding. It'll eventually find all the data. But the problem with finding all the data is they'll often then find the wrong data too. And they'll find the document that isn't actually the authoritative source of truth document for that decision that you just made or that contract that you just signed or that invoice that you have to be able to go send. And they'll either send the wrong information or answer the wrong question. Or even worse, if you don't have complete synchronicity between all of the security across all your systems, they'll inevitably answer a question for a user, revealing information that users shouldn't have access to, because maybe one access control was out of date in one of those systems. So think about how big of a problem this is. You send 100 times more AI agents than people in an organization to go roam around all of your data and find exactly what it's looking for and answer any question that the user has. Inevitably you're going to get security risk. You're going to get the wrong answer from information. It's not going to know the authoritative source of truth of data. So this is actually a data problem, it's not an AI problem. Interestingly, there's no amount of AI that you could deploy that will fix that problem. You have to start at the data layer. And so companies need to start to think about how do they govern their information and their unstructured data in a way that ensures that AI agents can actually be deployed in their enterprise to make the most use of that data and get the most value from it. So I would think about how do you have an architecture where you've got clear sources of truth for certain kinds of data sets, you've got very clear boundaries of information for access controls. You ensure that you're not mixing too many different, again, permissions, layers and access controls together. The more systems that you have to manage, the harder that will get. And so we generally recommend to customers that they think about these core systems of record per data type. So I know that all my CRM data is going to come from here. I know all my HR data is going to come from there. I know all my unstructured data and contracts and marketing assets are going to come from there. That will help you really have a cleaner way of having this clear demarcation point between systems. How do you manage the access controls effectively and at least get you to a better starting point with an AI strategy?
- [16:00]
B
Aaron, that's absolutely fascinating. I'm hearing that you've got to do quite a lot of preparatory work to be ready to get the real value out of AI agents? Am I thinking about it in the right way? It's almost the preparatory work. Your data issue, which I think is really important for so many companies, is there data in the right place. Do I have silos where things are just hidden from other parts of the business where they shouldn't be. So that's a good process to go through then I imagine. Do you then go to experimentation and then to measurable impact? I just wondered, Aaron, if you could talk us through how a company could move through the stages where it's getting itself to the point you make, which is that in the maybe not too distant future we will have hundreds more of AI agents working with us in our businesses. What's the process to get us to that, those kind of big staging posts?
- [16:53]
A
I think in general I tend to lean toward experimentation first, try many different things across an organization to figure out what's working, what's most effective and then figure out how to scale the things that are working. So I do take a lean oriented approach where you're building a flywheel of success that you can then find successes, scale those, kill the ones, kill off the projects that aren't working well. And then equally though, there are some areas where you have to say, you know what, this other company over here has this new interesting best practice that they're doing. We're just going to go big because we can now learn from the tests and the trials and tribulations that other companies have done. So often I'll find myself pushing teams to try lots of stuff as well as, hey, I saw this startup that really has re engineered this workflow. What if we go big and just reinvent how we're doing a certain practice, leveraging that new methodology? I think one, one tendency that I tend to see really kind of across the space generally is I think most enterprises aren't pushing AI enough. They're sort of happy with too limited of gains from AI models and they could actually be going way bigger with AI. They could actually be pushing the model to do far more for them. And this is I think a trend generally that those of us in AI will see within the enterprise is everybody sort of thinks that AI is only capable of, let's say, you know, X units, when actually AI is now probably capable of 5x units. And so you have to be both really tracking all of this so you stay very current. But you have to be up to speed with the latest new capabilities, new trends, new techniques to be able to push this technology as, as far as possible.
- [18:30]
B
Can you give us an example, maybe not a specific business, but an enterprise you have worked with, where they have gone from that experimental stage to the measurable impact stage.
- [18:42]
A
So we just launched this new capability called Box extract. And box extract. What it does is it takes AI agents that can now extract the structured data from unstructured documents and data. So you give it a contract, the AI agent reads the contract and it pulls out the critical data fields from that contract, like the party names and the counterparty and the amount and the due date and the address, all of this messy unstructured data inside of a contract that lets you have now valuable insight and valuable information from that contract. And so we're seeing a lot of customers that will start by saying, oh, I want to play with this data extraction. And maybe they'll individually prompt the document and pull out the right data fields. And then very quickly they realize, wait a second, I have these workflows in my enterprise where we're processing hundreds or thousands or tens of thousands of documents. Could be insurance claims, could be medical records, could be financial documents, could be contracts where they say we're spending an incredible amount of time wasted just reviewing these documents by hand. What if we could go and automate the full information extraction workflow from all of this data at scale? And so we'll have a lot of customers that again, experiment, they play around with the tool and then they very quickly scale it up to in some cases, tens of millions of documents that they want to be able to go and process, which is really a way of re engineering their entire workflow. Because now you're no longer bottlenecked by just again, human time and our labor on that, whatever that workflow is, we can free up that time to go do much more strategic, much more interesting and important work within that workflow.
- [20:16]
B
And Aaron, how do you understand the return on investment? You talk about time saving. How does it work to create a stronger bottom line in the end, more profitable revenue? That could be obviously one of the key KPIs. How do you help your partners and who you' with really understand ROI of whatever description?
- [20:38]
A
Yeah, so I think companies are going to take different forms of what their KPIs are. I talk to as many companies that actually want to generate more revenue or be able to drive more output as those that want to save money or do something just more efficiently for the sake of efficiency. And that's, I think the really powerful thing about AI is each company is going to kind of determine where they want to be on that slider of actually I want to use AI as a competitive weapon against my competition. And my only measurement is the rate at which I win more deals versus my competition or the rate at which I can drive the revenue cycle faster versus versus what I was doing before. And I would argue that you should really push yourself to find as many of those situations as possible instead of just, okay, I was able to save 2% cost on this particular workflow. Now there's going to be plenty of that as time goes on and as companies scale up AI. But in general, I would bias toward pushing your company to think about doing more better, serving customers, shipping more product, making the business be able to deliver better for again employees, customers at scale and the way that you will measure that will just be different by industry. At box. We think about it as can we improve our win rate? Can we reduce the cycle time of how long it takes to sell to our customers? Can we generate more leads and more demand for our service? Those are the primary things we're measuring, not necessarily can we save x percent of dollars? Now, there might be some areas where as we scale the company, we're going to shift our resource allocation because AI is allowing us to augment the current work that we're doing and get more done faster. And we want to go now put those freed up dollars back into the growth areas. But again, every company is going to be different in terms of how they determine what to invest in.
- [22:18]
B
You said that many businesses may be using what for some people are sort of that notion of first stage AI. An AI agent gives much more richness than that. Is there a very straightforward, what we might describe as a litmus test for a company, a firm, looking at some of its uses of AI at that rather basic level that you discussed at the beginning of this conversation and thinking, I really need to now move, what are you looking for in your business? Which is that sort of litmus test where you're, I need to get in touch with Boxer.
- [22:55]
A
Yeah, I think if you are, if you are, if you look around your organization and there's a lot of bottlenecks that are really about. The bottleneck is how long it takes a human to review information or generate information in some capacity. So review a contract or write a contract, review marketing assets and produce marketing assets, look through sales data and come up with an insight. If you have bottlenecks in your organization that relate to that, those are the areas that I think you will see the greatest amount of upside in the near and medium term with AI. And then you should, you should call us or you should call somebody who knows how to think about that problem, because that's where you're going to see the greatest amount of upside. If you have parts of Your organization where you're not really bottlenecked by information, it's just the human interaction with the outside world is too heavy of a dimension of the workflow, then you're gonna have limited upside by using AI, at least today. But right now, if you are spending a lot of time reviewing, creating, finding, working on any digital information task, that's where agents will come into play. And the amazing thing is this is gonna work across fields that we really wouldn't have thought about before. There's a recent study where doctors that use AI scribes. So this is sort of listening to the patient experience and then writing down all of the notes from it. There was something like a 30% improvement or plus or minus on the burnout rate of doctors when they use an AI scribe. Because guess what, lo and behold, if you're a doctor and you're spending your day meeting with patients, the last thing you want to then go do is spend hours and hours writing up all the notes. Putting it into a system like that is not strategic value, that's effectively homework. You're not improving the patient experience anymore because of that step in the workflow. And so what's amazing is that's largely traditionally an analog task that gets converted into a digital task. And now AI is like very good at doing that because it can listen to the conversation to get all the right data and then we can just pop that into an EHR or medical system. So those are the kind of use cases that we can bring automation to, that free up everybody's time. It improves the quality of the delivery because now you have a less burned out doctor and improves their job in general, which is great for everybody. Those are the kind of opportunities I would be looking across the organization for, which is how do we do more, how do we better serve clients or patients and then and ultimately deliver more for the market.
- [25:20]
B
What are some of the hurdles or just some of the caution that you're going to be hearing from business leaders, business organizations around how much can I use this? As you say, people often think 1x2x when it's actually 5x10x in terms of what it can actually do to help you. You've touched on the silos issue, which I think is clearly vital. You've touched on the sort of data hygiene issue, which I think is a big issue for many companies anyway. And actually moving into an AI agent situation would help you clean up those types of systems, which I'm sure many companies know they need to do. What else are you hearing or are Those the types of things that you're hearing?
- [25:59]
A
Yeah. So I would say all that. I'd add two more. One is the rate of change in the AI model space can be overwhelming for most companies because you have a new breakthrough AI model almost every single month at this point that exceeds and surpasses the last breakthrough model. And so companies are inundated with, well, should I go with Anthropic or Gemini or OpenAI or Llama? What, what do I do? That can be very overwhelming, which leads you to probably wanting an architecture that lets you take advantage of any of those breakthroughs. So that's a kind of quick architectural benefit of when you use an, a kind of a platform layer to manage your information and bring agents to that information. That's just sort of one side point, I think. Another one is this general question. About two years ago, we were talking a lot about hallucinations. And while hallucinations are being reduced quite effectively, you still have times where an AI model will, will answer a question wrong or pull the wrong information into the data or the workflow, but not necessarily any differently than a human would just making its own judgment, decisions about what to do, but it might be different than what you would have wanted. And so there's a couple elements to this. Some are the tractable elements, which are like, well, what can you do to reduce that as much as possible? And then the other parts are more psychological in our workflows and in business setting of what do we do about this? So the tractable part is that you're going to hear a lot about this concept of context engineering, which is how do you give the AI agent and the AI model the right context to basically best execute your task? And that could be like a very long prompt, a page worth of detailed specification of exactly what you're looking for that will dramatically improve the rate at which the AI model is able to go in and find the right information. And so writing long or at least detailed and clear specs and prompts will help that quite considerably and then making sure that you're giving the AI agent the right data to be able to make the decision. So those, those will help quite a bit for being able to get the right information back. The other category is sort of this idea that a human might make the same kind of mistake as this AI agent of either answering the question slightly differently than you would or pulling in the wrong information. And this is something which is much more about expectation setting, about really AI agents. We are probably, for the foreseeable future, not going to live in sort of you, you kicked off perfectly of this idea of this sort of sci fi utopian kind of state. We're not going to live in that state. It is not going to be a world where you tell an AI agent to do something and it just does it perfectly and you're just like so happy that it did it perfectly. We have to treat AI much more like you would treat a relatively new employee that has relatively new information about the organization they work in. And, and that means you have to be both very clear about what you want it to do. But you also know that you're going to have to review its work and you're going to have to go through and check the ultimate results. You are the manager of an AI agent in the exact same way that, that previously you may have been a manager of new people coming into an organization. So if you go tell an AI agent, please take these 20 documents and make an SEC filing. SEC in the US for public companies make an SEC filing with that data. It's going to get five things wrong. I don't know what they're, I don't know what those are going to be yet, but there'll be five things wrong with the filing that it creates. And so you will have to go through with a pen and look through to fix the mistakes that it made. But guess what? We did that with people. You're going to have to do that with AI agents. But what you saved was the 10 hours or the 95% of the time that it would have taken to go and collate all the information, review it all, pull it together, write a new document and do that from scratch. So AI is not going to get you the 100% workflow. It's going to get you the 90 to 95% workflow. Humans will still have to review it, you're going to have to orchestrate it, you're going to have to edit it. And that is the, that's what the future of work is going to look like for the foreseeable future. There's no panacea here, there's no magic pixie ducks. But the benefit is that you get a huge burst in productivity because we're doing a lot less of the drudgery work in that process.
- [30:05]
B
I think Aaron, you touch on a really interesting almost psychological point for human beings. You know that the first time a self driving car killed somebody, it was a much bigger story than all the people who die on human caused traffic accidents every day on the roads of America. And Every road around the world.
- [30:24]
A
Right.
- [30:24]
B
I think one of the big hurdles is this idea of if it gets one thing wrong, how can I trust anything it has done?
- [30:31]
A
Yep.
- [30:32]
B
I think you've spoken about or it's been broken by this idea of human in the loop. I just wonder how you encourage people through that process to continue experimenting and exploring.
- [30:43]
A
Yeah, well, this is where actually it's really important to have a general understanding of how AI works, where its pitfalls are, where it's not. AI literacy is going to be the new skill for the 21st century. I remember growing up in middle school and high school when Google was sort of taking off and we had librarians that would teach us how to Google and, like, how to effectively use the Google kind of way of doing search and how PageRank worked and how we should trust sources or not trust sources. And we just need the new version of that for AI. So everybody has their own, like, internal kind of mental model for how this works. But for instance, for me, if I ask AI a question and I'm tuned to sort of say, okay, when it comes back with something that is like, purely a fact, like a number, a date, a location, I'm going to have my spidey sense up a little bit higher to sort of say, okay, does that sound right? Does that check? I might sort of see, are there any sources for what it just answered? And then I'll very quickly understand generally whether I think that the source of information is correct. I already got 90% of the value even before doing that check, because there's probably a lot of other kind of pros in that answer. That was very useful. But maybe I need to make sure that that thing that happened in 1923 was actually 1923, and it was in 1926. But the general concept is there. I'm already off to my. Off to the race far faster than I would have been if I had sort of started that workflow from scratch. And that's more of a kind of a personal example. But the same thing is true in business. If you're asking, what's the best message to share with a customer in this industry, it's not going to hallucinate that because it's a lot of qualitative information that it can pull from different data sources. But if you said, please generate a financial document about all of our financial reports from the past three years, you're probably going to want to verify each of the. Each of the numbers that it comes with. For now, at this point in where AI is In two years from now my answer might be different, but we are certainly working to eradicate that last sort of final mile to your self driving car. Point of those mistakes.
- [32:45]
B
What Aaron, are the most common reasons for AI supported projects? Use of AI agents for them stalling, for them being sort of not driven. You said something which I think is really key to this conversation. Often just feeling overwhelmed and people tend to, particularly in very large businesses, do tomorrow what they did yesterday, as long as yesterday wasn't an absolute nightmare. What are the main reasons that you sense from all the people that you work with for AI stalling and not being used in the most imaginative and useful way for all of those companies?
- [33:24]
A
I think choosing your partners correctly is very important because the way that these products are constructed, the user experiences, the how good the models are, how good they are, working with data, those all matter a ton for having effective AI strategies. So I think unfortunately a lot of failure modes are when companies either choose the wrong vendor or maybe try and do too much of the work themselves and they don't sort of have a good clear partner strategy. Even for us. You know, we're constantly thinking about what should we build ourselves versus work with OpenAI to leverage or work with anthropic to leverage because it's sort of just not worth the time and energy if somebody else has solved that. So everybody has a version of that in the economy. And I would be, I would try and be very thoughtful if you're an IT organization, what do you really need to build? What is what needs to be proprietary and value add from your organization versus the market is going to solve this and you can take advantage of it for a time to market advantage. And so I think companies that think through how can they mitigate the technical risk, how do they mitigate the use case risk that would already evaporate the vast majority of failures. And then again, figuring out where do you want to deliver that kind of 1x value because it's just useful enough and then where do you want to push the system so you blow everybody's mind of now this is really the power of AI so it's not sort of stuck in this kind of like mediocre purgatory where it's useful but it didn't really game change anything. So I only go back to it so often you want to make sure you're picking out the right kind of use cases that blow people's minds on.
- [34:53]
B
That front, let's look forward a little bit. I want to touch on this jobs issue. People will be simply concerned. I will introduce this powerful new partner, this powerful new technology, and actually all these people I employ will become. I won't need them anymore. And suddenly you're into this situation where you're going to have to completely resize your organization. Is that itself a hurdle to adoption that people just think is going to create unemployment?
- [35:20]
A
I think there's probably truth in that for sure. What is my motivation as an employee? To effectively partner with and bring in a technology that maybe I perceive as being competitive with my own role. And so I can see that as a kind of a logical, tacit form of resistance. There's always this tension, though, because equally I think you, you ask most knowledge workers what do you like and what do you not like about your job. The part that they don't like is the stuff that AI is very good at. The stuff of. I spend four hours a day in PowerPoint files, moving information around, collating data, reading something and copy and pasting it in, modifying the graphics. That's not that enjoyable. Maybe it is the fifth time I do it, but not the 7,000th time. And so, you know, I think there's a lot of jobs that we do where you're just going to see a reallocation of time from the stuff that, that again, only people could do in the past to now AI is able to do. And now we can change the definition of what are people only capable of doing that will provide, I think initially some kind of psychological crisis at first, because you'll be like, well, I can't believe that. That thing that I used to. Only a person used to be able to do of read that information, modify it, transform it, put it into another document, share it. AI is now able to do that. But then if you really step back and you say, what was the value creation, you know, part of that journey? It was probably going and meeting with the client and understanding what are their problems and then ultimately executing on whatever that thing was that you're trying to deliver for them. The value creation wasn't the creating the PowerPoint and responding to the email or reviewing the document. You do have to do that because you have to understand what is the goal that that customer has. At least previously you had to. But that wasn't really the value creation. The value creation was understanding the problem, being there with the client, working with them hand in hand to execute that. So I think what you're going to see as AI for the most part is solving the kinds of problems it's very good at solving the Kind of problems that people don't really want to do. And yes, it will on the surface feel like some of the pie graph of our time gets shrunk because maybe that was a third of our time before. I would argue that actually what will happen is that third that goes away due to AI doesn't mean that a third of the people in an organization go away. It means that we will now shift what we are spending time on in our organization because we'll have reduced these bottlenecks that are just everywhere in our enterprise and we will spend more time on the value creating parts of our job, which is building more products, getting our message out there to clients, innovating more, better serving them. And for that I don't see there as being some kind of finite limit on how much the world wants more innovation, better health care, new medical discoveries, better consumer products, better education. This is a, we have an insatiable appetite for all of those things. And so we will spend incrementally less time moving data and responding to questions and researching something inside of a large data set. We'll spend more time doing those value creation activities that now people are the only ones that can go and deliver on.
- [38:25]
B
Aaron, it's funny hearing you say you were sort of taught how to use Google. I remember inputting data into sheets, actually typing in the numbers. That was one of my first ever jobs was data inputting. Can you imagine still doing that job?
- [38:37]
A
Yes.
- [38:37]
B
Today would be absolutely ridiculous. So I take, I take your point, Aaron.
- [38:41]
A
Yeah, but, and that's the interesting thing, right? Like if you were to look at, let's say the advertising industry, you said to somebody in 1980, making up a time frame, you know, take this advertising workflow you do now imagine now I'm going to show you technology from 20 years in the future, Photoshop and the Internet, where you can deploy the advertisement and automate the workflow around getting that deployed. Do you think that there's going to be more or fewer graphics designers and people in advertising? And I think if you were to take and snapshot exactly that mindset in 1980, you would say there's fewer because wait a second, I just took this workflow that took me a week or two weeks to generate this poster or campaign or billboard for this client. And now it seems to take 20 minutes. I take a template file, I modify something, I draw something, it takes what used to take weeks and it turns that into 10 or 20 minutes. And lo and behold, over the past 20 or 30 years, obviously we've had Just a massive increase in graphic designers in marketing automation, in advertising dollars generally. Why is that? Well, guess what, we took something that used to be a very scarce resource and a scarce activity and we made it available to 10 to 100 times more sort of customers and participants in the economy, which has then just transformed the amount of graphic production and advertising and media that we generally create and consume. And so the way to think about AI agents is actually not that dissimilar to this idea of think about all of the things that we do today that are relatively scarce because they're expensive, they're slow. You can't really hire that many people to do that thing because it's just hard. Now AI agents come on the scene and it's not that we have then fewer of those people doing that, it's that we do way more of it across all of our businesses. And now the 10 person startup that never really was able to get expert legal advice gets now expert legal advice. And the 10 person startup that couldn't have afforded the best graphic designer is able to do amazing advertising. What's the ripple effect across the economy when now companies can do all of that? It's actually going to lead to the creation of new jobs, new companies, new firms that are the response to what does this new environment look like when companies can move faster, ship more, better advertise, better serve their customers? Which is why I'm not very myopic on what kind of what happens to jobs as a result of that.
- [41:02]
B
You've touched on quite a few sectors in our conversation. Healthcare, law, advertising, media, marketing, graphics, very broad range. Where do you think the real sectors of focus are? If you're a leader in three or four big sectors and you're not yet really using AI agents to their full capability, which sectors are those? If you're a leader in those sectors, are you going to be talking to?
- [41:30]
A
I have a hard time answering that simply because I think the technology is pervasively useful. So I treat that as the inverse, which is which sectors are the least interesting and I can't find them, they don't exist. Yeah, There is not a sector that cannot use AI right now to really transform the way that they're working. Again, it's going to start small and we don't expect overwhelming broad transformation in the next one or two years. These things take years and years of change management to actually deliver on these outcomes. But I think you could just go through the list. Any company that has a large amount of information that humans process that's where you're going to see the impact of AI and AI agents and we're just in the beginning of that journey and often it's people.
- [42:11]
B
I think, well, I'll start that, but I might start it in 2026 or 2027. What is that starting point? It needs to be actually probably quite a low lift, as you suggest. First time is that 1x fee. Just do something simple and straightforward because I think a lot of leaders thinking about their next quarter will be thinking, oh well, I'll just put that back a bit and I don't need to do that yet because it's all a bit overwhelming, as you've said. Aaron, what would you say to someone about what they should do next quarter, I. E. The next three months? And how straightforward would it be to take advantage of all the things that you've explained on this fantastic. During this fantastic conversation.
- [42:50]
A
One of the benefits of moving relatively early, not like total first mover, but at this stage is you're going to get a great kind of feedback loop of lessons of what works, what doesn't. You'll build a sort of a better understanding of where you should invest your time and energy. So that's why starting relatively small, of trying out experiments, having teams kind of show off what's possible and then figuring out how to scale those. That's sort of why I always recommend companies just dip their toe in and get going, because you're going to get a really good feedback loop that sort of starts to snowball over time. This is the moment to get started and I would not wait two years to dive in. Doesn't mean that you need to transform the entire business today, but it does mean that you should start experimenting, start to develop a wherewithal of what this environment looks like because that will pay off massively over time.
- [43:39]
B
Aaron, last couple of questions. How far ahead can you look and is there something that we should already be considering? Because if you imagine this conversation even two years ago would have been completely different three years ago. You've obviously been in the technology innovation industry since you were a student, Aaron, so you know your stuff. If you look forward two to three years maybe, is that too far? Is there a big thing on the horizon that we should already be thinking about as business leaders that we need to start prepping for now, which are. You've given a very clear explanation about what you can do now, what's available now, Give us a bit of futurology.
- [44:22]
A
Yeah, unfortunately, as I answered some of these questions, I've compressed actually the Timeframe quite a bit. So the things that we're seeing right now at the very beginning are actually the things that will be more mature in that timeframe. So the general way I would just think about it is almost for any discrete task in an enterprise that deals with knowledge and information, we will be able to automate. That is again, very different than automating a job. A job is a collection of tasks. And what we're going to be able to do is automate individual tasks and still require people to coordinate all of that automation. Coordinate the work that's sort of happening from these AI agents. It changes by industry and by job function. But if your job is to generate marketing campaigns for 20 different markets you're going after, you'll be able to deploy that to a set of agents that can go and generate that set of marketing materials and assets. And you'll review them and spend time with your team, reviewing which ones you like and then figuring out which ones to deploy. Maybe you'll get an agent to help assess that. But in general, I think human preference and taste is going to become still be a very important factor. If you're in the legal function, you're going to have AI agents review all of the contracts coming in and then figure out how to redline them and which things do you want to go and modify. Or you'll have AI agents that are generating them, but then you're going to have a set of lawyers that go and review the decisions and the outputs from those agents to figure out what they think is sort of most correct about that. If you're building code, you're going to have engineers that deploy AI agents to write entire chunks or parts of a code base, and then they'll go and review the code and edit it and look for bugs and security issues and then coordinate that into the overall code base or whatever is being deployed. And so I think our more and more of our jobs will look like that, which is, again, not that different from when we moved, let's say, from the analog process to the digitized process. We used to have all of this work that we had to do by hand, and then we made it digital, and we started just typing in a keyboard and reproduced what we were doing digitally. Now we're doing the next phase of that, where we now have an AI agent that we deploy a task to and it does a bunch of work, and we go and review that work. And so I think this is sort of the trend of technology in general is we're always sort of Incrementally moving up the stack of what we are working on with higher levels of abstraction and we're able to use technology to do more of the kind of grunt work for us so we can do the things that are more interesting and more enjoyable in our processes.
- [46:48]
B
Okay, now imagine. Aaron, final question. You're an AI agent and I have asked you to. I'm the human in the loop and I've asked you to sum up this conversation. Three key takeaways for the audience in terms of how I should think about use of AI agents in my business. So, Aaron, I'm going to test you and see if you would come up with as good an answer. Oh God, an AI agent.
- [47:15]
A
I mean, unfortunately I don't think I would. Claude. 4.5 is very powerful at this. I would probably, I'll probably be a little bit biased to our world, but one, your AI strategy is your data strategy. Two, start small but get going right now. And three, push the limits of these AI models far past what you think is possible.
- [47:33]
B
I think if you were asking me the same question from the other side and I was the AI agent, I would agree with. I think those are three very, very clear ones. I think that idea of how you can de risk by starting small, and I think a really high impact point you've made to me is the silos of your data. You've probably already got a data problem you're fashioning of. You don't have an AI problem and an AI regulation problem or a fear of AI problem. You've got to make sure your data is being used in the right way and is being protected in the right way and then AI can help you. That's a real hygiene issue. I love that point.
- [48:09]
A
Exactly. Yep.
- [48:10]
B
And I think get on with it now. Experiment and then move towards what are the actual enterprise scale solutions that you can create which make your business run better. And I think the thing you said as well, which I loved, was that could be across all sorts of internal and external functions. It's not just about bottom line or profit, although it may be, but it's about work. It's about your internal culture. I love that as well.
- [48:39]
A
Awesome. Well, thanks for having me on.
- [48:41]
B
Aaron, great conversation. Thank you so much for joining us. I think the bottom line is winning with AI agents takes more than technology itself. It's about busting myths, breaking silos and building the skills to thrive as work reshapes. The companies that start now experimenting, learning and building the right foundations will be the ones future proofing. My thanks to Aaron for joining me in this conversation, and to you for listening. This podcast was brought to you by Intelligence Squared in collaboration with Box.