Episode Transcript
[00:00:00] Good morning, everybody. I want to talk about something that's becoming really, really big. And I don't think I completely understand it yet, but I have some perspectives because this is going to be a topic that you're going to be debating and hearing about and talking about a lot. And it's the idea of what is now called task intelligence.
[00:00:20] All of the major software companies have started to purchase and build tools to take a job title or a job and break it down into its tasks and then use those individual task items to decide what to automate, to look at the capacity of the company in various areas, to plan for future tasks and decompose work into tasks. I have been thinking about this for many years and I always called it activities, not tasks. But let's just forget that for a minute. And so Lightcast, this company, techworld, that was just acquired and all of the other labor market data providing companies have spent a lot of time drop who we were with this week, D R A U P Building intelligence systems, to take labor market data and break it into tasks. So, for example, if we know that there are 35 million waiters, and you know, waiter, by the way, is a job type in the Bureau of Labor Statistics, then we would know that for each of those waiters, there is a task to bus a table. So we could configure out how many table bussing tasks there are in the world. And theoretically, if we built a table bussing robot, we would know the size of that market. And if you're a restaurant, you could look at the table busing task as a company or as a restaurant individually, and you could say, well, you know, maybe we should automate that first because that's more important than the sweeping task because there's more table busing going on and that has a bigger impact on our revenue than the floor sweeping task. So you can see where this is going. Now, just to give you a sense of the data here, the best, most commonly used structure for this is the Bureau of Labor Statistics job categories of 835. It's about 832jobs. Lightcast has a similar one that they've developed called the Lightcast job architecture, but that's a little bit more proprietary. So let's just talk about the BLS one. So you take these 832 jobs jobs and you break them into tests. But unfortunately, what you find out when you talk to companies and this week I talked to McKesson about this, I talked to some banks about it, and some other companies are doing this Is that the tasks in the BLS or the job titles in the BLS are averages of many, many things. And by the way, how do those companies figure out what those tasks are? They read job descriptions and they look at publicly available data to aggregate together what the job requirements are for these tasks. So this is very, very, very similar to the hairball we've been working on for years of skills. It's the same idea. Tasks are an attribute of a job.
[00:03:09] So as an academic exercise, if you like doing mathematical things, you could take the 830 job titles and you could take the. Let's suppose there's 30 tasks per title, whatever that is, 24,000 or 240,000 tasks. And you could do all sorts of fancy charts and graphs on how many of these are automatable based on the maturity of AI, which a lot of academics have done. And you could build all sorts of tools for companies to say, here's how much of your workforce could be automated. And that's basically what's happening. So you can, you can buy these tools from all of the companies now. Workday now has one, SAP is planning on announcing one through the techwolf stuff. And you know, you're probably, it's probably going to be sort of an implicit tool from a lot of these vendors. And then you could run your company as a task capacity company. Forget about humans. Let's not, let's pretend there are no humans. Let's just look at all the tasks we have to do to run our company. And if we want to do more of a certain task, we'll dial it up, you know, with more robots or more agents. And then if it's, and you know, some of them will be done by human, and you could, you could sort of say, well, for that task, the human is 30% less efficient based on token consumption. So it's cheaper for us to use tokens than it is to use human energy for that task.
[00:04:24] You could do sorts of financial analysis because a token is a unit of work in a sense it's a unit of intellectual work, but it's a unit of work. And you know, you could sort of, it seems theoretically great, but let me warn you that we may be going down a very big dead end here. But before I tell you what the dead end is, let me just tell you about where this works. Where this works really, really well is in compliance related work, safety related work, regulated work, like accounting, where the actual TAs have to be written down. Because if people don't know how to do These tasks, they can't do the job, they're not qualified to do the job. If you're a nurse and there's a task to draw blood, you have to know how to do that task. Now I would call that an activity or I wouldn't call it a skill, but you could call it a skill. So there's a lot of things in work that are very task related and we need to qualify people for tasks and we can define them and they're easy to understand and we can manage them as activities, which I call them and you know, do work around the capacity planning of those activities. And I don't have any science to this, but I would estimate that that's at most 20% of the work that goes on in a company, in a manufacturing plant. I've talked about my little trip to Ferrari last week or two. There are a lot of tasks, I mean, you look at what these people are doing to assemble a car and I'm sure this is true for Tesla and Mercedes and Porsche and all the other companies.
[00:05:52] You could write a list of the things, screw this bolt into here, tighten this check available of this, make sure this isn't leaking and you could write all those down. And if you're a manufacturing process engineer, that would be helpful because you could better design the robots in those manufacturing plants to automate those tasks that are not yet automated. And so the argument makes perfect sense that in white collar work, if we wrote down the 29 steps that a salesperson does to close a contract, you would have all those tasks. And so, so I'm not arguing against these tools and I think in some cases they're going to be really, really useful. And we're doing a whole bunch of case studies on this. But the risk of this or the challenge, I maybe is the word is that the task is a means to an end. The task is not an end. It may be an end in some cases, but in other cases it's not. So and if I take the example of our company, we have dozens and dozens of tasks that are involved in producing, publishing, marketing, distributing and providing services around a body of research. So most of you know what research companies do, but I'll just tell you, we hypothesize a problem which we oftentimes just get from clients. We do a massive literature search of what the current state of thinking is on that particular domain. We determine if it needs to be evolved or not. We do a lot of interviews, we do a lot of surveys, we assemble that data and analyze it In a particular methodology that we use, we build maturity models and we decide how to produce that material in a form that you can use it, whether it's a report, a podcast, an article or a tool or other forms of digital or human distribution. Then we decide how to market it, then we decide the channels we're going to market and so forth. So if we did a task by task breakdown of all of that, including the publishing part, creating the graphics, doing the layout, testing the website and all that stuff, we would probably come up with, I don't know, maybe 300 or 400 tasks. However, when I look at the actual output of what we're generating around AI in Galileo, for example, we can throw a lot of that away. We can re engineer a lot of it. For example, the process of interviewing people, which we do used to do without recordings, we can do via recordings. And we can use those recordings to more quickly assemble interview data and consolidate interview data across lots and lots of companies that we've talked to. And that doesn't, that's not automating the task of interviewing. That's re engineering or reinventing the way we do this work.
[00:08:38] So the, the risk you have of the task analysis, task intelligence approach is you're institutionalizing the old way of doing it. So we actually did this. We have data in Galileo from a company called Rejig, which is another one of these companies that does task analysis. And you can go into Galileo and you can say, what are the tasks of an HR business partner and how much of that is automated? And let's evaluate our business partners versus these business partners and so forth. And it's a very interesting exercise and you learn a lot. And what you find out is that there's a bunch of stuff that people do that you never realized they were doing. And you look at that and you say, wow, that's ridiculous. Why are they doing this? That should be automated. Why are they doing this? That should be automated. So you end up automating and improving the job structure you have today.
[00:09:32] Now, there's nothing wrong with that because that's maybe the first step or the first generation of automation is to take the jobs you have to and automate them. But I really think it's a little bit of a, I wouldn't say fool's errand, but it's a little bit of a distraction because the really good manufacturing plants didn't automate the manual process that humans did. They started from scratch. And the manufacturing engineers looked at the car, not the jobs. They looked at the output of the car. And they said, how do we get that car developed in a way that's more efficiently? And by the way, maybe we need to redesign the way the car is designed in order to manufacture it more effectively using a process called design for manufacturer as opposed to automate the tasks of the, you know, 1920s Model T production line. You see where I'm going here? So I think these tools, this, this approach, this whole domain is a little bit risky. And the reason I say it's risky is because we just lived through this with skills. We just spent 10 years, maybe more as an industry building skills models because we were told by someone we don't know who wasn't me, that we need a skills model to run our company and that if we don't have a global skills taxonomy and we don't keep it up to date, we can't run our company well. And then our software partners built skills engines into their software to institutionalize the skills. Well, you know what happened with the skills work? It is, it was and is very useful. It's very useful for hiring, for selection, for capacity planning, for training, for professional development, for succession management. It's very good for all that stuff. But the word skill is not a static concept. My skills in doing my job today are completely different than the skills I used two years ago. Completely different. The conceptual skills are the same, but the actual granular skills are very different because I'm using AI in a way that I never even conceived of. So, you know, if you went to the API of all of the skills, quote unquote, that come out of Lightcast, and those skills come from scraping job postings. So Lightcast is not inventing this. They're looking for real skills in demand. You would find that every single day there are 20 or 30 or 40 new skills, new phrases of words that describe an activity or rather a capability that somebody's doing at work. So I think we have to be a little careful that we don't turn the task activity analysis project into an end of its own. Now I think what's going to happen, we're early days into this. I think what's going to happen is most of you are going to start using these tools. You're either going to get frustrated by them and stop and just give up because your job architecture and job titles are so messed up, or you're going to discover some things that are really surprising. You're going to discover that some of these sort of cross functional jobs like project management for example, or program management or data analytics are very, very widely used in different parts of your company and they are not at all standardized. And so one of the things that's, I think, valuable about this exercise is you will find these cross functional similar jobs and you can consolidate them. I remember Stuart Logan, who was at BNY Mellon at the time, did an analysis of the project management and data analytics in some of these big jobs and he discovered that every business unit had people doing jobs with the same title. I saw this in some other job architectures too, but the actual work was slightly different in each part of the company and there was no consolidated view of what those jobs were. So there is an area where task intelligence could be very useful. And you could say all the people with this job title, say QA engineer maybe for example, or pre sales, sales rep, sdr, that kind of thing. And then we're going to consolidate all these tasks and we're going to build a tool that automates that set of tasks or some of those tasks. And all of you guys who have this job title are going to use that tool and some of you aren't going to have to do this job anymore. WPP has been going through this for many years in the advertising industry. And so for most of you that have grown companies through acquisitions and you have lots of distributed business areas where there isn't a lot of consistency between the job titling and the job architecture in these different business areas, you're going to learn a lot. But I want to warn you that if you try to do an enterprise task analysis to figure out what to automate, it could go down the slightly dead end of skills analysis. Now I'm not at all negative on this because I think all of these kinds of projects are massive learning experiences about how your company works. And the interesting thing about AI is it's so powerful and so flexible, you kind of have to see the big picture before you decide where the heavy big investments are. So I think in a lot of larger, bigger, older companies like banks and insurance companies and any process oriented company, professional services companies, this falls into the category of what I would call process mining. Process mining, a different domain, but very related to this, where you actually, I think it used to be done by consultants and now it's done more and more by tools. SAP has a lot of tools that do this, where you process, mine and monitor and evaluate a process in detail. The real process, not the theoretical process. And you look for areas to deduplicate or streamline or eliminate steps. And that's part of the re engineering process too. And interestingly enough, let me tell you something that is also sort of a factor in this whole process. You can do process mining with AI without any of these task intelligence tools. And here's how. If you're a company, not some companies can do this, some can't. If you're a company that allows your AI to index emails, if you look at Sana or you look at Copilot, or you could do this with Claude. Any enterprise AI infrastructure, one of the data sets that it can ingest is email traffic, email data. It can ingest meeting data and actual conversations. And then obviously it can inject policies and you can put documents and other digital artifacts into it. But let's suppose you put the emails in there also. Well, in our company, I've told a lot of you this, we have a system which we call Digital Twin. It's from a really cool company called Vivint, which we can talk more about later. And what it does is it indexes all of the emails of all the people in our company, all of the documents in the Microsoft cloud, in the Microsoft graph rather, which is all of the documents we create in Microsoft Tools. And then we have all of our business related meetings in SANA and other data from things that we've generated or created on Galileo. So we have all of that in essentially one big corpus. So I went through our Digital Twin maybe about a month ago and I asked it, please describe our process for selling and closing a contract of a research membership, corporate membership. And the first statement it said was, you do not appear to have a process for this.
[00:17:06] And it said, here's the way it seems to happen, here's the way it's usually happening, here's the center points of control that seem to be common in all of the deals you guys do. Here's the areas where it seems to be not common. And literally I could talk to it, and through just literally talking to it, ask it questions about areas of the process that it thinks we could improve. Now, we're not a huge company, but we've been doing this a long time. So it has a lot of intelligence about all of the activities and deals and relationships and different corporate accounts that we have just by looking at the emails. Because if you look at the email of everybody in our company, which is much smaller than a lot of your companies, every detail of every transaction that we've ever done is in there in many forms. Not, you know, completely, but if you add the data from our conversations, our meetings, where we had meetings with clients or about clients. There's a lot of data there. So regardless of whether you bought techwolf or Rejig or you know, lightcast task intelligence stuff or any of that, you actually could do this yourself. And at the rate of pace of improvement of AI, I think that's going to be a very, very viable solution. Now, I don't know that there's one tool that does all of that indexing yet. And you know, a lot of companies will not allow their organizations to old emails around. We were with a big company this week that said that they have a 90 day retention policy and they don't want to keep old emails around. I don't know, I think it's probably for legal reasons, but I think from a business re engineering standpoint, this is a gold mine. And it may be ultimately much more useful than buying a bunch of task data from the outside world and going through and analyzing it and trying to figure out how many tasks we have and who's doing which one and what's our capacity and tasks. So we're going to do a lot of research on this. This is part of HR20, so we're going to keep digging into this and talking to you guys. If any of you are doing task intelligence capacity planning projects, please let us know. We're going to be doing a webinar with the company on this actually in a few weeks with drop and we'll, we'll continue to evolve what we're learning about it. I think the, I guess the encouragement I would give you is don't assume that just because there's a tool that seems to magically create this report that looks so useful that that's the right way to do it. All of these AI initiatives, all of these AI agents, all of these AI rollouts have to be done in the context of the business problem or the business goal. And I brought this up a lot this week at Workday Rising. We're not implementing AI for the sake of AI. We're not doing AI transformations. There's no such thing as an AI transformation. This is a business transformation to solve a business goal or mission or objective using AI. And if you don't know what the business goal is, doing a bunch of task analysis to sort of see what's going on and guessing where that's going to take you may or may not take you, unless you're thinking about the business goal down the right path. We all have plenty of efficiency and productivity opportunities in our companies. I mean, that's just the way companies are, human beings are not machines. Every human will do the same job differently. And in that humanity of our companies is the creation, the intelligence, the innovation, the good ideas that a machine may not have. So I don't want to assume that just because we can break all our work down into tasks that the humans are irrelevant and we'll just build a bunch of task execution machines and we'll be done. You'll end up with a very uninteresting, weird company if you do that. A company that's really good at doing a whole bunch of stuff that nobody wants to buy, for example, or that nobody wants to do business with. But nevertheless, robots are coming. Obviously our AI white collar robots are here. So this is a really, really valuable domain. And we're, because we're so connected to so many organizations, including Workday and SAP and all of the pioneers, Lightcast Drop, all the pioneers in this space. We'll keep you guys up to speed. And any of you that want to sort of debate and discuss what you're working on and share with us, we'll be, we'll be happy to talk to you and share that with the rest of the community. Okay, that's it for today. I'll see you guys next week in Europe.
