This is Your AI Copilot Speaking

Scott King

Well, hey, everyone. I am Scott King, and thanks for joining this episode of the Union Podcast. I’m joined by Chris Kraus and John Michelson. Hey, Chris. Hi, John. How are you? Good. Today, we’re going to talk about AI copilot. So, we’ve been receiving a lot of questions about what is an AI copilot. What is it good for? What is it not good for? Which one’s better? Which one’s worse?

So there’s a lot of confusion about what these things are. Chris, I would like to start with you just to understand what an AI copilot is. Maybe you can discuss what they’re good for, like what they do, and then maybe we’ll talk about some examples. So, you know, to kind of think about an example, like I obviously am thinking about the Microsoft example, but I would like to know, you know, one, what they are, and then let’s talk about some of them.

Chris Kraus

Yeah, so if you’ve been following The Union, you’ll know we had a thing where we talked about those different types of AI. So in general, when people talk about copilots, they’re talking largely about the generative AI part of AI, not predictors, not categorizers, which there’s tons of value in other places. But this one, an AI, think of it, they’re really saying an AI copilot that is based on a GenAI engine itself. So they really are focusing on that capability. You’ll hear people, and I saw many Watsonx commercials for AI code assistance, modernizing your old legacy couple code. You’ll see things that Salesforce is talking about; let us help you write an email back to a customer, and AI will read the five emails in the chain so you have the history of that. Summarizing the sentiment of a customer and things like that. So, copilots are gonna be around for a long time.

You know, this isn’t Clip -It. This is much better than that. But it’s not Nirvana. But realize it is. Yeah, everybody remembers Clip -It, right? Yeah, it was the original.

Scott King

Clip it. Is that the little paper clip guy? Oh my god.

John Michelsen

Don’t admit if you understand, you knew what Chris was saying referring to, don’t admit it. You don’t have to.

Scott King

Just means you’re old.

Chris Kraus

Yeah, and so, but they are they are specifically, you know, tasking with generative AI. So if you look at what’s out there, you’ll hear a lot of things like Salesforce is talking about theirs literally sick. When you have customer service, you have emails that customer sends you. It’s going to read those to help you summarize the case. And maybe it’s the call wash up afterward, or it’s like writing an email response to someone helping you write the email.

So it’ll give you the shell and then you edit it to go back. What I thought was really interesting is we’ve heard this concept of generative AI assistance in the legal area where it can actually draft a document. But then the problem is the people who learn how to draft documents never graduate from three to four-year lawyers to five-year lawyers who proofread those things. Well, that same problem that we talked about earlier is going to happen.

I mean, I remember my first job out of college. I was writing COBOL programs for the 911 dispatch system. So, you know, I would write my program, and then I would sit down with Steve Bishop, who was the head of our department, and he would tell me, Okay, you wrote this. It works, but this is what’s efficient. This is how you’re using memory. This is gonna cost swaps and CPU. So, the copilot does the same thing. It’s gonna get us that first cut of code quickly.

But that doesn’t mean it’s the most efficient code. It doesn’t mean it’s taken into account, like, how does a computer interpret these commands? So yeah, it’s going to help me get code on the page, if you will. But it’s not going to be the best code. It’s not going to be the most efficient. There’s still that human who’s going to approve and look at it and help with efficiencies.

John Michelsen

Well, Chris, you know what? A lot of stuff that I do, I’m just glad something else is going to be able to do it for me. And I will barely spot-check it because it doesn’t matter if it’s efficient or inefficient. So you’re absolutely on target, right? It’s not like it’s not actually attempting perfection, right? It’s a completion engine. So it’s going to do the most obvious: take the statistically most likely path through some kind of prose or some kind of coding activity or whatever. But I’m happy for that often.

Chris Kraus

Yeah, well, especially an email triage. If you think about it, do you want consistency of tone from the company to be empathetic? You don’t want the person’s, I’m having a bad day to show up in their email. If a computer is saying, if you asked to interview, respond with empathy or with concern of the customer’s problems, the good thing is you’re going to get very consistent results there. So you’re not going to have the human emotion there. It’s going to be more deterministic, which is probably a good thing, right? And so it’s never going to fire off a 2 a.m. email and flame someone out because it’s going to be more consistent unless you explicitly say, write this as a flaming email. So I think it’ll be good.

John Michelsen

Well, when you say never, the folks who are scared of AI are just convinced that’s exactly what’s going to happen as soon as I put that thing in my Outlook toolbar; then it’s going just to start generating flame emails on me and it’s going to it’s going to listen to what I really think of the boss and start writing emails anyway, but but no, we’re not we’re not we’re not there. That’s not what we’re talking about.

Chris Kraus

Yeah, it’s now. Yeah, this is not robots taking over.

Scott King

So Chris, I mean, you outlined a couple of different examples. And if I’m a certain role, if I’m a developer, I use a code assistant. If I’m a marketing person, I use a different one. Or in customer service, maybe they have one that runs inside Salesforce or desk .com. So it’s kind of interesting. It feels like there will be a certain role that will use, maybe one, right? But in a company, it sounds like there may be several to govern.

Chris Kraus

Well, you could end up with both, right? So, developers could end up with three or four code generators. So that whole proliferation of technology, let’s like, pick one, and everybody uses the same one. You know, you may use one in SAP, you may use one in Salesforce, you may use one in Zendesk. So, like, and then the marketing tools will have a different one. So it’s most likely if we’re thinking about task-level automation and you’re tasking in a piece of software, you could end up with five of them that you would use Scott because you could have one in Outlook, you could have one in your Team’s meetings, you could have one in HubSpot. And so we have to be concerned not that we’re using it, concerns like how much proliferation? Do we really need five of these to manage licenses and understand? Or should we like pick one and use it over and over? One that’s really good with the skill of tasking on coding, one that’s good on email responses, that could be customer facing or marketing facing. You want the same skills there.

How do we prevent that proliferation? Because you could literally end up with 20 of them before you know it.

Scott King

Yeah, 20 seems like a lot. Yeah, John, if we introduce, Chris said 20, so in between one and 20 or more, what’s going to happen when we start introducing all these AI copilots?

John Michelsen

And it’ll happen. It’ll happen.

Well, unfortunately, it gets confusing. It’s already an overwhelming desktop experience for most people who sit in a knowledge worker job at a desktop. And now we’re going to be hitting them up quite a bit with not only do you get your many apps, you’re going to get the copilot built into those apps, and you’re going to get a Slack or Teams experience full of another dozen, right?

And to some degree, if they’re providing a certain amount of value, maybe you’re going to suffer the consequences of having all of that disintegrated AI strategy, but you’re going to get something from it. Or, most likely, like what we do already with these enterprise apps, we ignore 80 plus percent of its capabilities, and we’ll probably ignore 80 % of the time what the AI co-pilot would be doing. And just, you know, sticking to our day jobs, right? And doing what we already know how to do. We, the, the, the fact that there are so many does also give you the struggle of, and we know this quite well. And I know that listeners have heard you guys talk about this already in a few podcasts that you have to talk about these different generative AI solutions differently. It’s not like there’s a universal English, right? It’s not like we don’t have, and even, even though they are, driven by unstructured English or hopefully reasonable sentence structure guys, right? Talk about these things with appropriate English and get a better answer. But even beyond that, right? Properly structured English that doesn’t work within their structure. And we have dozens of examples of this, actually hundreds of examples of this, where we can see prompting an LLM with this content works perfectly. Prompting LLM prime doesn’t get us an answer at all or gets us the wrong answer.

So I think we’re going to have a number of those kinds of proliferation problems. But maybe, like, I think where Chris was going, if you’re a developer, you’re going end up with one to prevent you from having to deal with email nearly as much as you have to now, and one to help generate template code, maybe we’ll say, get you rough starts on many coding tasks.

Those sound like pretty good things to me if I sit in a developer’s chair, right? Which I often get to do. And if I could make email go away for the next hour and a half so I could get something done, it’s really great, you know, that sounds great, right? So, clearly, we’re not describing co-pilots as bad. Of course, they have some struggles or challenges and we’ve identified some, but I think used judiciously, I think you got, you got to, it’s a net positive, right?

Scott King

Yeah, it seems easy to start with an AI copilot because it’s kind of an isolated use case, right? So I imagine it’s not the ultimate method to introduce AI into a company, but maybe you get your feet wet with something like this and understand the shortcomings. What do you think about using this versus something else?

John Michelsen

Well, it’s the motion we know how to do. That may not be a good thing, but when we get a new piece of tech, we throw it on the desktops and say, hey guys, sort it out. It’s something new for you guys that you should be able to use and it should be able to make you more productive. We do that all the time, right? Humans are the integration strategy. So if we think about every other major piece of tech we’ve brought in that we thought was gonna do us some good, it wound up, you know, the techs vendors’ goal was to be sticky on users desktops, and the consumers of that are basically overwhelmed with more screens and more who where do I go for what and what does it call that? Versus the other apps that call it the different things and all of that good old-fashioned complexity. They were all way too comfortable, I think comfortable with on the IT side, which is, Yeah, it’s complicated, but this is a really good feature. So you’re gonna deal, right?

Scott King

Yeah, it seems like we may get to a point where there’s a prompting expert to prompt from one AI copilot to another, right? If you think kind of like a business workflow and you have several of these, people are still integrating the data in their heads and retyping information elsewhere. Do you think that will happen?

John Michelsen

So, of course, the use of a copilot doesn’t automatically integrate it into back-end systems. It certainly doesn’t integrate it into more systems than the one that the copilot is built within if it’s an embedded one of those. And the coding copilot certainly isn’t going to help you with your email problem. And it’s certainly not going to go and get customer data out of your CRM to implement some kind of function, right? Humans are still, and this is the strategy that we obviously work against, humans are still the integration strategy. And this is just another motion that kind of proves that point. That’s not to say the copilots often don’t have an incremental productivity benefit. And we’re glad to see them. We’re actually users of them ourselves, right?

Scott King

Yeah, yeah. So, you know, people want to like to transform a business process, your digital transformation is just an overused buzzword. Is this going to do that, or do we need to do something else?

John Michelsen

Yeah, so that maybe is where this goes. It gets it becomes a struggle for me because, of course, we all love to see the new thing, and then we all love to prove that Gartner was right about the hype cycle, right? We run up the score on what we expect this thing to do only to realize that we were so overhyped, and we crash all the way down into the even below where it really can deliver value and that eventually we reach that point where we’ve recognized its capabilities, we leverage it for those capabilities, and we kind of sort it out, right? The classic Gartner hype cycle. And AI copilots have run up the hype cycle really, really high. I don’t even think we’ve reached the top of them yet. But Scott, you bring up a huge point. There’s nothing related to outcomes. There’s nothing related to digital transformation. I just don’t think it’s even in the conversation, right?

Optimizing a fairly finite task of an individual person on their desktop is, if the next thing I say is, the key to our digital transformation. Oh, come on. Right. I mean, that’s like a ridiculous statement. Or is it going to drive a bottom-line impact for my business? I mean, my company is going to see that extra two points of margin. No, they’re not. And again, hopefully, you’ve gotten us well enough. We like these, we use these, they have a role. There are some challenges, of course, with everything. But an AI copilot, at least nothing we have seen, just has the scope to do it, right? The one that I think probably most of the folks who would be watching or listening have played with is Microsoft’s Copilot, which is pretty cool, full of, and by the way, they’re gonna improve the heck out of this thing over time, right? This is a Gen One thing, and they will make it better. But some systemic issues are that it’s not a full Q and A system. It doesn’t go through the whole ingestion of content and the vectorization of that content, the chunking and all of the understanding, all of the stuff that a full solution in that space would. So you got a couple of thousand bytes in a document, and when you’re beyond that, it’s either chopping that content, or it says, hey, I better not do this task at all for you because it’s more than 3000 bytes. Well, I think the number is 3000. And by the way, that number can go up, right? It can go up over time. But we’re working with very finite amounts of content, even in the space where it should be excellent. And so when you think about it, I’ll just put this on all my customer service people’s desks, and then the gigabytes of content they have will be available to them. No, no, no, no, no. We just completely missed the scope of what, at least today, copilots are capable of doing. And it’s actually a very sophisticated solution, not centered on the desktop, but centered in the cloud, that is actually going to be capable of ingesting that kind of content in such a way that it then puts it into a solution that someone can query with natural language and get reasonable answers. So we got a lot to do there. And it always just reminds me, you know, again, we’re running up the hype cycle, so I can understand that the hey, well, why can’t it read my inbound emails from customers? Look, create the tickets in the CRM, look up the in the gigabytes of knowledge management to answer questions, or if it’s if the issue is related to or whatever, do so in the other systems around and then reply back.

We just mentioned the content size problem, but it doesn’t integrate to other systems. It doesn’t actually run unattended. It’s a human hitting a button at the end of a typing session is when that copilot actually starts to do something. So, that inbound email is not particularly useful to a copilot if you’re not sitting there asking if it can help me with this reply. So there’s just, it’s just, maybe it’s sufficient to say we’re not talking about an outcomes-oriented solution. And by the way, that’s not its designed purpose, right? So we’re, we’re not saying it’s a bad thing. We’re just saying it’s not applicable to the, we’ll say, digital transformation to reiterate a comment or the way that you really want to bring AI into an organization for a massive kind of effect. And I would also leave it as to reiterate a comment, expand on just a second. So Chris brought up the notion that it’s, let’s just start with, this is a generative AI type of AI, which is great. We’re all excited about it. Have been you know, and we have been for a long time, right? We were, we were seeking a better AI building our own and doing lots of things. We can reach a place where we are even better with generative AI. And, uh, you know, we’re all happy about it, but generative AI is one of countless styles of AI that businesses need to be able to exploit the calculation of dates and numbers, the prediction of classification systems. Those activities, truly groundbreaking transformations, are available there, and GenAI is none of those things. But when you bring all of those AI capabilities together with the integration of people and existing systems and all that good stuff, now you’ve got the ability to basically say, okay, I now have, I see a massively AI future truly groundbreaking transformations are available there, and GenAI is none of those activities exploiting and incredibly. That’s when you start to say, okay, I’m going actually to impact the bottom line. I mean, this company is going to have a completely different operating algebra. And it just, we’re not talking about any of that stuff if we’re talking copilot. So keep your, keep your, you know, keep your confident, you know, keep your, you know, stay real grounded here. Let’s, let’s, let’s, but, but of course.

We like the fact that maybe it’s introducing people to AI that hadn’t actually interacted with it before, right? Getting back to the original comment. And it’s going to do some things for me that I certainly appreciate. I mean, I really just, the number of people who just need me to validate that they’re, they’re okay is too high. So for me to basically say, validate this guy’s okay. And that’s it. I don’t have actually to go through the laborious process of speaking. I’m sorry. I’m being, I’m being rude, but it’s, but, uh, and, and that isn’t true. If only my emails were all that. But the fact is, whatever it is that you’re able to use it to take minutes off of your tedious schedule, great.

Scott King

Yeah, I just hope it would. The fact that you did talk about outcomes and user-level things, the fact that a lot of these are user-level pricing just shows that it’s really focused on the user itself. They don’t want to remove users. They need more. So it’s interesting. They are helpful. I mean, I use a couple of different copilot tools, and you have to spot-check a hundred percent of the time, and 90 % of the time, I’m fixing something, but it’s a huge jumpstart. It does save a lot of time. So, Chris, we talked about the, you know, what the copilots are, some examples. John talked about the workflows and actually transforming the way businesses make money, right? We’re here to make money. What can you leave us with? If someone was looking into copilots or they already had a copilot, maybe give us some advice on what we can do.

Chris Kraus

So if you look at your copilot and you say, hey, this helped me reply to an email. But then you say, wow, there’s stuff that I want to automate on my desktop. When I get a second meeting that conflicts with the meeting, it is if you’re asking why I have to manually reject this meeting because this meeting is more important than that one. Or why am I doing? You’re still doing a lot of tasks around that copilot. Think about if you’re saying, wow, could a copilot with some integration and some AI look at two conflicting meetings, categorize one as high priority and low priority, and then reject one politely and then confirm the other one? Then you’re ready to talk about doing great things because you’re actually removing that nonsense of having to look at every meeting and manually saying, which one do I take because I always get two or three that conflict? So I think when people start thinking it could do more than just reply and email, they’re ready to really start getting value in large adoption.

John Michelsen

So Chris is on the I’m a really popular guy theme.

Scott King

And you’re trying to get email to reply to people asking about Chris, right? So, Chris is our lowest common denominator here today.

Well, perfect. Hey, well, thanks, guys. I appreciate it. Hope everybody learned a little bit about what an AI copilot can and cannot do. Obviously, if you are into copilots and you like it and you want to take that next level of automation, please contact us at Krista and we’ll be happy to help you whatever your use case is. So, until next time, thanks so much.