Agent Washing: How to Spot If You're Being Sold an AI Agent That Isn't

Speaker A: So imagine you're buying a robot that's supposed to think for itself. It's not a far fetched idea, it's

Speaker B: the future that we're living in.

Speaker A: So the product comes in, the box, looks amazing, the paint is still shiny and it says, smart, independent, agentic agent.

Speaker B: Then you actually give the robot a

Speaker A: real job to do and that paint starts falling off. And underneath it's just the same old automation we've had for years. That's a real thing that's happening right now. And it has a name. It's called Agent Washing.

Speaker B: So here's the problem.

Speaker A: Thousands of companies are offering AI agents right now, but only a slice of

Speaker B: them are the real deal.

Speaker A: The rest is a fresh coat of pain. And the result is that many of these companies will suffer because their AI project, as a result of this, will be canceling their projects by the end of 2027. But if you're assigned with the task of finding one of these projects to implement in your company, well, if the project goes flat, you're the one on the hook. So how do you determine whether or not it's real or fake?

Speaker C: You should be asking questions such as, you know, what does this thing do? When does it do it? What is the outcome? How much is going to be the cost of operation afterwards? What's the billing model for this tool if you're buying something out of the shelf?

Speaker A: Well, it's not as easy as determining whether or not it's a real Chanel bag or a fake one purchased on Canal street, but that's what today's episode's for. Our guest joins us to determine whether or not the quality they're selling is the quality they're offering.

Speaker C: Um, my name is Mariana Horic. Uh, I'm a senior product leader, Making sense. Um, I work on the AI innovation department too. Um, basically what I do is I'll help companies understand, define, design, the best strategy to implement AI on their organizations.

Speaker B: Amazing. Mariana, I want to thank you, first off, for joining me on this episode of Brains Byte Back. And those of you who are listening. And today, we're going to be digging into something that I know has been happening for the last little while, but finally it, uh, has a term, and it was coined recently by Gardner, and the term itself is called Agent Washing. For those of you that haven't heard of it before, it's basically this growing gap between what vendors are selling as an AI agent and what actually holds up once they try to put it to work. So basically, think of it like a Vendor rebranding, old automation as new agents. And and Mariano, I know that you work directly with mid market buyers so they're the ones that I think are most exposed to this. Um, so we just kind of want to get some practical insight um, about basically what this looks like in terms of somebody just trying to put a fresh coat of paint on something that is basically automation. But I want to start by asking you, you framed a lot of enterprise AI as basically plausible noise mistaken for real value. When every vendor right now is basically slapping autonomous agent and I say this in quotations, um, on their product. What is actually driving that right now?

Speaker C: Yeah, I think that's uh, a, it's a good, good question and uh, I think it goes a little bit behind that and is more towards can a company mutually agree in terms of what an agent is today? Um, I feel like the term is used so broadly and um, with all the you know, advances that uh, have been happening lately, uh, especially on the entropic side that they, those guys have been releasing features like non stop and they've been pushing the ways in how things are done very rapidly. Um, I think that it comes to first getting an agreement to what an agent is. And for me, um, for the listeners, um, an agent is composed by mainly four things. Um, it has to be autonomous. Mainly that means that the agent doesn't need you to be putting individual prompts to get to the goal that you're expecting, but rather uh, pursues the goal on its own. Uh, the second characteristic will be that it should reason and plan M mainly what I mean by that is uh, the system should be able to think the different scenarios and then choose the most proper uh one based on the task that is being given. Then the third big difference is uh, when you're talking about agents and agentic AI normally this systems have access to the digital real world. Um, mainly they can interact with uh, external systems, they can browse webs, they can run code, they can manipulate files, they can call APIs, they can integrate with different MCPs. So there is a layer of access to the real world to make actions and to use the available tools that you provide to the agent. That I think that's also very relevant. And then the fourth one will be the capacity to learn and adapt. Um, normally with an agent there's going to be two feedback loops. One the agent is going to evaluate themselves and try to do better next time. But also you can give them active feedback on plain English, which is amazing. Um, so I think that's what it defines um, agentic, if that makes sense. And it's been very mainstream and many, mainly I think it's because agentic is associated with autonomous and autonomous is associated with less headcount. Right. So every company that we're working with, they are trying to get their gentic AI strategy in place. But at the same time I think people are, and I completely understandable, I mean they're a little bit behind uh, in terms of what that means and how does it work. So they ended up calling third party vendors that help them, you know, figure uh, that out. And some of those vendors have this, you know, pack solutions, um, that end up being more like RPAs, uh, raw process Automation or work automations rather than like a pure agentic approach solution.

Speaker B: It's interesting that you say that because it makes me think of this um, video that I came across and I think it was on TikTok and the content creator, she basically kind of described what's going on right now. Like the crypto age where you got a bunch of guys that I guess are crypto bros. They realize that this is a product that is really hot. People are like interested in it and then they're kind of like selling you uh, coins that you don't really, that really have no value. So at the same time it's like there's a lot of people trying to get into this market and it seems very saturated and a lot of people are like, hey, I'm going to fall behind or the sales tactic is I'm going to fall behind if I don't have this, you know, in my workflow. And you're connecting, you know, with you know, people that you see online or you know, the algorithm that kind of throws you this, you know, sales guy and they're basically selling you a product that you don't necessarily need. And the way I see it is that it's challenging because some, some of the people that are being tasked in these companies to find these um, you know, solutions I guess is the best way of saying it, aren't necessarily, uh, well, like have a lot of knowledge when it comes to like AI and you know, what is the best solution, especially because it's growing so fast. So what, what's this? I mean, how would you define what is kind of like the marketing tactic like early on, right before maybe getting into this demo or while you're getting into the demo. How do you define like they're just trying to sell you something that, that maybe you don't really need instead of like um, it actually just, you know, being something that you actually need at the end of the day.

Speaker C: Yeah, I don't know if it's like bad intentions. I feel like it's more a situation where many companies are pushing blindly towards AI and um, whatever is the latest and greatest to be implemented at the companies. And sometimes they forget that programmatic solutions or even work automations are still great options and they have, uh, a time and space and the solutions. Actually, many of the companies that I work with, we strongly suggest to go with workflow automations because the problem that they need to solve has a, um, clear input, clear set of steps and a clear outcome. And it's always the same and it never changes. So on those cases, a workflow automation approach is much, much better than an agentic AI approach. Um, and you can think about, for instance, compliance, uh, validations. Right when you're validating the same thing over and over again on different files, but the validations necessarily change. Um, that's a great example. Like invoice consolidation, for instance. That's another great example for workflow automation where agentic is probably an overkill. And I will encourage companies to always think about the problem first, uh, and what they're trying to solve. And I'll give you two quick examples to represent this. Um, if someone comes to us and tell us that they want to do an agentic solution for their employees to be able to check, how many days of PDO do they have left? Well, that's for sure an overkill. M. I mean, that's an API call. That's all that it needs to be. It doesn't need to be any more than a good polished programmatic solution. But now if the problem is different and it's more complex, for instance, a business owner comes to us and says, hey, I figured that during summer, um, after the pto, my employees are slow and sometimes they are sick because they get sunburned. So I want to give them tools to know exactly what type of, um, sunscreen and how much should they apply depending on where they're going. So now, as you can see, this problem is much more complex because now you have, uh, the person itself, how much does he weigh, how high the person is, uh, what time of the year are they traveling, what part of the world are they traveling, how many days, what is going to be the forecast on that place, uh, how much exposure are they putting on getting to the sun? Um, and there is all of these variables that makes the answer very complex to get to where, if you try to Put that in a programmatic way, it's going to be very difficult, very, very complex. And if you put that on a workflow automation, it's not going to work either because the variables are different. Right. And what if the forecast service provider that we set up, uh, does not have the location because this person is going to go to a remote island in the middle of nowhere. Uh, so the workflow breaks. Right. So you need the agentic capability to be able to solve that problem and continue towards the solution. That is a good example for an agentic solution, for instance. But again uh, it's more about what is the problem that I'm trying to solve and what is the right solution to solve that problem and not over engineering. I think that's the main advice that I will give to the audience.

Speaker B: And just to add on that, I think, because one of the interesting things with the Gardner report is that they mentioned that about 40% of agentic AI projects will be canceled by the end of 2027. And that's a mix of like, because the costs are escalating because it could be very costly, obviously unclear business value as more people are realizing that this is an issue.

Speaker A: Right.

Speaker B: How are you kind of helping these mid market businesses make sure that the types of solutions that they're adopting would most likely be successful? Because I imagine people are seeing these numbers and their fear right now is that they're going to be part of that 40% and obviously they're like, hey, we want to m, make sure that if we're investing in this that it's going to add value in the long run.

Speaker C: Yeah. I think that there are two parts of that. Uh, uh, the first one is if there is something out of the shelf that solves your business, probably is a great idea to get it. Uh, of course you should do some research to make sure that the company's not going to disappear tomorrow. Right. It's not one of those um, thousand startups a day that we get an AI. But um, if there is something out of the shelf that solves your problem, probably that's going to be the most cost effective way of solving things. Now if you need a custom solution because it's um, a key part of your business or it's very strictly related to how you perform business or how you deliver the service, then there are good practices that you can follow, um, to control that. Um, cost of operation is a big thing that nobody thinks about in terms of burning tokens pretty much. Um, and there are a lot of things that you should do along the way while you implement these solutions to prepare for that and optimize it in the best way possible. Um, things such as um, running, setting good parameters, access and controls and guardrails for the agencies is very important. But also having a good evaluation pipeline at the end of the process so you can actually determine how, how accurate are you being and how costly your prompts are being, uh, on the execution basis. And probably those concepts are things that people don't think about when they think about these solutions. They think about like the good things, oh, I'm going to do this and it's going to do everything by itself and it's going to solve all my problems and I'm going to be able to be much more efficient. But uh, there is, there is a cost associated with that, uh, that needs to be um, you know, considered and I feel that that's where, you know, a good consulting firm could help you understand, define and design that solution correctly for your business.

Speaker B: So Mariana, I'm, I'm sorry to put you on the spot, but I'm going to ask you this at I guess a lot of the conversations that you have with you know, the big companies right now there, there was another report saying, and I've seen this for the past year, that companies that are kind of getting rid of staff and bringing on AI to take on these roles that what's going to happen is it's, they're going to backtrack, they're going to go back and they're going to rehire a lot of these people. What are these conversations that you're having? Like, are there legit, like, and I'm not saying name names here, but are connecting with some companies that are being very honest with you from the get go to be like, hey, we're looking to reduce overhead, cost and things like that. And part of that is that we want to invest in agentic AI to take uh, up most of these roles because I'm not sure what a lot of people are talking about is that right now it's being inflated by the media. It's kind of fear mongering in terms of what's going on. But some reports say one thing and then the media says another. But because you're talking directly with these companies, what's your insight based on that?

Speaker C: I think there is a little bit of everything, um, for sure there is the more mindful approach where the conversation is more leaning towards there is this new technology and we don't want to fall behind so we want to Just see how can we leverage this for a business, which is a much more healthier conversation than the one I want to wipe high my whole operation tomorrow and I want to just put agents to work for them. Um, I think a good comparison is, um, I feel like the error comes when companies treat this as a IT forward initiative, as uh, it could be a cloud migration process, for instance, when in reality it's completely different. And the main reason why is because, for instance, when we were on the cloud wave, um, everybody was moving from on premise to the cloud. Um, the approach was a little bit simpler because it was like you leave what you have, you shift it to cloud and then you optimize it as much as you can. Right?

Speaker A: And

Speaker C: all the discipline was measured in how reliable and scalable your infrastructure is. Back then you didn't need to understand how, um, a, ah, person performs their daily task or how are, uh, let's say an underwriter think and then rethink that process to optimize it, to put it and embed all that knowledge on an agent, which is a completely different thing. Right? Uh, because it's not. It, ah, LED is. It sounds weird, but for me it's like very human led because you need the people to be on board. You can get the best solution in the world, but if people don't adopt it, it's worth nothing. And then you need people to be on board and collaborate with you. Because many times we go in the company, they're like, we want to optimize or automate this workflow. And the workflow is A, B and C, and then you get D and you're like, good, okay. And then you go, you talk with people and then people are uh, like, yeah, that's not how it works. I mean, you go from A to Z and then from C to Y, and then you get to B, but then you go to B1 and then you get to C, but then you got to go to M2 and then you get to, you know, D at the end of the day. So the workflow is completely different, the knowledge is completely different. And all of that knowledge needs to be embed on the solution. And I think that that's the biggest challenge on the difference between cloud modernization and agentic AI solutions. And that's one of the main reasons why companies shouldn't approach it as a completely IT initiative, uh, the forefront.

Speaker B: So I feel like that's a great response, especially directly. I mean, speaking with the people directly that do the job, I think is what's Most important because that's what provides you the most insight. But let's get into how mid market buyers basically protect themselves a little bit more. Um, so I remember there's this quote that was sent to me by. So it was buy commodities, build different differentiators. It can get tricky when every commodity is now sold as an agent.

Speaker C: Right.

Speaker B: I'm sure you'd agree. How do you draw that line for a client? Like what's the red flag that exposes at first like a washed age agent the fastest?

Speaker C: Yeah, that's my, that's my rule of thumb for um, building versus buying uh, buy commodities, build differentiators. That's a good, has a lot of, yeah it has a lot of caveats though. But that's like the, the, the big picture. Um, yeah I think it's again uh, I will suggest to either get someone that is knowledgeable in the technology or a good partner. Um, but uh, for the most part um, um if, if you can buy something out of the shelf and your competition can also buy the same thing out of the shelf and whether or not that is agentic, it still seems to be like a commodity for me. So what you should be focusing more is on what the thing actually does. And I um, feel like if you can describe what the tool does on an if then do that type of sentence then it's probably not agentic 100%. Um, so you should be asking questions such as you know, what does this thing do? When does it do it? What is the outcome? How much is going to be the cost of operation afterwards? What's the billing model for this tool if you're buying something out of the shelf. Um, and those things let you again understand if this is something that is going to be able to pick up your problems and figure out a way out to the solution or if it's more uh, like a standard process that is going to be some sort of like adapt to what your business does. Like like with tweaks here and there on the steps.

Speaker B: Um, where would somebody even start looking like, like, like um, most of the time you're looking for something you go into Google now people are going to go into like LLMs and chat. I need to find uh, like you know. Yeah like an engine, like a, like a engineering company that can help me with like software engineering company that can help me with this uh, so or with this issue. But like how do you recommend people really go about in finding a trusted um, partner um to help them implement uh, something into the workflow?

Speaker C: Well the Easiest way is called Colors. That will be the easiest route of course. But, but aside from that, um, again I feel like you should have some sort of, I mean you gotta be trained somehow on these technologies. Either you or some other your organization. I completely uh, disagree with the fact that some companies want to completely outsource this. Uh, there's something that you should be part of your stack at somehow even ah, if it's one person in the organization that can make that judgment and I ask the right questions whenever you're having this introduction calls with these people. Um, but again, um, if it looks like it's um, um, A plus B then C, then probably it's not an agent. But still if it solves your problem, that's amazing. I mean you should definitely go with that. Um, again at this point in time I will worry more about does this solve my problem and does this falls under my budget for cost of operation rather than if it is agentic or workflow automation or on RPA or you know, whatever it is. Um, at the end of, they are all AI technologies, uh, that self a purpose. So I'm always focus on the problem. That's the most important thing. Um, that's what's going to get you the right solution. Uh, don't start from the solution backwards to the problem because you want it to be a fancy AI agentic solution when you probably, you don't necessarily need that. Um, having a good understanding of the problem I think is, is how you prevent yourself to walking into something that you don't need or is not a right fit for you.

Speaker B: Of course. And I think it speaks to this last episode that I did where it was talking about the rise of the AI manager, which is people basically in the companies right now that their role is to make sure that everything is you know, going to plan. Um, but they're, they're internal. Even though they're working with people that could be like an external third party, um, obviously they have the company's best interest in mind. Um, and making sure that you know everything from the guardrails, you know, um, you know, recording their own forms of roi, right whichever way, defining what that looks like. But I think that's really important for people to kind of uh, bring somebody on. And that's an eye opening thing for people that are looking to get into this industry. Whether or not you're straight up, you know, going into university or looking for a career change. You know, people that are kind of in the consulting world because it's not, I mean there's a technical aspect of it, but at the same time there's things that you can learn along the way and the technology is still growing, so they'll be learning things as, as the technology itself grows.

Speaker C: Yeah, of course, of course. And I, I always tell people that, uh, don't try to, you know, know everything and anything. It's uh, it's very hard to, it's very hard to keep up with, with all the progress. But it's good for you to at least understand how the technology works. Uh, sometimes when I tell people that AI is something that gets text and return text, they look me like, what are you talking about? But in reality it's just that, I mean, if you think about it is you input text, you get text back. Uh, with the genting things is a little bit different because you have tools and they have external, you know, access to tools and things from the outside world. But it's just that. Right. So having that into consideration, understanding that is probabilistic, it's not deterministic. Understanding what are the good practices to keep, you know, consumption down? Understanding what are the good practices for getting the most consistent set of answers for all the problems that you're making it? Uh, um, since again it's not deterministic how you make it, you know, as, as consistent as possible, I think those are the important things and then understand very well how that maps out to business value directly. I mean, I think that that's the most important thing. And many of those, um, I think, uh, pilots that you mentioned, I think those could be things that sound grain on paper and they work amazing on the demo environment with all the, you know, this uh, structure, predetermined data, but whenever you put them in the real world, you understand that they don't bring, you know, as much business value as you expected at the beginning. And I feel like having someone that can understand at that level and ask the right questions to the right people, uh, it's a great role to have on organizations right now.

Speaker B: Maybe I want to leave this off on a note for somebody. Imagine somebody that's listening to this podcast, came across it because right now they're looking for a new solution. Maybe they're about to sign a contract.

Speaker C: Mhm.

Speaker B: Knowing that they're going to get into this, is there anything specific that somebody should put on paper before they sign to kind of save themselves? Maybe when signing up with this company, whether or not it's to be like, hey, this is something we're going to try for a year, like, you Advising these companies as like a consultant just to make sure, because I know some contracts can go longer and obviously this is like a long term plan, but something that maybe somebody should have in mind, like as a takeaway. Yeah.

Speaker C: I will highly recommend people to understand exactly why they're trying to get out of this solution and set up some, um, KPIs or goals on the 30, 90 days type of mark. Uh, so, uh, what I'm trying to say is make sure that you have in the contract a very, ah, clear way that cannot be dialed to determine if this implementation was successful or not for the business. Um, I think that that's very important because these things could scale a lot. There's always new things to do, there's always new features to implement. So you can, you know, get into that kind of like rat race, you know, behind, uh, new features and making it better and faster, more scalable and, and this and that. But in reality, again, does it solve the problem? Does it deliver the business value that you're expecting? Can you measure it? And can you say, yes, this was a success. No, this was not a success. I think that that's what matters the most. Um, aside from lines of code, aside from, you know, um, whatever else you know, you get out of this, what is your definition of success for this implementation from the business standpoint? And remember that AI is just another tool that you have in your toolbox to solve a problem. And that's why you should be treated as.

Speaker B: Those are very good points, Mariano. Uh, we're reaching the end. I just wanted to thank you again. If anybody is looking to, I guess, pick your brain a little bit more, what's the best way that they could reach out?

Speaker C: Yeah, I mean, they can reach out me on LinkedIn, um, or we can share, um, my m, my personal contact information. I'm always happy to, uh, discuss different points or be taught new things. So, uh, yeah, open to both amazing

Speaker B: and best of luck to Argentina, because you are from Argentina. You did mention that.

Speaker C: So, yeah, yeah, yeah. So go Messi. And yeah, hopefully, hopefully we'll be, uh, winning the World cup again.

Speaker B: I hope Canada gets pretty far because they've done pretty well. Thanks again, Mariana.

Speaker C: Yeah, thank you, Eric.

Speaker B: That's a wrap for today's episode. If you've benefited from what you've learned today, feel free to leave us a review on itunes. You can also find all our past and Future episodes on YouTube and Spotify under Brains.

Speaker C: Bite back.

Speaker B: And if you'd like to connect with us directly, you can always reach out

Speaker A: to us by email@infoocable.com Remember, your feedback helps us grow.

Speaker B: We welcome it. We appreciate you joining us. Until next time,