Understanding the MIT NANDA Report
AUGUST 25, 2025
Ramsay Brown: So this is the report from MIT's NANDA group entitled the Gen AI Divide, State of AI in Business 2025.
And we found this originally from a writeup of it done in Fortune with the incredibly evocative and market shaking headline of 95% of money that's spent on generative AI pilots ends up being a massive waste.
And you and I have gone through and read the report and that's not at all what it's actually saying.
Andrew Melville: No, not at all. So the report itself goes in depth about the executives that they talked to and they discussed everything from how are they implementing these pilot programs in what areas of their business are they building them internally or are they using external partners to build them.
AM: So what you find in the report is that there's a whole spread of different reasons why these pilots may be succeeding or failing and you really do have to go into the details.
But when it rolls up to the headline of 95% of AI pilots at large companies fail it makes it sound like the tech is not ready organizations are not ready it makes it sound like the entire thing is about to collapse and when you dig into the details you find that well actually 67% of pilots that are conducted with specialized vendors startups and other vendors those succeed 67% of the time. Whereas internal builds succeed far less. And so yes, the headline itself makes it sound like the hype bubble is ready to burst. And then when you look into the details, to me it makes it seem like we're actually just getting started.
And this is really what you would expect from a new general purpose technology for companies.
RB: I want to go into the paper pull out the key findings that I think we want to know, and that the folks we work with would want to know as well. If I had to sum up the entirety of this paper into anything other than absolute "sky is falling" narrative that drives clicks into what's really here - into a single word - it would be "approach".
And that the approach that the enterprise is taking to generative and agentic AI rollout determines its success.
Which sounds like "yeah, duh approach is going to influence how you do this", but the counter narrative to that would be that actually the technology itself is deeply flawed and no one's seeing any value because the tech's just not there.
And what the paper really keeps emphasizing is that the tech is there. It turns out that organizations are going about implementation and going about their approach in a spectrum of different ways, but a lot of those ways end up being incredibly unproductive.
RB: And so there is a way to do this. There is a way to build internally if you want to go down that path [though the report clearly emphasizes work with a vendor partner to do this and do that kind of specialized approach where you're good at one thing your vendor's good at another] and they highlight some takeaways here.
I want to explore these with you, but I want to kind of get 10,000 foot view first.
What jumped out at you? Because you read the thing front to cover many times. What jumped out at you about it?
AM: So, what jumped out to me was that first off, it's that the areas of a business where generative AI tools would be the most effective at driving actual P&L value: back office operations, business process outsourcing, some of these types of very not flashy but fundamental parts of a business. Those areas were not what generative AI pilots were run on.
They were typically run on sales, marketing department, sort of customer facing and front-facing use cases. Things like chatbots and various things helping the sales team respond faster.
There's a quote in the article from a CIO that says, "well, aside from it processing some invoices a little bit faster, we haven't really seen this have much of an effect in our business".
And so back to your statement about approach. So right out the gate, right, the technology is being applied to the wrong areas of a business most of the time or rather to lower value areas of a business. And then when those projects don't deliver some sort of measurable value, very quickly they point to, well, the technology is not ready.
Well, in reality based on the data, it sounds like it is being directed at the wrong use cases.
RB: So I go back to this all the time, this metaphor of this paper that was really fundamental in AI itself, at least here in the Bay around the "bitter lesson". And it's this metaphor about scale: just throw more parameters and more training data and more energy usage at the problem. Even if you've got a general purpose deep learning transformer architecture neural network that's not specialized to anyone particular task, scale alone will beat really sophisticated algorithmic approaches and exquisite systems.
And they call this the bitter lesson because it angers a lot of people who have been in AI for a really long time that this generalist "just make the model massive" approach actually wins. In fact, it wins nearly across the board. And they call this the "bitter lesson".
But I'm reminded that there's about a dozen other bitter lessons embedded in the actual practice in a business of transforming using AI.
And this feels like one of them. And that is that someone comes back from a conference or from some sort of summit or they were on a board meeting and then someone panics saying "we got to have an AI strategy". We got to do something here. And so they end up paying 50 million bucks for a bad PowerPoint from one of the Big Three that suggests a really mediocre idea that this flashy frontline topline front of house implementation that then falls apart and it's a bitter lesson of where would value have actually been created.
"I don't know, did you try month-end close? Did you try inventory management? Did you try ticket resolution? Did you try the actual business operations that drive your firm?"
And that's what this report captures: those places buried deeper that you used to give to BPO or agencies, those actually are hot beds of ROI compared to what everybody's doing.
And that's the bitter lesson.
AM: And I think you have to sympathize in the sense of the current news cycle around AI and so it is understandable that in an effort to say that your company is AI-capable, AI-enabled, etc. It's not a very exciting story to go in front of the board or to go in front of the camera and say "we've just digitally onshored our IT ticketing operations that used to be international and we've now brought that on board we're now using AI to do that and it's saving us x amount of money and the turnaround time has decreased by that" and it's some function that is not very interesting it's much more exciting to say "we've given it to our sales team our marketing group" and then it's not able to deliver the results.
RB: Yeah, this feels like if we're lucky - for listeners who are familiar, probably everybody, with the idea of the Gartner hype cycle - this feels like if we're lucky, this is one of the milestones on the plateau to productivity of "oh, so it turns out those incredibly quotidian tasks your org actually just runs on that are in the guts of the organization. Turns out to be a really great place for sufficiently advanced autonomous automation" as opposed to these incredibly flashy frontline use cases. So I appreciate your empathy.
So if we dig in beyond approach, there is something that they call out in the report that I do think is worth discussing and that is that there are two major flavors of requirement that the enterprise has that isn't getting met: and that is that 1) a lot of the technology they're buying from vendors is brittle. Can't be customized or modified into their ways of working. and 2) that the technologies don't learn as they go in and that you are constantly having to refresh gobs and gobs of context and the model isn't customizing to their own behaviors over time.
And those two major requirements are, according to the report, where the teams that are succeeding in rolling-out things that can learn or are flexible enough to meet the workflow requirements of actual ways of working. Those are where teams are winning.
And then the rest of this outside of those two narrow things, implementation, it's where people are wanting this to make sense.
AM: One of the things that really jumped out was that employees within organizations don't experience a pervasive fear about jobs being automated away by AI, as I think a lot of people suspect. They actually found that a lot of organizations have very high adoption rates of AI already within the organization but it is being done essentially offbook.
It is individual employees often using their own private accounts to use ChatGPT or Claude and to use it to draft emails or write code or do whatever they're able to do. So, it's limiting it to fairly basic tasks.
Again, you don't have memory storage. You don't have the ability to construct complicated agents that are able to actually do work. And so employees are clearly very interested in adopting these technologies. And the report actually calls out that they often find that implementations of enterprise AI tools actually aren't as good or good enough as just using ChatGPT. And so they essentially say, well, this isn't as good or better than ChatGPT, so I'm just going to continue using ChatGPT. And they just reject the custom-built tool.
RB: Which is something that I'm glad you pointed that out because they've got a pullquote in there from a legal team that suggested that there's this incredibly expensive piece of software that they procure that does summarizations of sophisticated legal text. And the quote emphasizes that what they get out of this expensive tool is incredibly constrained, as if the vendor really bent over backwards to try to create a hard guardrail on how the tool would behave, but it came at the cost of what they really needed, which was the ability to be a little more flexible and iterate.
And so this legal professional quoted reported defaulting to the $20/month ChatGPT subscription that she was paying for out of pocket to use because it got her what she needed.
And I think that's really compelling because from the perspective of the doctrine of the market which is that a bunch of incredibly constrained vertical SaaS or vertical AI platforms where the founders say we're going to be the ChatGPT but just for real estate contract negotiation and then it turns out you go from this report talk to folks who are using technology like that and their answer is yeah they've kind of overrotated on the specialization of the tool And in doing so, they made it brittle against how I really needed to work. And now I'm going to go reach for the more powerful generalist tool.
AM: The report refers to this as the they call it the AI shadow economy. And I thought of something as I was reading about the AI shadow economy, which is that so if you imagine that a large percentage of people at your organization are using ChatGPT below the radar to do their work.
One, it shows that people are agreeable to adopting these types of tools, assuming that they work and actually help and aren't hard to use.
But the second piece that I think is really interesting impacts directly the shareholder value of an organization which is that if you have an existing workforce design and you have your org structure, you have tasks assigned, you have timelines and time budgets for different types of things and you have productivity targets and everything across just standard accounting stuff within organization. If a big percentage of your employees are becoming more productive and finishing their existing work faster or in different ways, but they're doing so essentially untracked and off booked. So, they're using these tools outside of your system.
And this disregards just simply the risk of them putting sensitive or private or HIPAA information just through that. Never mind that. That is a strategic risk.
But from a financial perspective for a company you now have your employees that are becoming essentially more productive and some of that may register but in many cases it probably won't because projects will still be staffed and scheduled assuming that people aren't using these tools because these aren't officially implemented tools within an organization.
And so I started to think, I wonder if this shadow economy really is firms are leaving billions of dollars of shareholder value kind of off the table because it's not being reflected in actual productivity gains or changes in headcount or new roles and responsibilities, upskilling people into higher value added activities etc. And so I think there is actually a risk there for leaders.
RB: And what you're pointing to I think gets to one of the punchlines of the entirety of the report when they interviewed teams that have done really well with their AI rollout and said "what worked for you, if your organization didn't fall into the shadow realm here of shadow AI" and you talk to teams about what worked and how they won, you find that there was a few things that they did, especially teams that to what you pointed out earlier saw a 2x improvement of the success of moving from pilot to full production use case.
And there were teams that won because they did it by partnering with startups or vendors who just specialize in this one piece of the value chain in just building these AI tools. They highlighted four things that jumped out at me that moved this from shadow to light. And that was the emphasis on this needs to happen as federated experimentation. We need to try a lot of experiments fast. We do not do the wait and see. We do not do the doing nothing is safe. Doing nothing is not safe.
They worked with partners closely.
They held those partners accountable to the business outcomes the tool should drive. Not just the model eval metrics but the actual business value outcomes.
And they treated the deployment as an evolutionary process in which they said "look there's going to be stops and starts. We're going to try something. It's not going to work. There will be a failure. That's okay. That's not a showstopper because we know that the transformation potential is here and we're going to keep working on this together."
And so there were winning strategies for teams to move this from "we've got this all off the books, we can't account for this, we don't know how it plays into our productivity, we don't know how to plan around it" to "okay, now we have a way to win."
And I love this because this is buried in this report totally obfuscated by this headline of 95% of all this just goes to waste because this is actually a playbook.
RB: Andrew, you have done a really deep dive on this report. What do you think is the punchline here for what the buyers out there who are trying to figure out how to make heads or tails of this need to know based on what's worked for winners here from who got interviewed?
AM: So the main points or factors that executives and buyers should be looking for when they're thinking about their AI adoption journey is finding AI solutions that are flexible, that are able to be adapted to changing workflows and circumstances within an organization. Tools that are capable of retaining context and learning and improving as they go.
AM: That's a key thing with human employees.
Somebody does something a few times, they usually figure out a better way to do it. AI tools need to have that capacity of not just essentially repeating the rote process over and over. This was called out in the report as well. Not surprisingly, a big factor is around data security and data privacy. Basically we're using organization is using an AI tool and the AI vendor says, "Oh, yeah, your data is safe."
How do you verify that? How does an organization actually know that their data is not being co-mingled with other people's data or not being stored in a safe environment? Another key factor is minimal disruption of current tools. So within especially within a large organization, you have all manner of ERPs and marketing tools and dashboarding tools and workflow managers and project management tools and a million other SaaS products that your employees are currently using to get work done.
AM: And if the AI solution is not able to integrate either with those tools or use those tools and instead it is this separate thing it is bound to fail because ultimately as part of the change management into these tools employees are it's very hard to get employees to learn one additional tool that doesn't dance with the rest of their tools that they're already using. And then the last two very key are that the AI solution and the AI vendor needs to have a deep understanding of workflows. It's kind of the classic consultant faux pas that a young consultant comes into a business and talks with a veteran about their workflow and thinks that they understand it very quickly. And the reality is is that when you're integrating AI tools into complicated workflows that may involve many different people and handoffs and tools and other things, the AI tool needs to be able to the AI team and the AI tools need to achieve a really deep level of integration into those workflows.
AM: And then last, no surprise, is just it has to be a trusted vendor. There's a lot of AI startups out there. There's a lot of incumbents that are offering various AI tools. And ultimately the buyer has to feel confident that they can trust the vendor to deliver on what they promise and then deliver on the other factors that I mentioned.
RB: I love this. So, if I were to cherry-pick admittedly the order of thinking about reflecting on each of those from what we've seen building Mission Control, I'm just going to start with the data security one because it's my favorite joke at least as many organizations are building out their agentic AI strategy and they're turning towards the admittedly the trap of trying to roll their own here. So, they're standing up MCP servers. I'm reminded of the industry joke that the S, the letter S in MCP stands for Secure. Because teams are rapidly realizing, "holy cow, all of our basic data governance rules are going out the window as we've started adopting these technologies because the data we're putting in or granting access to these tools is immediately running counter against what we've expected out of our human operators."
And the teams that are winning here are those that say, "Nope, from the get-go, we're going to be able to implement things like our role-based access control or whitelist blacklist directly onto the agentic AI or the generative AI systems that we're either going to build ourselves or we're going to buy."
And so I'm glad to see that this report is dovetailing with some of the things that we've seen. I think back to some of the experiences we've had around the use of legacy software. Gosh. So, you and I have discussed for one of the teams that are using our technology. Just by way of illustration, how often someone says, "Well, we've got this [insert system] that was very state-of-the-art maybe 20 or 30 years ago that we have all of our data inside and yeah, we're going to be on a data migration path against it to modernize. But the practical reality is we've got a lot of folks both internally and through business process outsourcing that know how to use this mainframe technology. So anything we do needs to be able to fit within our way of working and operate on this mainframe technology."
So we say "great" and then this synthetic worker can go and manipulate against this mainframe technology.
But the use of legacy tooling to your point of the dozen ERPs or dashboards has to happen for anyone to say yes to this. I think about what you've described of the understanding of ways of working from some of the implementations that we've assisted for our partners and them showing up with 40 to 50-page SOP documents that involved a few tools and a dozen people helped build this over a long time so they could give it to someone offshore and have them run it even if they themselves were just learning the ropes of the business.
For organizations that are relatively mature, these SOPs may be in place, but for organizations where you need Steve or Michelle to sit down with someone and show them, so here's really how we do this. And you need to have a deep amount of empathy with how the organization operates and be attuned to their ways of working and not your tool, but rather how they need to get this done.
That's been make or break. And I think all of these things ladder up to that fundamental of trust that you described because if you're going to be in partnership here as look this is not buy it here's the license key and here's the knowledge base good luck but rather this is a new way of doing work. Your people are going to learn how to work alongside it or work directly with it. The tools will increasingly be shaped like colleagues and less like just pure software and we're going to partner with one another on this. That needs to be an earned trust relationship. So everything you're saying dovetails with the things that we see from our partners who then aren't falling in this 95% trap. But there's something you and I discussed which is around one of your philosophies about what this transformation is like and I want to give you last word here.
AM:AI solutions are about getting work done.
AM: It's not about using the tool. The tool is you're using the tool in order to get work done. And so if you have an organization your organization not if your organization is already getting work done through using human processes, paper processes, SaaS processes, reporting process, right? Your organization is already getting work done in a certain way. And so AI needs to be able to come in and needs to be able to assist and accelerate how you're getting work done because the point of AI is not to use AI.
RB: Yeah.
AM: The point of AI is to use AI to get work done. And I think that sometimes gets lost especially in the current hype cycle where everybody is very eager just to have an AI solution whereas again the focus is always on what is the work that needs to get done. But looking out to the future.
So one of the things that I really took away from this report was that we are at really the early stages of this technology that really what this report was talking about was how does an organization get to successful pilots?
Well, before you end up with an AI transformation or an AI enabled organization or you look at a company like MADNA that recently merged its IT and HR departments, which to me seems kind of a signal of what may be to come in terms of how organizations operate and who's working at organizations, right? A mixture of AI and human workers. And what I really took away from this report is that what we're really talking about is it's the transformation from being a non-AI enabled organization or maybe a shadow economy organization to running successful AI pilots to then moving towards full integration and AI transformation that comes later. Organizations that aren't able to make the jump and start launching successful pilots and figuring out how to integrate these solutions into their workflows and things. The organizations that succeed at this and that figure this out are going to be able to rapidly accelerate away. So I think a really good parallel to what's happening is at the start of the last industrial revolution, the 1910s, Ford Motor Company invented the assembly line.
I wrote about this recently. It took them 12 hours to assemble a car before the assembly line. And then a team of people at Ford got together and they invented the moving assembly line. That instead of people going back and forth building the entire car part by part, you had the car move down an assembly line and then you had people at different stations adding parts to the car. And so they were able to go from 12 hours to build a car down to 90 minutes to build a car. And I often think if you were a competitor of Ford at that moment and you see this, orders of magnitude faster, right? Ford can now build cars in 90 minutes instead of 12 hours. And if your response to that was, gee, maybe if we give our employees some power tools, they can build cars slightly faster. And they probably could. So maybe you go from being able to build a car in 12 hours and you get it down to 11 hours or 10 hours and so you can put a few more cars out.
And meanwhile, Ford is able to build many times more cars than you. And I think that within the near to medium future, I really think that's what a lot of organizations are going to be seeing is really the organizations that successfully navigate these AI pilots and then are able to start transitioning into beginning to rethink the workforce for the AI age in the same way that the assembly line rethought the workforce for the industrial age. And there's not exactly a clear picture of what that future workforce of humans and synthetics working together in different ways and how exactly that is going to play out. That is literally being built and is emerging as we speak. And I think this is a really critical moment for a lot of organizations that the wait and see approach may be a big risk because it won't just be from we don't have AI to AI transformation if you miss the step change of implementing successful pilots and understanding how AI can do work within your organization. If you miss that step change, it may be very difficult then when other organizations around you you find are starting to completely redesign their workforces around human synthetic hybrid models and you haven't yet implemented successful pilots around some basic business processes. It's going to be very hard to get caught up culturally, technologically, organizationally. And so that's perhaps the biggest thing I took from the report was really looking forward to where I think this is going and why right now is such a critical time to figure out how to get these pilots.
And we found this originally from a writeup of it done in Fortune with the incredibly evocative and market shaking headline of 95% of money that's spent on generative AI pilots ends up being a massive waste.
And you and I have gone through and read the report and that's not at all what it's actually saying.
Andrew Melville: No, not at all. So the report itself goes in depth about the executives that they talked to and they discussed everything from how are they implementing these pilot programs in what areas of their business are they building them internally or are they using external partners to build them.
AM: So what you find in the report is that there's a whole spread of different reasons why these pilots may be succeeding or failing and you really do have to go into the details.
But when it rolls up to the headline of 95% of AI pilots at large companies fail it makes it sound like the tech is not ready organizations are not ready it makes it sound like the entire thing is about to collapse and when you dig into the details you find that well actually 67% of pilots that are conducted with specialized vendors startups and other vendors those succeed 67% of the time. Whereas internal builds succeed far less. And so yes, the headline itself makes it sound like the hype bubble is ready to burst. And then when you look into the details, to me it makes it seem like we're actually just getting started.
And this is really what you would expect from a new general purpose technology for companies.
RB: I want to go into the paper pull out the key findings that I think we want to know, and that the folks we work with would want to know as well. If I had to sum up the entirety of this paper into anything other than absolute "sky is falling" narrative that drives clicks into what's really here - into a single word - it would be "approach".
And that the approach that the enterprise is taking to generative and agentic AI rollout determines its success.
Which sounds like "yeah, duh approach is going to influence how you do this", but the counter narrative to that would be that actually the technology itself is deeply flawed and no one's seeing any value because the tech's just not there.
And what the paper really keeps emphasizing is that the tech is there. It turns out that organizations are going about implementation and going about their approach in a spectrum of different ways, but a lot of those ways end up being incredibly unproductive.
RB: And so there is a way to do this. There is a way to build internally if you want to go down that path [though the report clearly emphasizes work with a vendor partner to do this and do that kind of specialized approach where you're good at one thing your vendor's good at another] and they highlight some takeaways here.
I want to explore these with you, but I want to kind of get 10,000 foot view first.
What jumped out at you? Because you read the thing front to cover many times. What jumped out at you about it?
AM: So, what jumped out to me was that first off, it's that the areas of a business where generative AI tools would be the most effective at driving actual P&L value: back office operations, business process outsourcing, some of these types of very not flashy but fundamental parts of a business. Those areas were not what generative AI pilots were run on.
They were typically run on sales, marketing department, sort of customer facing and front-facing use cases. Things like chatbots and various things helping the sales team respond faster.
There's a quote in the article from a CIO that says, "well, aside from it processing some invoices a little bit faster, we haven't really seen this have much of an effect in our business".
And so back to your statement about approach. So right out the gate, right, the technology is being applied to the wrong areas of a business most of the time or rather to lower value areas of a business. And then when those projects don't deliver some sort of measurable value, very quickly they point to, well, the technology is not ready.
Well, in reality based on the data, it sounds like it is being directed at the wrong use cases.
RB: So I go back to this all the time, this metaphor of this paper that was really fundamental in AI itself, at least here in the Bay around the "bitter lesson". And it's this metaphor about scale: just throw more parameters and more training data and more energy usage at the problem. Even if you've got a general purpose deep learning transformer architecture neural network that's not specialized to anyone particular task, scale alone will beat really sophisticated algorithmic approaches and exquisite systems.
And they call this the bitter lesson because it angers a lot of people who have been in AI for a really long time that this generalist "just make the model massive" approach actually wins. In fact, it wins nearly across the board. And they call this the "bitter lesson".
But I'm reminded that there's about a dozen other bitter lessons embedded in the actual practice in a business of transforming using AI.
And this feels like one of them. And that is that someone comes back from a conference or from some sort of summit or they were on a board meeting and then someone panics saying "we got to have an AI strategy". We got to do something here. And so they end up paying 50 million bucks for a bad PowerPoint from one of the Big Three that suggests a really mediocre idea that this flashy frontline topline front of house implementation that then falls apart and it's a bitter lesson of where would value have actually been created.
"I don't know, did you try month-end close? Did you try inventory management? Did you try ticket resolution? Did you try the actual business operations that drive your firm?"
And that's what this report captures: those places buried deeper that you used to give to BPO or agencies, those actually are hot beds of ROI compared to what everybody's doing.
And that's the bitter lesson.
AM: And I think you have to sympathize in the sense of the current news cycle around AI and so it is understandable that in an effort to say that your company is AI-capable, AI-enabled, etc. It's not a very exciting story to go in front of the board or to go in front of the camera and say "we've just digitally onshored our IT ticketing operations that used to be international and we've now brought that on board we're now using AI to do that and it's saving us x amount of money and the turnaround time has decreased by that" and it's some function that is not very interesting it's much more exciting to say "we've given it to our sales team our marketing group" and then it's not able to deliver the results.
RB: Yeah, this feels like if we're lucky - for listeners who are familiar, probably everybody, with the idea of the Gartner hype cycle - this feels like if we're lucky, this is one of the milestones on the plateau to productivity of "oh, so it turns out those incredibly quotidian tasks your org actually just runs on that are in the guts of the organization. Turns out to be a really great place for sufficiently advanced autonomous automation" as opposed to these incredibly flashy frontline use cases. So I appreciate your empathy.
So if we dig in beyond approach, there is something that they call out in the report that I do think is worth discussing and that is that there are two major flavors of requirement that the enterprise has that isn't getting met: and that is that 1) a lot of the technology they're buying from vendors is brittle. Can't be customized or modified into their ways of working. and 2) that the technologies don't learn as they go in and that you are constantly having to refresh gobs and gobs of context and the model isn't customizing to their own behaviors over time.
And those two major requirements are, according to the report, where the teams that are succeeding in rolling-out things that can learn or are flexible enough to meet the workflow requirements of actual ways of working. Those are where teams are winning.
And then the rest of this outside of those two narrow things, implementation, it's where people are wanting this to make sense.
AM: One of the things that really jumped out was that employees within organizations don't experience a pervasive fear about jobs being automated away by AI, as I think a lot of people suspect. They actually found that a lot of organizations have very high adoption rates of AI already within the organization but it is being done essentially offbook.
It is individual employees often using their own private accounts to use ChatGPT or Claude and to use it to draft emails or write code or do whatever they're able to do. So, it's limiting it to fairly basic tasks.
Again, you don't have memory storage. You don't have the ability to construct complicated agents that are able to actually do work. And so employees are clearly very interested in adopting these technologies. And the report actually calls out that they often find that implementations of enterprise AI tools actually aren't as good or good enough as just using ChatGPT. And so they essentially say, well, this isn't as good or better than ChatGPT, so I'm just going to continue using ChatGPT. And they just reject the custom-built tool.
RB: Which is something that I'm glad you pointed that out because they've got a pullquote in there from a legal team that suggested that there's this incredibly expensive piece of software that they procure that does summarizations of sophisticated legal text. And the quote emphasizes that what they get out of this expensive tool is incredibly constrained, as if the vendor really bent over backwards to try to create a hard guardrail on how the tool would behave, but it came at the cost of what they really needed, which was the ability to be a little more flexible and iterate.
And so this legal professional quoted reported defaulting to the $20/month ChatGPT subscription that she was paying for out of pocket to use because it got her what she needed.
And I think that's really compelling because from the perspective of the doctrine of the market which is that a bunch of incredibly constrained vertical SaaS or vertical AI platforms where the founders say we're going to be the ChatGPT but just for real estate contract negotiation and then it turns out you go from this report talk to folks who are using technology like that and their answer is yeah they've kind of overrotated on the specialization of the tool And in doing so, they made it brittle against how I really needed to work. And now I'm going to go reach for the more powerful generalist tool.
AM: The report refers to this as the they call it the AI shadow economy. And I thought of something as I was reading about the AI shadow economy, which is that so if you imagine that a large percentage of people at your organization are using ChatGPT below the radar to do their work.
One, it shows that people are agreeable to adopting these types of tools, assuming that they work and actually help and aren't hard to use.
But the second piece that I think is really interesting impacts directly the shareholder value of an organization which is that if you have an existing workforce design and you have your org structure, you have tasks assigned, you have timelines and time budgets for different types of things and you have productivity targets and everything across just standard accounting stuff within organization. If a big percentage of your employees are becoming more productive and finishing their existing work faster or in different ways, but they're doing so essentially untracked and off booked. So, they're using these tools outside of your system.
And this disregards just simply the risk of them putting sensitive or private or HIPAA information just through that. Never mind that. That is a strategic risk.
But from a financial perspective for a company you now have your employees that are becoming essentially more productive and some of that may register but in many cases it probably won't because projects will still be staffed and scheduled assuming that people aren't using these tools because these aren't officially implemented tools within an organization.
And so I started to think, I wonder if this shadow economy really is firms are leaving billions of dollars of shareholder value kind of off the table because it's not being reflected in actual productivity gains or changes in headcount or new roles and responsibilities, upskilling people into higher value added activities etc. And so I think there is actually a risk there for leaders.
RB: And what you're pointing to I think gets to one of the punchlines of the entirety of the report when they interviewed teams that have done really well with their AI rollout and said "what worked for you, if your organization didn't fall into the shadow realm here of shadow AI" and you talk to teams about what worked and how they won, you find that there was a few things that they did, especially teams that to what you pointed out earlier saw a 2x improvement of the success of moving from pilot to full production use case.
And there were teams that won because they did it by partnering with startups or vendors who just specialize in this one piece of the value chain in just building these AI tools. They highlighted four things that jumped out at me that moved this from shadow to light. And that was the emphasis on this needs to happen as federated experimentation. We need to try a lot of experiments fast. We do not do the wait and see. We do not do the doing nothing is safe. Doing nothing is not safe.
They worked with partners closely.
They held those partners accountable to the business outcomes the tool should drive. Not just the model eval metrics but the actual business value outcomes.
And they treated the deployment as an evolutionary process in which they said "look there's going to be stops and starts. We're going to try something. It's not going to work. There will be a failure. That's okay. That's not a showstopper because we know that the transformation potential is here and we're going to keep working on this together."
And so there were winning strategies for teams to move this from "we've got this all off the books, we can't account for this, we don't know how it plays into our productivity, we don't know how to plan around it" to "okay, now we have a way to win."
And I love this because this is buried in this report totally obfuscated by this headline of 95% of all this just goes to waste because this is actually a playbook.
RB: Andrew, you have done a really deep dive on this report. What do you think is the punchline here for what the buyers out there who are trying to figure out how to make heads or tails of this need to know based on what's worked for winners here from who got interviewed?
AM: So the main points or factors that executives and buyers should be looking for when they're thinking about their AI adoption journey is finding AI solutions that are flexible, that are able to be adapted to changing workflows and circumstances within an organization. Tools that are capable of retaining context and learning and improving as they go.
AM: That's a key thing with human employees.
Somebody does something a few times, they usually figure out a better way to do it. AI tools need to have that capacity of not just essentially repeating the rote process over and over. This was called out in the report as well. Not surprisingly, a big factor is around data security and data privacy. Basically we're using organization is using an AI tool and the AI vendor says, "Oh, yeah, your data is safe."
How do you verify that? How does an organization actually know that their data is not being co-mingled with other people's data or not being stored in a safe environment? Another key factor is minimal disruption of current tools. So within especially within a large organization, you have all manner of ERPs and marketing tools and dashboarding tools and workflow managers and project management tools and a million other SaaS products that your employees are currently using to get work done.
AM: And if the AI solution is not able to integrate either with those tools or use those tools and instead it is this separate thing it is bound to fail because ultimately as part of the change management into these tools employees are it's very hard to get employees to learn one additional tool that doesn't dance with the rest of their tools that they're already using. And then the last two very key are that the AI solution and the AI vendor needs to have a deep understanding of workflows. It's kind of the classic consultant faux pas that a young consultant comes into a business and talks with a veteran about their workflow and thinks that they understand it very quickly. And the reality is is that when you're integrating AI tools into complicated workflows that may involve many different people and handoffs and tools and other things, the AI tool needs to be able to the AI team and the AI tools need to achieve a really deep level of integration into those workflows.
AM: And then last, no surprise, is just it has to be a trusted vendor. There's a lot of AI startups out there. There's a lot of incumbents that are offering various AI tools. And ultimately the buyer has to feel confident that they can trust the vendor to deliver on what they promise and then deliver on the other factors that I mentioned.
RB: I love this. So, if I were to cherry-pick admittedly the order of thinking about reflecting on each of those from what we've seen building Mission Control, I'm just going to start with the data security one because it's my favorite joke at least as many organizations are building out their agentic AI strategy and they're turning towards the admittedly the trap of trying to roll their own here. So, they're standing up MCP servers. I'm reminded of the industry joke that the S, the letter S in MCP stands for Secure. Because teams are rapidly realizing, "holy cow, all of our basic data governance rules are going out the window as we've started adopting these technologies because the data we're putting in or granting access to these tools is immediately running counter against what we've expected out of our human operators."
And the teams that are winning here are those that say, "Nope, from the get-go, we're going to be able to implement things like our role-based access control or whitelist blacklist directly onto the agentic AI or the generative AI systems that we're either going to build ourselves or we're going to buy."
And so I'm glad to see that this report is dovetailing with some of the things that we've seen. I think back to some of the experiences we've had around the use of legacy software. Gosh. So, you and I have discussed for one of the teams that are using our technology. Just by way of illustration, how often someone says, "Well, we've got this [insert system] that was very state-of-the-art maybe 20 or 30 years ago that we have all of our data inside and yeah, we're going to be on a data migration path against it to modernize. But the practical reality is we've got a lot of folks both internally and through business process outsourcing that know how to use this mainframe technology. So anything we do needs to be able to fit within our way of working and operate on this mainframe technology."
So we say "great" and then this synthetic worker can go and manipulate against this mainframe technology.
But the use of legacy tooling to your point of the dozen ERPs or dashboards has to happen for anyone to say yes to this. I think about what you've described of the understanding of ways of working from some of the implementations that we've assisted for our partners and them showing up with 40 to 50-page SOP documents that involved a few tools and a dozen people helped build this over a long time so they could give it to someone offshore and have them run it even if they themselves were just learning the ropes of the business.
For organizations that are relatively mature, these SOPs may be in place, but for organizations where you need Steve or Michelle to sit down with someone and show them, so here's really how we do this. And you need to have a deep amount of empathy with how the organization operates and be attuned to their ways of working and not your tool, but rather how they need to get this done.
That's been make or break. And I think all of these things ladder up to that fundamental of trust that you described because if you're going to be in partnership here as look this is not buy it here's the license key and here's the knowledge base good luck but rather this is a new way of doing work. Your people are going to learn how to work alongside it or work directly with it. The tools will increasingly be shaped like colleagues and less like just pure software and we're going to partner with one another on this. That needs to be an earned trust relationship. So everything you're saying dovetails with the things that we see from our partners who then aren't falling in this 95% trap. But there's something you and I discussed which is around one of your philosophies about what this transformation is like and I want to give you last word here.
AM:AI solutions are about getting work done.
AM: It's not about using the tool. The tool is you're using the tool in order to get work done. And so if you have an organization your organization not if your organization is already getting work done through using human processes, paper processes, SaaS processes, reporting process, right? Your organization is already getting work done in a certain way. And so AI needs to be able to come in and needs to be able to assist and accelerate how you're getting work done because the point of AI is not to use AI.
RB: Yeah.
AM: The point of AI is to use AI to get work done. And I think that sometimes gets lost especially in the current hype cycle where everybody is very eager just to have an AI solution whereas again the focus is always on what is the work that needs to get done. But looking out to the future.
So one of the things that I really took away from this report was that we are at really the early stages of this technology that really what this report was talking about was how does an organization get to successful pilots?
Well, before you end up with an AI transformation or an AI enabled organization or you look at a company like MADNA that recently merged its IT and HR departments, which to me seems kind of a signal of what may be to come in terms of how organizations operate and who's working at organizations, right? A mixture of AI and human workers. And what I really took away from this report is that what we're really talking about is it's the transformation from being a non-AI enabled organization or maybe a shadow economy organization to running successful AI pilots to then moving towards full integration and AI transformation that comes later. Organizations that aren't able to make the jump and start launching successful pilots and figuring out how to integrate these solutions into their workflows and things. The organizations that succeed at this and that figure this out are going to be able to rapidly accelerate away. So I think a really good parallel to what's happening is at the start of the last industrial revolution, the 1910s, Ford Motor Company invented the assembly line.
I wrote about this recently. It took them 12 hours to assemble a car before the assembly line. And then a team of people at Ford got together and they invented the moving assembly line. That instead of people going back and forth building the entire car part by part, you had the car move down an assembly line and then you had people at different stations adding parts to the car. And so they were able to go from 12 hours to build a car down to 90 minutes to build a car. And I often think if you were a competitor of Ford at that moment and you see this, orders of magnitude faster, right? Ford can now build cars in 90 minutes instead of 12 hours. And if your response to that was, gee, maybe if we give our employees some power tools, they can build cars slightly faster. And they probably could. So maybe you go from being able to build a car in 12 hours and you get it down to 11 hours or 10 hours and so you can put a few more cars out.
And meanwhile, Ford is able to build many times more cars than you. And I think that within the near to medium future, I really think that's what a lot of organizations are going to be seeing is really the organizations that successfully navigate these AI pilots and then are able to start transitioning into beginning to rethink the workforce for the AI age in the same way that the assembly line rethought the workforce for the industrial age. And there's not exactly a clear picture of what that future workforce of humans and synthetics working together in different ways and how exactly that is going to play out. That is literally being built and is emerging as we speak. And I think this is a really critical moment for a lot of organizations that the wait and see approach may be a big risk because it won't just be from we don't have AI to AI transformation if you miss the step change of implementing successful pilots and understanding how AI can do work within your organization. If you miss that step change, it may be very difficult then when other organizations around you you find are starting to completely redesign their workforces around human synthetic hybrid models and you haven't yet implemented successful pilots around some basic business processes. It's going to be very hard to get caught up culturally, technologically, organizationally. And so that's perhaps the biggest thing I took from the report was really looking forward to where I think this is going and why right now is such a critical time to figure out how to get these pilots.