Lightly edited transcript of the video’s captions: speakers identified, names and terms corrected, filler, false starts and crosstalk trimmed. Timestamps mark where each passage starts in the video.

[0:00] Hugh Abbott (HA): Hey, Amie, how are you doing?

[0:02] Amie Devero (AD): I’m okay. After all our technical problems, I finally got my headset on, so I can look like a 1920s aviator. I’m ready to go.

[0:10] HA: Great. Amie, I wanted to talk to you because I’m looking for great uses of AI, great uses of large language models, and I think what you are producing is very interesting and different. I think it is a business process, or a product, which would not have been possible had it not been for the invention of large language models. So perhaps you can tell us about it, and how you had the insight to think: this is a business need which can be served by a tool that uses large language models.

[0:53] AD: Well, I actually want to take us back to a more macro level, because there are two things I think are momentous about what I’m doing. The first is that I’m building a software product, and I’m not an engineer. I’ve worked with startups and technology companies for many years, and I’ve led two of them, but apart from a few classes in JavaScript, I am not a coder. So the fact that somebody like me can build a fully operational enterprise product is miraculous.

[1:26] HA: So you’re using vibe coding, or whatever they call it?

[1:30] AD: Exactly. I am vibe coding this on Base44, which means I don’t need DevOps, I don’t need AWS, I don’t need an engineer. At some point I may pull it off that platform because it becomes too complex, but it is extraordinary to be able to build, test and deploy this thing entirely using my own creativity. I know some technical things, so I can help the AI find some problems, but fundamentally somebody else is coding this, and that somebody else is an AI.

[2:04] HA: Yeah, that is incredibly liberating. You’ve had the business insight, saying here’s a need, I can deliver it once I’ve got the large language models, but how can I actually create the tool?

[2:16] AD: Right.

[2:16] HA: That in itself is being made possible by the existence of brilliant large language models. So I presume you can choose your model within Base44?

[2:28] AD: I built this from scratch. This is completely my idea, and I have built every portion of it. It’s just that I am giving instructions to an AI that is executing on my behalf. I can send you a login, and you can log into the product and use it, and nothing in your experience would tell you it’s any different from any other product you log into, like the one we’re recording this on, or Google Mail, or any of a million cloud-based products you use every day.

[3:01] AD: And this is not the only app I’m building. I’m also building a gardening app, which I’ve talked to a bunch of my friends about, including our mutual friend Chris. So I am really astounded by the degree to which vibe coding, and in particular these platforms that really do make it possible for lay people to create products, have freed people like me. I’ve had this repository of ideas for years, so many startup ideas that were all technological, that I could not execute on because I didn’t have seed funding or an engineer in my pocket. So I just want to say that, because I think what I’m building is a very interesting product, but that I can do it at all is really what’s earth-shattering.

[3:50] HA: Can I just ask, the tool you use is Base?

[3:53] AD: It’s called Base44.

[3:55] HA: And within Base44, can you choose the model that you decide to code with? Or does it come with a standard model?

[4:02] AD: It’s using multiple models.

[4:05] HA: Oh, that’s interesting.

[4:06] AD: So I’m not choosing which LLM runs in the background. It’s usually using ChatGPT, but you can also have it use Claude; it’s got several. I think it’s mostly using ChatGPT because we had a little glitch when they did that last upgrade, but I’m not involved in that side, if you know what I mean. Although I am helping the Base44 AI design prompts for the ChatGPT AI. Within my product there are many, many prompts that the AI is writing on my behalf, and when things go wrong, it’s usually because the AI didn’t write a good enough prompt. So then I have to get involved and say, “Wait, you need to instruct it to do this and to do that and to make these distinctions,” and so forth.

[4:50] AD: So that’s in the background. In the foreground, what this product is: this is not a new need. It’s just newly able to be addressed by something other than human beings. What this product does is something I’ve essentially been working on my entire career as a strategy consultant. When an organization creates a strategy, the CEO or the board of directors or the senior team has a problem: now they have to get everybody in the organization working with that strategy as the context. There are a lot of very traditional, recognizable ways we do that. We go on a road show, do town halls, send out emails and PowerPoints, train the managers to train their employees, or whatever.

[5:36] AD: But all of those processes have several problems. One is that they are largely one-directional. It’s the CEO or the VP delivering information to workers, usually virtually, who have many different levels of education, expertise and functional knowledge, or lack thereof. And there’s no practical way to determine, or even to help, employees use that strategy to make decisions. You just put it out there, teach it to them as best you can, and hope that they internalize it and use it to make decisions. And mostly they don’t, because mostly we make decisions in our day-to-day jobs based on expediency, or metrics, or who’s hustling us the most, or some kind of power differential. We just need to do something because someone told us to, or someone’s bugging us more.

[6:32] AD: But you really want your employees to make decisions based on the strategy, on the strategic principles and the strategic goals, because employees constantly face choices, even at the most basic level. Think about somebody at a call center deciding how long to stay on a call to solve someone’s problem, with what demeanor, with what degree of care and attention, or whether to escalate it to a manager or offer a refund.

[7:04] AD: And we’ve typically used metrics to do this. Metrics are a tool, but they’re a very clumsy tool, because all they can do is incentivize one kind of outcome over another. They don’t determine the behavior that generates that outcome. Everybody knows the story of Wells Fargo in the US, which used metrics to force its employees to generate greater and greater numbers of accounts. Employees were under so much pressure that they created fictional accounts, ten accounts for customers who knew nothing about them. It was huge. It was fraud. Most metrics aren’t that bad, but you can see it’s pushing for an outcome, not for a behavior. What you really want is for your employees to behave and make decisions in ways that help you fulfill the strategy.

[8:03] AD: And when you do it the old-fashioned manual way, you do all this one-way delivery. The other problem is that it’s usually a finite thing. You deliver the messages, you do the road show, you do the town hall, and after a month or six months, it ends.

[8:19] HA: Yeah. You deliver it once, you’ve delivered the strategy, you’re happy.

[8:23] AD: And then you do it again in five years, or annually, or some incremental thing like that. It’s not an ongoing experience in your life.

[8:30] AD: So what this tool does is allow a leader, founder, CEO or whoever to provide all of their documents to the AI, and then it walks them through a dialogue to ingest the strategy. The dialogue is extensive; it takes about 45 minutes to complete. It uses your documents to populate its information, and then it asks you questions about your strategic tradeoffs, your position in the marketplace, everything about your strategy and your theory of how you’re going to win in the market.

[9:09] HA: So it’s got a complete understanding of your strategy, from the documents and from a conversation at the moment of ingestion.

[9:16] AD: Much more than a conversation. It is a very complex, structured dialogue, up to and including an entire section where it gives you difficult strategic dilemmas and asks what you choose and why. Say you have a vendor that is about to raise their prices a great deal. You have customers waiting for the product who have already agreed to a price. There’s another prospective vendor, but switching is going to delay delivery. What do you do? Do you stick with your vendor and eat the losses? Do you raise the prices on the customers? Do you switch vendors and delay the delivery? Those are real choices; that’s how life really is. You give your answer and your explanation, and the AI adds those principles, those examples, the way you calibrate priorities, into the strategic schema.

[10:13] AD: So it builds an entire schema. Then you can deploy that to your employees, and your employees get taught the strategy in an engaging, interactive way that is specific to their job. And then the AI also asks them similar kinds of dilemma questions.

[10:34] HA: So you’re teaching via conversation.

[10:38] AD: It is conversational, but it’s not open-ended chatting. It’s a combination of a form and questions and answers. I know we think of LLMs like that, but this is much more structured, because remember, this is uploading the strategy and teaching the strategy.

[11:06] AD: Once an employee has gone through the process of learning the strategy and being able to make decisions that are aligned with it, there’s a second feature that allows the employee to use the tool to make decisions on the fly, on an ongoing basis. So they can go to the tool and say…

[11:27] HA: You query the tool about a real problem.

[11:30] AD: Only after you’ve done all that other setup, so that it has the background and the CEO has gotten the strategy right. Then it acts as a sort of strategy ombudsman for decision-making in the moment.

[11:46] HA: So let’s wind back a bit. We are using LLMs and AI to create the tool; that’s an amazing conversation in and of itself, so we’ll park that. Within your tool, we’re using LLMs to completely ingest all the documents and, through some kind of interaction, ingest the whole of the strategy. Then we can use that in some interactive way, a sort of guided learning, which was not previously available, because we couldn’t interact in such an unstructured way.

[12:25] AD: Well, because it’s not algorithmic; it’s intelligent. It’s not that the AI has an algorithm for what the right answers are. I’m creating the model, but the AI is populating the educational model not just with the actual words from the strategy, but with the inferences it has drawn from its understanding of the principles. So it invents dilemmas for the…

[12:54] HA: Yeah. So it’s a far more sophisticated method of teaching, or learning, than was previously available. And then finally, the third thing you have is that you can use it to ask a real-world question: this is my dilemma, you’ve got all the knowledge, can you help me? That seems to me very much like the way some people are using AI to create the first draft of an email or a document. You ask it, and it might give you the answer, but it’s the first draft of the answer. You say, “Oh, yeah, I need to think about this.” And it’s so useful to have something you just ask that gives you the first draft.

[13:42] AD: But this doesn’t do quite that. It acts more like a Socratic interlocutor. Say you tell it, “I’ve got this problem. I need to get this project done, but now we have five bugs that came in from big customers. Do I have my team stop working on the next iteration of the roadmap, or do we handle these bugs? I’m under a lot of pressure. My boss is out of town. How do I make this decision?” And the AI will say, “Okay, let’s look at the specific strategic principles at play here. The customer needs this, we have the deadline for the release,” and it will walk you through that, because there aren’t right answers in these situations. There are only better and worse answers. Sometimes you have two not-so-great answers and you just have to call it. And the AI can help an employee call it in a way that’s consistent with the strategy rather than inconsistent with it.

[14:38] AD: And then it has a third tier, which I think is the most useful. As your employees are using this tool, the AI is amassing a body of knowledge about what their problems are, what employees are reckoning with in the world that is germane to the strategy. Are they constantly getting complaints from customers about some aspect of the product that isn’t broken but doesn’t fit their needs, and that needs to be added to the product roadmap? Are they constantly having problems with deployment because one part of the product is too difficult to use? Employees come back to the AI with these dilemmas, and the AI can then create reports for the leadership saying, “Here are problems that may be indicative of inconsistencies or suboptimization in your strategy. Maybe you want to look at this or that.”

[15:33] HA: So that information was previously unrecorded, or maybe it was recorded, but in unstructured ways that were difficult to interrogate or build into the rest of the management. You can record it, extract meaning and deliver it back up to the management in a…

[15:54] AD: Right.

[15:54] HA: …a wonderful feedback loop.

[15:55] AD: Well, the thing is, in an organization it’s not that it isn’t recorded; it’s that it lives as a complaint. Everybody on the sales team is nattering and gossiping together and complaining about something that is fundamentally a strategic problem. But because it gets heard as complaining, it doesn’t get distilled into a genuine piece of feedback that ought to inform an iteration of the strategy. This tool takes all those problems and turns them into useful feedback. And I hope, though I don’t know if this is true, it is entirely a hypothesis, that if that virtuous cycle really kicks into gear, it makes the environment of the organization less toxic, because persistent problems that are strategic get dealt with as such rather than…

[16:52] HA: There’s a pressure valve. You can let the pressure off: you can say something, and it is being listened to. And you’re absolutely right. I’m sure a lot of these complaints are recorded, but then there’s a big document with a whole lot of stuff in it, and it’s quite difficult to ingest, master and make actionable. A tool like yours can understand it and feed it back in a virtuous cycle, in a way that somebody can do something about it.

[17:26] AD: Right. So the CEO has a dashboard, and when he goes to his strategy dashboard, it says, “Three issues you should consider,” gives the issues and shows how each one is showing up. First of all, how do I help employees with this? And second, maybe the strategy needs some tweaking. Maybe we change the roadmap. Maybe the product needs this. Maybe we shouldn’t be going into that secondary market right now; we’re getting so much pushback. This is information that’s really valuable, and it usually comes way later, on the fat tail of unintended consequences, instead of in a programmatic way, when you need it and can do something to fix it.

[18:04] HA: Yeah. And I suppose that comes back around to your first insight: companies do strategy, they do it in a year, they deliver it, they forget about it, and they come back in a couple of years’ time and say, maybe our strategy needs updating, but they haven’t done anything in the meantime.

[18:26] AD: It’s not that they haven’t done anything, but it’s like you throw the strategy out there, and five years later you come back, look at the results and ask, did that work? Did that not work? In the middle, you’re just assuming the strategy was fully executed. And of course the projects might have been executed, but with how much noise along the way?

[18:48] HA: It’s a bit like sending your boat off into the sea and expecting that the direction you push it in will be the direction it ends up sailing in, no matter what waves and storms come to knock it off course. So obviously tweaking your strategy, because stuff happens, is what you should be doing.

[19:11] AD: And I don’t want to make it sound like we’re doing nothing right now. Obviously leaders tweak their strategy when the environment changes, or when they get staggeringly bad feedback, like they can’t sell something. Metrics do inform strategy. This doesn’t replace metrics, but it also doesn’t deploy or use metrics. This is really about the qualitative alignment of an organization with the strategy rather than the quantitative alignment. The quantitative alignment matters, and I don’t want to sound like I’m suggesting it doesn’t. But it is a different thing from having people who have internalized the strategy and have it as their context for decision-making, because they have a tool to do that consistently. I think you can get away with just the metrics, but the possibility of holistic strategy alignment is so much richer if you also have this qualitative alignment piece.

[20:11] HA: Well, I think it’s a brilliant use of a large language model, because, as you say, a service such as this would not have been possible four years ago. But obviously you’re close to the market, you understand what the needs of the market are, and you thought, actually, this could be delivered.

[20:34] AD: We’ll see if I’m right about that. It’s all up for grabs whether anybody really wants to buy it. But I am looking for beta testers. So if anybody who’s listening has a company or has a strategy and thinks this tool might be of interest to them, I would love to have more beta testers who would be willing to upload their strategy docs, use the tool and give me feedback so that I can improve it.

[21:03] HA: I think what you’ve done is fantastic and inspirational, because I just love the fact that, as you said, you’re not a coder.

[21:09] AD: No, not at all.

[21:10] HA: And you knew there was this need, and you thought, actually, large language models can help address it. But just the vibe coding step, throwing yourself in and thinking, I think I can do it, having the courage to do it: I think it’s brilliant. I love it.

[21:27] AD: Well, I appreciate that, Hugh. I’m very grateful for that. I hope you’re right that it turns out to be useful.

[21:35] HA: In some respects it reminds me of the liberating things about technology in my lifetime. One of them, I think, is the rise of YouTube and other platforms for learning. That was an unintended consequence. Nobody said to me 20 years ago that YouTube was going to be great for learning, but actually you can learn anything. And I love Coursera as a platform where you can really learn things; it just takes your energy and endeavor to access it and do some learning, which I think is very liberating. In some parallel way, this vibe coding is also very liberating. You’ve got an idea for a product or a service, and you can go and create it.

[22:23] AD: I really do feel like I have been given a second life, in a certain kind of way, because I have been a frustrated product person my whole life. Not any kind of product; I’m not an inventor of devices. But I have always had all these ideas for how to deliver something using digital technology, and I’ve always been frustrated, just thwarted in any attempt to get anywhere with it. So it’s extraordinary to be able to disgorge my brain into this tool and have it create something that actually is my idea, manifest. It’s amazing.

[23:10] HA: That I think is great. I wasn’t looking for that. I was looking for great uses of AI and LLMs, but to describe it as the gift of a second life…

[23:21] AD: A second professional life, I mean. I don’t mean a second wife.

[23:26] HA: But I think that’s the order of magnitude of what you’re saying. It’s so liberating it’s given you a second life, and that’s brilliant. Well, thank you very much, Amie. That’s been great, and I wish you much luck with your second life.

[23:42] AD: Thank you very much, and I appreciate you taking the time to chat with me.

[23:47] HA: Right, thanks a lot. Bye-bye.