For the past two years, the enterprise AI conversation has largely revolved around one thing: bigger models. Organizations raced to integrate the latest frontier models into everything from customer support to internal operations, believing that more capable AI would naturally lead to better business outcomes.
Now, as those pilots transition into production, a different challenge is emerging. Enterprises are discovering that deploying AI at scale isn't simply about choosing the most powerful model. It's about building systems that remain reliable, economically sustainable, and under the organization's control.
Few founders have watched that shift as closely as David Villalón, the co-founder and CEO of Maisa. After working with large language models long before generative AI entered the mainstream, Villalón built Maisa around a different vision for enterprise AI: one where accountability, governance, and operational efficiency matter just as much as raw intelligence.
In this interview with TechNews180, Villalón explains why AI costs are often misunderstood, why enterprises are over-relying on frontier models, and why the next competitive advantage won't come from accessing smarter models, but from building AI architectures that businesses can truly own.
1. The conversation around AI costs often focuses on token pricing. You argue that's the wrong framing altogether. If a CEO is worried about an exploding AI bill, what question should they be asking instead?
The first question shouldn't be "How do I reduce my AI bill?" It should be "What business outcome am I actually trying to achieve?" Because in most cases, the issue isn't the price of the model at all. What we're seeing is that companies have quietly adopted a strategy where every new workflow gets assigned to the biggest model available, regardless of whether that level of intelligence is actually required. That approach makes perfect sense during the pilot phase because the objective is simply to prove that AI can perform the task, and nobody is paying much attention to economics yet. The problem comes later, when those same workflows move into production and organizations suddenly find themselves paying frontier-model prices for tasks that could have been handled just as effectively by a specialized small model designed for that exact use case. That's why I think the conversation around AI costs is often framed incorrectly. The companies that succeed won't be the ones negotiating the lowest token prices. They'll be the ones that become disciplined about matching the right model to the right task in order to achieve the business outcome they actually care about.
2. Many enterprises seem to be treating frontier models as a universal solution, applying them to everything from customer support to internal workflows. Why do you think so many companies have defaulted to this approach, and what risks are they only beginning to discover?
I think companies defaulted to frontier models because, for a while, it was actually the right decision. When ChatGPT and the first wave of powerful models arrived, the priority wasn't efficiency, it was capability. Everyone wanted to know what AI could do, so naturally they reached for the most capable systems available. What many organizations underestimate is that a model performing well in a controlled environment doesn't necessarily mean it will perform reliably once it's embedded into a real business process. That's where the real challenges begin. Hallucinations become much more costly, scalability becomes a concern, costs start growing faster than expected, and dependency on external providers increases. Suddenly, decisions about pricing, availability, compliance, or even whether a model continues to exist are being made by someone else. Getting AI to work in a controlled environment was never the hardest part. The real challenge is making it reliable, scalable, and economically sustainable when it's executing thousands or even millions of business-critical tasks.
3. We've seen previous technology waves where infrastructure became more important than the application itself. Are we reaching a similar moment in AI, where architecture matters more than model selection?
I think we're already seeing that shift. Two years ago, the main question was whether a model was capable enough to perform a particular task. Today, most frontier models are remarkably capable, and organizations can access them with just a few API calls. That means the conversation is becoming less about intelligence itself and more about how that intelligence is orchestrated inside real operational workflows. The real challenge isn't choosing between Model A or Model B, it's deciding which model is appropriate for each step of a workflow, making sure it produces reliable outputs, minimizing hallucinations, and ensuring the system can repeat that performance consistently at scale. That's where I think the industry is heading. Models will continue to improve and become increasingly interchangeable, but the organizations that create the most value will be the ones that know how to orchestrate those models into systems that remain reliable, controllable, and economically sustainable over time.
4. For years, enterprises assumed AI would become cheaper as models improved. Why do you think many organizations are now finding the opposite once they move into production?
I think people often confuse the cost of accessing AI with the cost of running AI in production. The assumption was that as models improved and competition increased, AI would naturally become cheaper. In reality, the opposite has happened for many enterprise deployments. Today's models are significantly more capable, but they're also reasoning through much more complex problems before producing an answer, and that additional intelligence comes at a cost. On top of that, companies are no longer making a single call to a model. They're building complete AI systems with multiple models, orchestration layers, validation steps, and business logic working together to deliver a reliable outcome. As those systems become more sophisticated, their dependency on AI grows, and so does the cost of operating them. That's why I think the conversation has to move beyond model pricing. The real question isn't whether individual models are becoming cheaper, but whether the overall system has been designed in a way that remains economically sustainable as it scales.
5. Many organizations are trying to balance performance, cost, and control. Is the future of enterprise AI likely to be built around multiple models working together rather than a single frontier model? What would that shift mean for how companies design AI systems?
I think that's exactly where the industry is heading. One of the biggest concerns enterprises are starting to recognize is that if all of their business logic lives inside third-party models, they gradually lose ownership of one of their most valuable assets: their own institutional knowledge. At the same time, they become increasingly dependent on decisions they don't control, whether that's pricing, model availability, compliance policies, or even whether a model continues to exist. That's why I don't think the future is about finding a single "best" model. Businesses don't operate as single tasks; they operate as systems made up of dozens of decisions, validations, and workflows. It doesn't make much sense to use the same frontier model for every one of those steps any more than it would make sense to hire the world's best lawyer to answer every email in your company. Instead, we're moving toward architectures where different models play different roles depending on the complexity of the task, the level of risk involved, and the cost constraints of the business. Some processes genuinely require frontier intelligence. Others can be handled by specialized models designed for that exact use case, allowing organizations to retain greater control over their own knowledge while reducing cost and dependency. Ultimately, I think the companies that win won't just be the ones using the best models. They'll be the ones building AI systems that they can truly own, evolve, and improve over time.
6. The AI industry often celebrates benchmark performance and model capabilities. Yet CFOs care about something much simpler: return on investment. How big is the gap between how AI vendors sell their products and how enterprises ultimately evaluate success?
I think the gap is still enormous. Most AI vendors sell capability, while most enterprises buy outcomes. The industry spends a lot of time talking about benchmark scores, reasoning performance, context windows, and model rankings because those are easy things to measure and compare. But when I sit down with business leaders, that's rarely what they're interested in. They want to know whether a process is becoming faster, whether costs are going down, whether their teams are becoming more productive, and whether the economics still make sense six months after deployment. I've yet to meet a CFO who has asked me about benchmark performance. What they care about is whether the AI system creates more value than it costs to run. The challenge is that many organizations are still evaluating AI projects through a technology lens rather than a business lens. That's why we're seeing so much attention shift toward ROI, because eventually every company reaches the same point where they stop asking "Can the AI do this?" and start asking "Does it make economic sense to keep doing it this way?" I think that's the transition the industry is going through right now.
7. If we fast-forward three years, do you think the winners in enterprise AI will be the companies with access to the most powerful models, or the companies that become best at operationalizing and optimizing them? Why?
I would bet on the companies that become best at operationalizing AI and building something they actually own. Access to powerful models is becoming increasingly democratized, which means almost every company will be able to use the same frontier models. That's not where long-term competitive advantage will come from. The organizations that stand out will be the ones that use those models to build their own proprietary layer of business logic, workflows, and institutional knowledge. That's the part that compounds over time and becomes unique to them. Models will continue to improve, and new ones will replace old ones, but the way a company captures knowledge from every execution and turns it into its own intellectual property is much harder to replicate. I think in three years we'll look back and realize that the real race was never about having access to the smartest model. It was about building something on top of those models that competitors couldn't simply access through the same API.
8. There's an interesting parallel emerging between cloud computing and AI. Many organizations initially embraced the cloud for flexibility, only to later face unexpected cost challenges. Are enterprises at risk of repeating the same mistake with AI, and what lessons should they learn from that experience?
I think there are definitely parallels, although AI could become an even bigger challenge because the cost structure is less predictable. With cloud, a lot of companies initially optimized for speed and flexibility and only later started paying attention to efficiency. We're seeing something similar with AI today. Organizations are racing to deploy new use cases, experiment with agents, and automate workflows, which makes sense because the technology is still relatively new. The risk is that many of those decisions are being made without a clear understanding of what the economics will look like once those systems are running at scale. What companies learned from the cloud era is that flexibility is valuable, but flexibility without governance eventually becomes expensive. I think the same lesson applies here. Enterprises should absolutely move fast with AI, but they should also think early about ownership, control, observability, and how costs will evolve as adoption grows. Otherwise, they risk waking up in two years with AI embedded across the business and no clear path to making the economics work. That's why I believe the winners won't be the companies that adopt AI the fastest, but the ones that build it in a way that remains sustainable as usage scales.
9. As AI becomes a core business capability, what parts of the stack do organizations need to control themselves, and where does reliance on external providers still make sense?
I don't think companies need to own everything, and I certainly don't think every enterprise should be training its own frontier model. The reality is that foundation models are becoming increasingly powerful, widely available, and difficult to justify rebuilding from scratch. Where I think organizations need to focus is on making them owners of their own IP: their business logic, their workflows, their controls, and the institutional knowledge they've accumulated over time. That's the part that creates a competitive advantage and compounds with every execution. If all of that lives inside an external model provider, then you're effectively building your future on top of someone else's asset. That's why we advise companies to treat models as interchangeable infrastructure and focus instead on owning the layer that captures how work gets done inside the business. Reliance on external providers still makes perfect sense for raw intelligence, but the decision-making processes, operational controls, and learning generated by those workflows should remain under the company's independence. As AI becomes embedded in core operations, I think that distinction will become increasingly important.
10. Historically, technological revolutions create new metrics for measuring business value. During the internet era it was traffic, then engagement, then cloud efficiency. What do you think will become the defining KPI of successful AI adoption, and why are many organizations still measuring the wrong things today?
I think many organizations are still measuring activity instead of value. They track the number of prompts, the number of users, the number of workflows automated, or even the amount of money spent on AI, but none of those things tell you whether the technology is actually improving the business. If I had to pick one metric that will matter most over the next few years, it will be the value to their own clients, also understood as time to value. Ultimately, companies don't buy AI because they want more prompts or more models. They buy it because they want work to be done faster, cheaper, or to serve their own clients. The organizations that succeed will be the ones that can measure exactly how much value an AI system creates relative to what it costs to operate. That's why I think the industry is entering a new phase. For the past two years, the focus was on proving that AI could do the work. The next phase is proving that it can do the work repetitively and at scale. Once that happens, many of the metrics we obsess over today will start to feel as outdated as page views from the early internet era.
