Last Updated: 17 September 2026
Up until not so long ago, discussions about enterprise AI mostly focused on capability. What mattered when introducing a new model was whether it was smarter, faster, or could automate processes that were out of reach just a few months earlier.
Cost was not often a major area of focus because, in most cases, AI was not being relied on. It was being experimented with. That is gradually changing as enterprises start putting AI into their regular operations. Companies no longer have the luxury of simply asking what their latest AI implementation can do. They need to understand whether they can run the systems they need economically and whether the numbers still add up when AI becomes part of the business. According to David Villalón, CEO of Maisa, this is usually the wrong way to look at the problem.
In this interview, David explains why the real challenge is not model pricing but architecture, why enterprises have become dependent on frontier models, and what separates organisations that successfully scale AI from those that remain stuck in experimentation.
Maisa AI is an enterprise AI automation company that builds “Digital Workers” for complex, regulated business processes. Its platform, Maisa Studio, lets teams create and deploy autonomous workers in natural language, connecting them to company systems, data, policies and workflows. Unlike general-purpose AI agents, Maisa emphasizes traceability, auditability, reliability and compliance through its proprietary Knowledge Processing Unit (KPU). It serves enterprises in banking, insurance, energy, manufacturing, mobility, engineering and infrastructure, with deployments across 21+ countries and offices in San Francisco, Madrid, Valencia, Barcelona and London.
Q1. Enterprise AI costs have suddenly become one of the industry's biggest talking points. Why now?
I think the assumption that AI has suddenly become expensive is slightly misleading. Costs have become a topic of concern mainly because companies are now trying to deploy AI in their operations rather than just run pilot programmes. These are two very different things.
With a pilot, the goal is to validate the technology and prove that it works. Once that same process becomes part of the business, the questions change. Can you run it every day? Is it reliable? Can you govern it? And, importantly, does the economics still make sense a year from now?
That is why AI economics has become such a big part of the conversation. Enterprises have reached the point where the economics of running AI in production matters just as much as the intelligence of the model itself.
Q2. Why have so many organisations defaulted to using the largest available models for almost every task?
I don't think companies made the wrong decision at the beginning. If you go back two years, everyone was trying to understand what these new models were capable of. Naturally, they reached for the most powerful model available because they wanted to remove as many limitations as possible.
But the situation has changed. Models have evolved, and enterprises now have many more options to choose from. The problem is that very few organisations have stopped to ask whether every workflow still needs that same level of intelligence once it moves into production.
Q3. If a CEO tells you their AI budget is spiralling out of control, what question should they be asking instead?
My first question is usually around how they’ve designed the system. It is easy to assume you have a cost problem, but in many cases, it is actually an architecture problem.
We still see companies using the same frontier model across every workflow because it is the easiest way to get something working. That can make sense during experimentation, but it gets expensive very quickly once those workflows move into production.
Before worrying about token prices, I would look at what each workflow is actually trying to achieve and how much intelligence it needs. In many cases, it does not need a frontier model. The biggest savings do not come from making AI cheaper. They come from designing the system so you are not paying for intelligence you do not need.
Q4. Is there any example of a workflow where companies are currently over-engineering AI?
Document-heavy workflows are a good example because they show how easily companies default to using a frontier model simply because it is the most capable option available.
A lot of organisations process contracts, invoices, compliance documents, internal policies and countless other documents through the same model, even though the level of reasoning required can vary enormously from one task to the next. Reviewing a complex legal agreement is obviously not the same as extracting structured information from an invoice, yet they are often treated in exactly the same way.
That is where I think companies start over-engineering AI. They are solving every problem with the highest level of intelligence available instead of asking what the workflow actually requires. Once you break a workflow down into the different decisions it needs to make, the answer becomes much clearer. Some decisions genuinely require frontier intelligence. Others can be handled much more efficiently by models built for that specific task.
That is how you end up designing a system where each model has a clear role, instead of expecting one model to solve every problem.
Q5. You've previously described this behaviour as "tokenmaxxing." What do you mean by that?
Tokenmaxxing refers to the idea of going all out with AI use and treating the number of tokens you burn as a measure of how much value you are getting from the model.
You see this mindset when companies assume that more tokens, more reasoning and more model usage must mean a better result. But that is not how it works. Just like how lines of code written is not an accurate measure of software development quality.
The better approach is to look at the outcome you need and then work backwards from there. If a smaller or more specialised model can deliver the same result, using a frontier model and burning more tokens does not create additional value.
Q6. The AI industry often assumes bigger models are better models. Why do you disagree?
I think we have started to associate model size with business value, and I don't think that is the right way to look at it. Larger models are incredibly capable, and there are plenty of situations where that extra intelligence genuinely makes a difference. The mistake is assuming that because a model performs better on a benchmark, it will automatically produce a better outcome in every business process.
That's not how organisations work. Most enterprise workflows are made up of lots of different decisions. Some are complex and require a lot of reasoning. Others are repetitive and predictable. If we treat every step as though it needs the same level of intelligence, we end up with a system that is more expensive and harder to scale.
Q7. What's the biggest misconception enterprises have about model performance today?
I still see a lot of organisations treating benchmark performance as a direct indicator of business performance. In practice, those two things don't always move together.
Benchmarks are useful because they tell you what a model is capable of. But capability on its own doesn't tell you whether a particular workflow will perform any better. A model can be objectively more capable and still deliver almost the same outcome because the task simply doesn't require that extra level of reasoning.
That's the distinction I think a lot of organisations still miss. We have become very good at comparing models against each other, but we don't spend nearly as much time asking whether that additional performance actually changes the outcome for the business. If two models deliver the same result, but one is cheaper to run, easier to control and easier to scale, it is difficult to argue that the more capable model is automatically the better choice.
At that point, you are evaluating the model in isolation rather than in the context of the business problem you are trying to solve. Those are two very different conversations.
Q8. Model providers are increasingly competing on price. Do cheaper models solve the problem enterprises are facing today?
Using cheaper models may solve part of the cost problem, but it still leaves enterprises dependent on the terms and conditions of model providers.
The only real way forward is for enterprises to own the intelligence that runs their work. There are a few reasons for that.
First, it frees them from depending on somebody else's pricing. Token costs are currently the one line item that nobody can reliably forecast. You can optimise the price per token, but you are still paying someone else every time the system does work.
Second, it frees you from depending on somebody else's data terms. If your institutional knowledge is constantly flowing into a model you rent, you are building your business on intelligence that someone else ultimately controls.
And third, it frees you from decisions being made entirely outside your business. That could be a government regulation, a change to a provider's terms, a new pricing model, or even decisions made by the people running the company.
So cheaper models are useful, but they don't solve the underlying dependency. If AI is going to run critical parts of an enterprise, the question is not just how cheaply you can access intelligence. It is whether you actually own it.
Q9. The AI industry celebrates benchmarks. CFOs care about ROI. How far apart are those conversations?
They are often talking about completely different things. When people in the AI industry discuss a new model, the conversation naturally revolves around benchmark scores, reasoning capabilities, or how much better it performs than the previous generation. Those are metrics a CFO is not looking at.
A CFO wants to know whether the company is becoming more productive, whether a process has become more efficient, and whether the investment will still make sense a year from now. Neither perspective is wrong. They are just asking different questions.
The problem is that a lot of organisations still assume better model performance will automatically translate into better business performance. That is simply not always the case. At some point, every company has to stop asking how intelligent the model is and start asking whether it is creating enough value to justify becoming part of the business.
Q10. What does a successful AI deployment actually look like from a business perspective?
One of the things we have learned is that making AI work and making AI work for the business are two very different things. It is fairly straightforward to build a pilot that shows what the technology can do. The real test is whether the business still sees the value six months or a year later, once that system is part of everyday operations.
That is where success is actually measured. Is it robust enough? Can people trust it? Will it keep working as usage grows? Does it still make economic sense?
If the answer is yes, then you have created something valuable. And at that point, what matters is the improvement to the business operation, not the technology itself.
Q11. Why do you believe organisations should eventually own part of their AI stack?
One thing I don't think we talk about enough is what happens after AI has been running inside a business for a few years. Models will continue improving, new ones will appear, and today's frontier model will eventually be replaced by another one. That's simply how this industry works.
What shouldn't be interchangeable is the way your business operates. Your workflows, your decision-making, your controls and everything your organisation learns over time are part of your competitive advantage. I don't think companies should be comfortable leaving all of that in the hands of external providers.
That's why I believe organisations should gradually start owning the layer that contains their business logic, while treating the underlying models as components they can replace whenever a better option comes along. That gives them more flexibility and control, and ultimately creates a path towards building custom models and owning the intelligence that runs their operations.
Q12. What are the risks of becoming entirely dependent on external model providers?
There are a few, and the first one is that, to get useful work out of a model, companies have to give it their most valuable asset: their IP.
That means the data on how the company operates, the know-how it has built over the years, the exceptions to its processes that only its teams know, and everything that sits inside the heads of its people. Without that context, the model output is generic and doesn't help the business.
At the same time, companies lose control over what happens to that information because they don't control the model lifecycle. The provider decides which models exist, how long they are supported, what replaces them and the economics of accessing them.
Then there is the way companies pay for that dependency. They pay per token, not per outcome. The bill tells you how much the system was used, not whether the work was actually done correctly. In fact, the better the system works and the more your teams rely on it, the more expensive it becomes.
And after years of spending money and feeding institutional IP into someone else's systems, the company doesn't own the intelligence it helped create. That's the part I think enterprises need to think much more carefully about.
Q13. Where do smaller and specialised models fit into the future of enterprise AI?
I think the only path that can produce positive ROI at enterprise scale is eventually owning the model that runs the work. But almost no company is actually building custom models today. There are three fundamental roadblocks.
First, they don't have the right data. Companies often think they do because they have decades of documents, transactions, and outcomes sitting in their systems. But building a useful custom model requires a much deeper record of how the work is actually done, step by step, including the exceptions and one-off judgement calls. A lot of that knowledge still lives inside people's heads.
Second, they don't have the people. Building a model and keeping it useful in production is a specialised job, and many of the people who know how to do that well are working at the big AI labs today. A bank in Madrid is unlikely to win that hiring race.
Third, even if a company manages to build a model, it still needs to be able to trust it. A small fine-tuned model is more likely to hallucinate than a frontier LLM, so enterprises need a way to make that intelligence reliable enough to run real work.
Those are exactly the three problems we are solving at Maisa. Our architecture, powered by our KPU, breaks processes down into discrete steps backed by traceable and verifiable code. That makes the system reliable by design.
Because every step is explicitly expressed as code, each run also creates a full trace of how the work is actually being done, including the exceptions and one-off judgement calls. Those traces become the data that has historically been missing from most companies.
Over time, that data can be used by our team of AI experts to build custom models that enterprises can own and run with full autonomy. So the goal isn't simply to use a smaller model instead of a frontier model. It's to create the path towards owning the intelligence that runs the work.
Q14. What belief about AI do you think most enterprises have misguided today?
I think a lot of enterprises still believe that adopting more AI automatically makes them a more AI-driven business. It doesn't.
The companies that are getting the most out of AI are much more intentional about where they use it and what they expect it to achieve. They look at the business problem first and then decide whether AI is actually the right tool for it.
That's an important distinction because AI is not a strategy by itself. It is a technology that should help a company achieve a specific goal. Sometimes that means using a cutting-edge frontier model. Sometimes it means using a small, specialised model. And sometimes the right answer is not to use AI at all and stick with a traditional approach.
The goal shouldn't be to maximise how much AI a company uses. It should be to maximise the value the business gets from it.
Q15. If you had to make one prediction about enterprise AI in 2030, what would it be?
By 2030, I don't think we'll be talking nearly as much about which model a company uses. That conversation will feel a bit like asking which cloud provider someone runs on today. It will still matter, but it won't be where companies create their competitive advantage.
What will matter is how well they've designed the systems around those models, how much of that intelligence they actually own, and how quickly they can adapt as the technology changes. The best systems will also improve as the business generates more data and makes more decisions.
I think we'll look back at the first few years of the AI era and realise we spent a lot of time competing on intelligence, when the real competition was always about execution.
