Why the biggest business decision about AI isn’t which model to choose

THE e-mail from your artificial intelligence provider arrives. The model your business has built into its applications is being retired. A replacement is on the way, and there’s enough time to migrate.

It all sounds routine.

Then your teams start asking questions. Will the replacement produce the same results? Will it cost more? Will applications built around the old model still work as intended? How much needs to be tested before the migration deadline?

This is the less glamorous side of the AI boom. Everyone talks about which AI model to choose, as if that were the hard part. But changing the model could be more challenging than choosing it.

Across the Asia-Pacific, more firms are bringing AI in-house. Globally, F5’s 2026 State of Application Strategy Report found that 78 per cent of organisations now run AI inference in-house. In other words, they now run an AI model on their own hardware or private cloud infrastructure.

But bringing inference in-house does not remove the switching problem mentioned earlier. It multiplies it. When a business hosts and runs its models internally, it also inherits full responsibility for what happens when one of those models needs to change.

Consider the scale of it. The average organisation now runs seven AI models in production or active evaluation, each one turning a request, such as summarising a document or answering a customer enquiry, into a response.

Three forces are driving this shift in-house.

Cost is one driver, as usage-based pricing from third-party hosted services grows harder to predict. Data sovereignty is another, as regulators tighten rules on where information can be processed. Lastly, geopolitical uncertainty now shapes technology decisions as much as it shapes trade ones.

Different jobs need different models

Bringing inference in-house does not mean favouring one model over others. No single model fits every task, a routine enquiry may suit a smaller model, a complex problem a more powerful one, and sensitive data a locally hosted one.

Consider a bank operating across the region. It might use a lightweight model for routine loan applications in English, another tuned for Bahasa Indonesia or Mandarin, and a third, locally hosted model reserved for anything touching customer financial records. Three tasks, three models, three very different requirements.

For businesses across the Asia-Pacific, that flexibility matters. Language performance, regulation, data requirements and cost can vary by market, affecting which model works best.

Choice gets harder at scale

Managing that mix is straightforward with a handful of AI pilots. It becomes much harder when AI spreads across dozens or hundreds of applications.

By then, these systems may be quietly working behind the chatbot handling customer complaints, the candidate screening tool or the system flagging suspicious transactions.

Many organisations build their choice of model directly into an application, so replacing the model means rebuilding the application. At scale, what sounds like a simple model change becomes a development project spanning dozens of applications, each needing updates and testing.

Models will change, too, as cheaper or more capable options may emerge, preferred models become unavailable, or regulation shifts where information can be processed.

What begins as a technology choice then becomes a business constraint.

Centralise the rules, not the models

As businesses use more models, they need consistent rules for where data can go, which models can be used and how that use is monitored.

Consistency does not mean putting every AI model in the same cloud or data centre. Models can run in different places. But the rules governing them should be centralised, directing routine tasks to lower-cost models and keeping sensitive information within approved environments.

Applications should use the model best suited to the task, without being redesigned each time, and without the business losing control over cost, data or risk.

Businesses need a control layer that lets a business swap the model underneath an application without rewriting the application itself.

The questions worth asking early are simple: Which model does this task actually need? Where is the data allowed to go? What happens if we need to change models tomorrow?

Build for the model that comes next

Business leaders cannot know which AI model will come out on top. Their priority should be to prepare for change, with the ability to adopt better, cheaper or more suitable models as they emerge, without disrupting the business.

As today’s preferred models give way to new ones, that flexibility will become more valuable. The businesses best placed to adapt will be those that have built in the freedom to choose differently tomorrow.

Ask your technology team: If our AI model changed tomorrow, how many applications would need to be rebuilt? If the honest answer is “we don’t know”, that is the real issue needing attention.

The writer is chief technology officer for the Asia-Pacific, China and Japan at F5