You shouldn't have to choose between quality and control

Three things to remember

• The perceived trade-off between quality and control is often a sign that the task is not yet fully understood.
• Frontier models create the greatest value during testing and development, where they enable rapid learning and iteration.
• As workflows become better understood and more narrowly defined, a smaller – preferably open – model is often enough, delivering full control at a lower cost.

The false trade-off

The most capable AI models are often hosted on cloud infrastructure governed by foreign jurisdictions, where legislation such as the US CLOUD Act may apply regardless of where the data is physically stored. At the same time, many organisations must maintain strict control over sensitive information. When those two requirements are presented as opposites, organisations are left with an impossible choice – and that perception is often what drives employees towards their own AI tools.

The problem is that the comparison assumes the model remains constant: that the capability required on day one is the same capability required in production. That is almost never true. What a task demands from the model depends on how well the organisation understands the problem – and that understanding grows over time.

Understanding moves intelligence from the model to the solution

At the beginning of an AI initiative, much is still unknown. What is the real task? What questions will users ask? What data is required, and in what condition? Where are the boundaries? At this stage, a powerful model is the right tool – not because the task itself demands it, but because the model's broad capabilities compensate for everything that is not yet understood. That is why the most advanced AI models, often referred to as frontier models, belong in testing and development. They allow organisations to iterate quickly, explore ideas and discover what the task really involves before building a production-ready solution.

With every iteration, something changes. Knowledge sources improve. Prompts become more precise. Business rules are documented. Interfaces to surrounding systems are defined. Step by step, intelligence moves from the model into the solution itself. What the organisation learns becomes embedded in the process instead of being improvised by the model every time it is called.

As the solution moves closer to production, the embedded knowledge does more of the work, and the model has less to figure out on its own. At that point, a smaller model can often deliver the same quality – not because expectations have been lowered, but because the task has become better defined. Quality depends less on the model and more on the solution surrounding it.

Map the process – and identify the repetitive work

The same principle applies at the organisational level. Organisations that truly understand their business processes can identify where AI creates the greatest value – and that is remarkably often in repetitive work: the same type of case, the same type of summary, the same validation, over and over again.

That is good news for two reasons. First, repetitive work can be broken down into narrow, well-defined tasks, and those are exactly the kinds of tasks that right-sized models excel at. Second, narrowly defined tasks are easier to govern, easier to evaluate, easier to trace and well suited to running on open models hosted on infrastructure you control. What started as a choice between quality and control becomes a solution that delivers both – at a fraction of the cost.

Conclusion

The way out of the false trade-off is understanding. Use powerful models to learn quickly. Embed what you learn into the solution through better knowledge sources, clearer rules and stronger boundaries. Map the process and reduce the task to its repetitive core. Production can then often be powered by a smaller, preferably open, model running on infrastructure you control – where quality, control and cost no longer compete with one another.