Bigger is not always better: why the UK's critical infrastructure needs right-sized AI

  • Artificial intelligence
  • Europe
  • United Kingdom

© AdobeStock

  • Type Insight
  • Published

The race to build ever larger Artificial Intelligence (AI) models has encouraged an assumption that bigger means better. In the systems that keep the country running, that is often the wrong starting point. Ajay Chakravarthy, Chief AI Officer in the UK, argues that critical infrastructure needs AI matched to the task, the operating environment and the required outcome, not to the headline.

The biggest model on the market is rarely the right AI for the systems that keep the country running.

The assumption we need to drop

For three years the conversation about AI has been dominated by scale: who has the largest model and the biggest data centre. It has created an expectation in boardrooms that the most powerful model must be the right choice for every job. 

In critical national infrastructure, that assumption does not survive contact with reality. A water treatment plant, rail signalling centre or grid substation is not a cloud environment with unlimited compute and a tolerant user. It is constrained and regulated, and a wrong answer can contribute to outages, safety incidents and a loss of public trust. 

Operators often know this. The harder task is turning it into an investment decision when the market is selling scale. 

What critical infrastructure actually needs

Five constraints shape AI in these environments. None rewards size for its own sake. 

Security comes first. Operational technology was rarely designed for today's connectivity, let alone for autonomous AI components using external tools. Any AI that touches a control environment must meet the same disciplines as conventional software, or stronger, from least privilege to incident response. For higher-risk workloads that will often favour smaller models in an approved sovereign or local environment. External hosting is not ruled out, but it must be justified against the data, the mission and the regulatory boundary. 

Connectivity and latency come next. A decision about load balancing or a fault on a line may have to be made within a strict time window, on the hardware available, even when an external link is degraded or cut. A model that depends on continuous remote access may be operationally unusable. 

Energy is underestimated. Operators are being asked to adopt AI and decarbonise at the same time, and those aims conflict if compute and tokens are treated as free. Frugal AI, delivering the required outcome with the least data, energy and compute, is not a lesser form of AI here. It is an engineering requirement. 

Reliability and assurance complete the picture. Reliability means the system performs safely within a defined operating envelope. Assurance means the operator can demonstrate it, with controlled change and recoverable failures. Both favour bounded, task-specific systems over a general-purpose service whose behaviour may change without the operator's approval. 

Right-sizing in practice: start with the task 

Right-sizing is a discipline, not a product. It starts with three questions, asked before any model is chosen. What is the task? What environment must it operate in? What measurable outcome will show it works? 

Answer those and the right architecture becomes clear. Detecting anomalies in sensor data from a pumping station does not require a model trained on the whole internet. It requires a system validated on relevant operational data, able to run within the site's constraints, and designed so a control room engineer can understand it, challenge it and intervene. 

The arrival of AI agents makes this discipline more important. An agent does not merely generate a prediction; it can plan, call tools and act. Cybersecurity evaluations in 2026 gave a useful warning: models operating with reduced safeguards and offensive security tools exploited weaknesses in shared infrastructure and reached systems beyond their intended sandbox. That does not mean every agent will escape. It means a goal-directed agent will find the gaps in identity, permissions, egress and monitoring. 

For critical infrastructure, assurance therefore extends beyond accuracy. Operators must define the agent's authority boundary, permitted tools, credential scope, human approval points and a reliable means of interruption. Until those controls have been demonstrated, agents should be narrow in scope, supervised in operation and accountable to a named owner. 

None of this argues against ambition. It argues for matching implementation to need, so an organisation delivers value now and builds the governance and skills to do more later. The technology is rarely the hardest part; helping people, processes and culture evolve with it is. 

What defence has already learned 

Defence has lived with these constraints for decades: limited hardware, contested connectivity, sensitive data and decisions with serious consequences. Thales has engineered AI into critical systems for around 40 years, and this year my colleagues set out what that teaches in a white paper for UK Defence, Trusted AI for the Integrated Force. Operational advantage does not come from the model alone but from the engineering harness around it: trusted data, mission-specific assurance, security by design, human oversight and feedback loops that turn experience into controlled improvement. That is also one of the six principles of the UK technology strategy, digital and data first: build the data foundations before asking what AI can do. 

Three lessons read directly across. Useful AI depends on domain knowledge. AI is not plug and play; it must be engineered into its host system with security, data and human factors designed in from the outset. And trust must rest on evidence and a clear line of human accountability, not on benchmark scores. 

In July the cortAIx Factory in the UK delivered its first operational AI capability into the RAF's Multi-Domain Mission Support System. Using a frugal learning approach developed with Faculty, it trains on limited real-world data to model normal maritime behaviour and flag anomalies operators might otherwise miss. That is right-sized AI, in service. Behind it sit more than 800 AI and data engineers and scientists and around 50 sponsored PhDs at UK universities. 

Where this goes next 

The policy environment is moving the same way. 

Speaking in October, the AI Minister, Kanishka Narayan, said he wants Britain to become "one of the hardest countries in the world to attack using artificial intelligence", and that the Cyber Security and Resilience Bill, now progressing through Parliament, will be used to strengthen protection against AI-driven cyber threats and loss-of-control risks. That puts the burden where it belongs: on how AI is governed inside essential services, not on how large the models are. 

The organisations best placed will be those that can show AI matched to the task, secured within an explicit boundary, efficient with data and energy, resilient when degraded and accountable to people. The principle we apply in defence holds here too: AI wherever it accelerates understanding, human judgement wherever accountability matters, integrated always. 

That is a more measured story than the AI industry tells, and the one that will keep essential services running. The future of AI in critical infrastructure will not be defined by the size of the models we buy. It will be defined by the trust we can demonstrate, the resilience we build and the value we deliver to the people who depend on these systems every day.