Baringa halves document drafting time with Azure AI Foundry

As enterprises move AI into production, the question is no longer what models can generate but whether their outputs can be trusted. System One models, which produce structured answers with calibrated confidence, are a welcome step forward, but only if the confidence scores themselves can be trusted.

The past few years have seen extraordinary progress in generative AI. Large language models have demonstrated unprecedented capabilities in reasoning, content generation and knowledge synthesis, creating a wave of experimentation (and often panic) across virtually every industry. As organisations move beyond pilots to production deployment, however, attention is shifting from what models can generate to how reliably their outputs can be trusted.

Most enterprises are designed around decisions, not prose generation. Organisations need to determine whether a document complies with policy, whether an asset requires maintenance or whether a customer interaction should be escalated.

System One models, a category introduced by TypeSafe with its Jev model in September 2026, represent an interesting evolution in this direction. Rather than treating a problem as a conversational task, these models are designed to produce structured outcomes directly. This is an architectural change, and it reflects a different view of how AI systems should interact with business processes.

A welcome addition of structured outputs

Many enterprise agentic AI solutions rely on prompting generative models to produce structured outputs. While this is effective, it creates a mismatch between the technology and the task. This is because the model is still fundamentally generating text, with structure imposed through prompts and rules. As a result, robust guardrails, error handling and post processing are often required to turn responses into reliable operational decisions, with the potential for hallucination.

Constrained decoding tools such as OpenAI’s Structured Outputs already guarantee valid structure, but they work by masking a model’s output. That can distort its probabilities and push it towards answers that are well formed, yet potentially incorrect.

TypeSafe’s Jev takes a different route. It is trained to answer directly from a set of options defined by the user, on which, TypeSafe says, it cannot hallucinate. In practice, the hallucination guarantee covers structure only. The model cannot produce an answer outside the options it has been given, which removes a common failure model in agentic AI systems. It retains the ability to pick the wrong option and to do so with confidence.

Nevertheless, separating structural reliability from factual accuracy is one welcome step forward, providing organisations with greater confidence that model outputs can be more safely integrated into operational workflows.

A shift towards calibrated decisions

As AI systems move into increasingly regulated and business-critical environments, there is growing recognition that organisations require more than impressive benchmark performance. Research demonstrates that LLMs possess some ability to assess correctness of outputs, but these confidence estimates become unreliable when generalising across tasks. This is driving a growing interest in uncertainty quantification and confidence calibration, with the aim of ensuring that a model’s stated confidence genuinely reflects its likelihood of being correct. A model that predicts an outcome with 90% confidence should be correct 90% of the time; otherwise, the confidence score has little operational value.

Autonomous systems must continuously make decisions, evaluate information and take actions with varying degrees of uncertainty and subsequent consequences. It is therefore crucial that an agent can accurately assess not only what it believes to be correct, but also how certain it should be about those beliefs. Well-calibrated confidence enables agents to distinguish between situations where autonomous action is appropriate and those where additional verification, information gathering, or human oversight is required. Without this, agents may act confidently on incorrect assumptions, compounding across multiple steps with consequences that become increasingly difficult to reverse. Effective and trustworthy calibration provides the foundation for proportionate decision-making that balances automation with appropriate caution.

However, providing confidence alone is not enough. Building true trust in these systems ultimately requires transparency into how they have been trained and calibrated. As this capability matures, we expect greater scrutiny not just of model accuracy, but of the evidence underpinning the confidence scores that increasingly drive automated decision making.

Implications for enterprise Agentic AI adoption

The emergence of System One models is not a departure from text generation, but rather than an evolution intended for specific purposes. These models retain the flexibility and reasoning capabilities that have made large language models so powerful, whilst placing greater emphasis on structured outputs, calibrated confidence and uncertainty estimation.

This is a pertinent development for enterprises building agentic AI. Over the past three years, much of the focus has been on improving model capabilities, such as larger context windows, stronger reasoning benchmarks and increasingly sophisticated agentic behaviours. For many organisations, the primary barrier to production deployment is now trust instead of capability. Understanding what the model predicts as well as the confidence behind that prediction is increasingly important when AI systems are used in operational or regulated environments.

An area where we see System One models playing a critical role is evaluations. Confidence-aware models provide a more robust mechanism for assessing the quality, faithfulness, safety and accuracy of AI-generated outputs, helping organisations move beyond LLM-as-a-judge assessments that are merely plausible. Models’ confidence scores can also act as an additional control signal, informing escalation, human review and autonomous decision-making.

The long-term impact, however, will depend on whether confidence estimates themselves can be trusted. A confidence score is only meaningful if it is demonstrably calibrated against representative and correct evaluation data. Whilst providers are increasingly sharing confidence metrics, far less information is shared on training data and calibration methodologies, which are needed to build confidence in the models.

A step in the right direction

Regardless of current transparency in specific System One models, there is huge significance in what they represent and enable. Structured outputs, calibrated uncertainty and explicit confidence measures build towards systems that understand the limits of their own knowledge, transparently communicate those limits and make decisions in ways that organisations can measure, govern and trust.

Baringa is partnering with Cosine, the AI lab building Lumen Sovereign – the UK’s first sovereign frontier AI model – to help develop artificial intelligence technology for the country’s critical infrastructure and other highly regulated environments.

Baringa will help test and enhance Cosine’s model, drawing on its deep experience in the energy sector and across critical national infrastructure, government and regulated financial services. Its work will help shape how Lumen Sovereign performs in environments where enhanced security, privacy, strict governance and niche legacy systems are the norm, not the exception.

Lumen Sovereign is backed by the UK Government’s £500 million Sovereign AI programme and is being trained on Isambard-AI, one of Europe’s most powerful supercomputers, with deployment targeted by the end of 2026. Cosine has already assembled a coalition of major UK institutions – including BT, Lloyds Banking Group, NatWest Group, LSEG, BAE Systems and Thales UK – to help shape the model’s design, a group so far weighted towards finance, defence and telecoms. Baringa’s role brings deep energy-sector expertise to the project.

Baringa has long experience in artificial intelligence, and this partnership is the latest in a long line of AI projects it has undertaken with clients and partners around the world.