AI at Volue

Every energy operator can broadly map themselves into these buckets. The question is not which column you are in, but how fast you can safely move right.

There are many excellent responsible AI statements out there, many of which we agree with - particularly that fairness, transparency, accountability, privacy are paramount.
However, we wanted to build on this, sharing our perspective on the five principles for responsible AI in energy operations, where the stakes are too high to experiment with black boxes on live operations.
This is a sneak peak into our thinking, ahead of a set of new product launches coming in time for E-World next February.

1. Critical infrastructure needs deterministic control boundaries

• Critical decisions and operations should not be handed to probabilistic large language models without the right control boundaries in place
• Regardless of how good LLMs have become, they are probabilistic by default: ask the same question twice and you can get two answers, and making them repeatable takes significant engineering effort. The deeper issue, though, is calibration: models do not reliably know when they are wrong, which is the root of hallucination and overconfidence. No amount of testing rules out the edge case where the model does something nobody intended; this is why clear boundaries and controls are critical
• Whilst not every process in energy operations has to be deterministic what must be both deterministic and independently enforceable, are the control boundaries around execution, for example a traditional LLM analyses and proposes, while separate layers validate, authorise, execute and protect
• Critically, though, our deterministic systems cannot be static. Coding agents let us improve them faster than at any point in Volue's history, and where a newer technique proves itself, in forecasting for example, we adopt it

2. Autonomy is earned through time and solid evidence

• Unbounded AI is the sorcerer's apprentice - tireless, literal, flooding the room before anyone notices
• The answer to this challenge is not less ambition; it is agents put to work in clearly bounded ways
• To scale in production, agentic workflows should move through defined stages: observe, recommend, execute under supervision, execute within bounds. Shadow operation and human verification at each stage build the evidence; time spent in a stage, on its own, proves nothing
• Oversight should also be uncertainty-aware. A control layer that holds calibrated beliefs about how reliable each agent is, and updates them with every human review, can decide when to proceed, when to check and when to escalate. Recent research argues this is where probabilistic reasoning belongs: in the orchestration, not the model
• Intelligence and authority are different things. An AI agent can have sophisticated reasoning and optimisation capability without any authority to execute an operational action; the execution and protection layers around it may have far less intelligence but far more authority over the physical system. More intelligence should never automatically mean more authority
• The decision to remove a person from the loop should follow three things: the consequence and reversibility of the action, the strength of the control environment around it, and the maturity of the organisation operating it
• Bounded does not necessarily mean small. An agent within clear limits can work every hour and every market window without a person watching each step.

3. Auditability has to be built in by design

• Every action that touches operations leaves a tamper-evident record with full decision lineage: the model and version used, the data and context it saw, the checks applied, the authority behind the action and the resulting outcome, regardless of whether the action was taken by a person or an agent
• We're building this because we believe it's the right infrastructure to enable our customers to move faster and be more competitive: a system that produces its own evidence can be extended without re-auditing everything around it, and it answers a regulator's question with a query rather than a project. Compliance done this way is a competitive edge, not an overhead.
• The EU AI Act will require this record-keeping of high-risk systems from December 2027. NIS2 and REMIT II already ask our customers for the same evidence. Building it in now means neither we nor they retrofit it later.

4. Sovereignty is critical, but it's a spectrum

• Sovereignty is important but can be abstract
• In energy operations we think what matters is retaining control, resilience and credible choice across the technology stack: control over data, identities, models, critical operations and decision rights; regulatory compliance; continuity if a service becomes unavailable; and a credible exit path if conditions change
• Open standards, common semantics, interoperable interfaces, portable workloads and exchangeable AI models are strategic controls, not technical niceties; rather like TCP/IP: its strength is not that every component is identical, but that a common standard lets different systems and vendors interoperate
• From our experience and conversations with customers, we see four main "dials" for control:
  • Where the system runs
  • Whose law governs access
  • Whether the model can be changed (e.g., for LLMs, open weights let you adapt the model to your own operational data while that data stays in house, which is where sovereignty and performance line up rather than trade off)
  • Whether anyone outside can switch it off (e.g., a proprietary API can be withdrawn by a provider or its government)
• We recognise that our customers sit at different points on that spectrum. Some want everything inside their own perimeter; others accept a US hyperscaler on EU soil
• Whilst you can tune each dial separately, the settings have to work as a whole. This is thinking we do actively, alongside our customers, to suit their needs and requirements
• Sovereignty should also be risk-based in the business context. A workload with no direct impact on production, safety or system stability does not need the same controls as SCADA, virtual power plant control or autonomous market and grid actions. The objective is not to localise everything for the sake of doing so, but to ensure that critical capabilities remain governable, portable and recoverable

5. Outcomes trump tokens

• We have moved past the era of AI adoption as the most important measure: usage figures say nothing about whether an operation is safer, cheaper or more effective
• What matters is AI's impact on the numbers that your business runs on: forecast accuracy, trading performance, reliability, stability
• Technology suppliers will create the most value where they put tokens to work most efficiently
• Across our AI product strategy our target is maximum business impact for the least compute; stay tuned to see more from us on this over the coming weeks and months