Energy networks are leaving hundreds of millions of pounds in innovation capital stranded because technically viable artificial intelligence pilots lack the digital trust required for full operational deployment.
DNV warns in its new paper, ‘Managing AI system risk within energy networks through digital trust’, that scaling these solutions safely is critical.
It says energy regulators must evolve into continuous assurance orchestrators to audit adaptive AI systems.
Operators require vendor-agnostic frameworks with explicit guardrails, ensuring human override capability and strict data-driven compliance as safety-critical networks integrate volatile new fuels.
In an exclusive interview with OGN energy magazine, Graham Faiz, Head of Digital Energy, UK and Ireland, Energy Systems at DNV, says: “Ultimately, the energy transition will have to be digital, but it will only succeed if it is trusted.”
Following are excerpts from the interview:
The paper notes that excellent AI solutions frequently remain trapped in the pilot phase. To what extent is this a result of cultural risk-aversion rather than actual technical limitations?
The lack of pilot scaling is not primarily due to intrinsic technical limitations, but it is due to a lack of evidenced confidence in how AI systems behave in operational, safety-critical contexts; it’s less risk-aversion in control rooms and more about risk tolerance.
Arguably autonomous systems already outperform human operators for certain tasks in both efficacy and speed, but without trust, they cannot be adopted in safety-critical systems.
This is where DNV’s digital assurance comes in: Far from being a burden or barrier to innovation, assurance acts as an enabler by supporting strong solutions bridge the trust gap into operations.
Given that digital trust is framed as a practical enabler of innovation, can you quantify the volume of stranded innovation capital currently tied up in utility pilots that fail to scale to business-as-usual deployment?
Industry experience suggests that a material proportion of innovation spend does not reach operational scale.
Across the UK and EU network innovation programmes (for example, NIA/SIF-type portfolios), it is common for only a minority of pilots to transition fully into business-as-usual deployment.
Given the scale of annual innovation investment by energy networks, tens to hundreds of millions of pounds of capital may be tied up in solutions that demonstrate technical feasibility but do not progress to operational adoption.
Energy regulators must evolve into assurance orchestrators of complex digital systems
When moving from decision support to fully automated network control, what are the precise operational trigger points or guardrails that dictate when a human operator should legally and safely hand over real-time control to an autonomous AI system?
Handover from human-led operation to autonomous AI is only justified when the assurance case demonstrates, with evidence, that the system is bounded, stable and operating within defined guardrails. It should be monitored and overridable and supported by clear accountability and fit-for-purpose data.
AI systems, like any other, operate within the relevant safety regulatory environment.
The onus is on the operator, in conversation with the regulator, to determine the operational trigger points for autonomous control.
They must be able to justify their decision and demonstrate a thorough risk assessment process.
Traditional regulatory frameworks are designed to audit static, predictable physical assets. How must bodies like Ofgem evolve to effectively audit adaptive, learning AI systems whose operational behaviours fundamentally change over time?
Regulators like Ofgem must evolve from compliance authorities into assurance orchestrators of complex digital systems.
Ofgem must move from point in time approval of static systems to continuous, risk based assurance of evolving AI systems, requiring evidence of safe behaviour in operation, lifecycle monitoring, and human overridable control, with increasing autonomy granted only where supported by a robust assurance case.
We expect that the industry will place a greater emphasis on continuously monitoring solutions; hence attaining data-driven compliance.
We see a strong role for 3rd parties here, to provide independent monitoring and compliance services digitally.
If an AI model optimises a safety-critical asset but inadvertently causes a localised network outage or consumer harm, where does the ultimate legal liability rest?
Autonomous AI systems do not transfer accountability; they operate within boundaries defined and approved by the operator through an assurance case.
If harm occurs, liability rests with the party that authorised deployment and permitted the level of autonomy, with potential secondary liability sitting with vendors or assurance providers under contractual arrangements.
How can operators build standard compliance and digital trust frameworks when constantly updating, patching, or swapping out proprietary third-party software vendors?
Operators should build compliance around a vendor-agnostic assurance framework, not individual solutions.
This means defining system-level claims, requiring evidence-based validation of behaviour under real conditions, enforcing standardised data, monitoring and audit interfaces, and maintaining continuous lifecycle assurance.
Vendors can then be swapped or updated, provided they meet the same evidential and operational guardrails, ensuring trust persists even as underlying technologies change.
Ensuring any digital solution is up-to-date and security patched is essential to demonstrate digital trust, as well as to maintain a cyber-secure system.
In the context of the Intelligent Gas Grid (IGG) project, how do data-driven AI models maintain predictive accuracy and safety assurance when historical training data lacks examples of abnormal or stressed network operating conditions?
Unsupervised learning is an established technique in machine learning (ML) whereby the model attempts to recognise patterns (or anomalies in this case) in the data stream without prior knowledge.
Such anomaly detection systems can identify deviations from normal operating conditions and alert users when they exceed a threshold.
Then field trials and in-operation experience can verify whether the detected data anomalies relate to a physical condition on the system.
As operational experience grows, the model can be retrained to account for the new knowledge, improving the detection characteristics.
In DNV’s role on the IGG project, we achieved this by applying our digital trust claims-based assurance framework, which accounted for model capability and safety expectations.
How does the digital twin framework adapt to the shifting physical asset behaviours and volatile chemistries of new fuels?
A digital twin is a continuously assured, data-driven representation whose validity is bounded, monitored, and re validated as system conditions change.
As networks introduce variable green gases, such as biomethane and hydrogen, digital twins adapt through a combination of continuous data integration, explicit definition of model limits and assumptions, and claims-based assurance.
New operating regimes are only incorporated where evidence demonstrates safe behaviour, with autonomy constrained to validated envelopes and supported by continuous monitoring, traceability, and human override.
This allows the digital twin to evolve alongside changing asset behaviour while maintaining safety and regulatory confidence.
How does a digital trust framework resolve the commercial friction and data-sharing hesitation inherent between distinct commercial entities?
As the industry and regulation mature, we see increasing focus on data standards and common information models.
Ofgem already publishes a Data Best Practice guidance. Our framework provides a structured approach to ensure that digital twins and data infrastructure conform to the market requirements, as well as verifying their functionality.
This includes requirements for data sharing, including the data-security aspects, for example establishing that only necessary data is shared between parties, suitably aggregated or anonymised where required.
What does a stress test under DNV’s Technology Qualification look like to separate genuine physics-informed industrial AI from corporate hype?
Technology qualification provides a claims-based framework to establish first feasibility and then capability of novel technologies where no agreed standard exists.
Proprietary AI systems clearly fall into that category.
Supporting DNV-RP-A203 (Technology Qualification), DNV has produced a set of Digital Trust recommended practices, which provide detailed baseline evidence expectations for digital systems, including DNV-RP-0671 (AI-enabled systems). We evaluate how the AI-enabled system behaves under real, including abnormal, operating conditions.
The recommended practices establish evidence expectations around accuracy, bias stability, system interactions and the effectiveness of human oversight.
AI introduces emergent risks, which are still poorly understood, and vendors must consider the whole system design, not just the AI component. In its KTP project with the Centre for Assuring Autonomy at the University of York, DNV aims to design a safety case approach for autonomous AI operations in the energy sector, building on York’s strong understanding of AI safety in other sectors.