AI in Utilities: Myths, Magic, and the Slightly Hallucinating Future  

 
1 Why utility leaders must turn artificial intelligence into sovereign operational capability

Artificial intelligence has arrived in the utility sector with remarkable speed. Only a short time ago, most discussions about AI in utilities were experimental: a chatbot here, a forecasting model there, a pilot for customer service, an innovation team testing a few prompts. Today, AI is moving into the center of strategic conversations. Boards ask about it. Regulators observe it. Technology providers promise it. Employees use it, sometimes officially and sometimes quietly. Customers will soon expect it.

This momentum is real, and it should not be underestimated. AI will change the way utilities work. It will influence how power, water and gas companies manage customers, interpret data, detect anomalies, plan field work, monitor assets, identify losses, generate reports and make operational decisions. The potential is enormous.

But utilities are not ordinary businesses. A utility does not sell convenience. It provides essential infrastructure. Electricity, water and gas are not optional services; they are part of the foundation of modern life. A wrong recommendation in a generic business process may be inconvenient. A wrong recommendation in a utility operation can be expensive, unsafe, disruptive or damaging to public trust.

That is why the discussion about AI in utilities must be more serious than the usual technology hype. The question is not whether utilities should use AI. They will. The question is whether they will use it in a way that strengthens their operational capability and sovereignty — or in a way that creates new dependencies, new risks and new black boxes.

The slightly hallucinating future of AI is not a reason to reject the technology. It is a reason to lead it properly.

2 The first myth: AI is magic

AI feels magical because it speaks our language. We can ask a question and receive an answer. We can upload a document and receive a summary. We can describe a process and receive a workflow. We can ask for code, a report, a chart or an email, and the system responds in seconds. For many users, this is the first time software feels truly conversational.

This is a breakthrough. Language has become a universal interface to digital systems. That alone is a major shift for utilities, where many processes still depend on complex applications, technical terminology and fragmented data sources.

But language is not the same as understanding. A model can produce a fluent answer without understanding the operational consequences of that answer. It can sound confident while being incomplete. It can produce a plausible explanation that ignores the reality of field work, regulation, data quality, customer impact or network constraints.

This matters because utilities operate in the physical world. A transformer is not merely an object in a database. It is part of a network, connected to customers, affected by load, voltage quality, communication status, weather, maintenance history and investment decisions. A water leakage signal is not just an anomaly. It may represent non-revenue water, pressure changes, asset deterioration, field inspection cost and customer impact. A fraud score is not only a number. It may trigger an inspection, affect revenue recovery, create customer conflict and require evidence.

AI in utilities must therefore move beyond impressive answers. It must become operationally grounded. It must be connected to trusted data, embedded in real workflows, governed by clear rules and measured against real outcomes. Magic is not a strategy. Operational intelligence is.

3 The second myth: one big model will solve everything

Much of the public conversation about AI is dominated by large general-purpose models. They are powerful, and utilities should not ignore them. They can be useful for document understanding, knowledge access, coding assistance, summarization, translation, training and many forms of productivity improvement.

But the idea that one large model will become the single intelligence layer for every utility is too simplistic.

Utilities are specific. They are local. They are regulated. They are physical. They operate with different market models, different network conditions, different customer behaviors, different legal obligations and different infrastructure histories. A water utility in Africa, an energy company in the Gulf, a distribution operator in Europe and a multi-utility in Asia do not live in the same operational reality.

Even within one country, utilities differ significantly. Some manage dense urban networks, others serve rural areas. Some have strong GIS data, others still struggle with incomplete topology. Some have advanced smart metering infrastructure, others are still building the foundation. Some face high non-technical losses, others face aging assets, extreme weather, capacity constraints or regulatory pressure.

A general-purpose model may understand the language of these problems. That does not mean it understands the reality well enough to support operational decisions.

The future of AI in utilities will therefore not be one model. It will be an ecosystem of specialized models, applications and agents. Some will understand country-specific regulations. Some will understand utility domains such as electricity, water or gas. Some will support fraud detection, leakage intelligence, transformer monitoring, prepayment, customer operations, topology validation or field prioritization. Some will be small, focused and highly governed. Others will be larger and more general, used as interfaces or reasoning engines.

The winning AI architecture for utilities will combine both: the flexibility of large models and the reliability of specialized intelligence. The question is not which model is biggest. The question is which model is relevant, explainable and useful for the operational task.

4 From answers to consequences

The real opportunity for utilities is not AI that answers questions. It is AI that understands consequences.

Consider fraud detection. A generic AI can identify unusual consumption behavior and suggest that fraud may be possible. That is helpful, but it is not enough. A utility-grade AI must understand customer type, revenue exposure, consumption history, inspection capacity, commercial rules, legal boundaries and previous field outcomes. It must help decide which cases deserve attention first and why. The value is not the detection alone. The value is better prioritization, better field productivity and measurable revenue protection.

The same is true for water leakage. Detecting an anomaly is not the same as reducing water loss. A useful system must combine smart meter behavior, night-flow patterns, continuous low-level consumption, GIS context, probable network structures, satellite evidence, inspection cost and confidence levels. The operational question is not “Is there an anomaly?” It is “Where should we send the team first, and what evidence supports that decision?”

Transformer monitoring follows the same logic. A warning about overload is useful only if it can be interpreted in the context of voltage imbalance, load trends, power factor, communication health, customer impact, asset risk and maintenance planning. The goal is not simply to produce another dashboard. The goal is to help the utility act earlier, more precisely and with better justification.

This is the core shift: AI must move from information to judgment, and from judgment to action. It must support the complete operational loop: signal, interpretation, decision, action, feedback and learning. Without that loop, AI remains a presentation layer. With that loop, AI becomes capability.

5 Sovereignty is the strategic question

AI creates opportunity, but it also creates dependency. Dependency on data quality. Dependency on vendors. Dependency on cloud infrastructure. Dependency on model behavior. Dependency on external expertise. Dependency on architectural decisions that may be difficult to reverse later.

For utilities, this is not a minor concern. Utilities operate essential infrastructure and must remain accountable for the services they provide. They cannot outsource responsibility. Even when they use external technology, the operational duty remains with them.

This is why sovereignty must become central to the AI discussion.

Sovereignty does not mean isolation. It does not mean rejecting cloud, global AI providers or partners. It does not mean that every utility must build its own models from scratch. That would be unrealistic and, in many cases, inefficient.

Sovereignty means staying in control. It means that the utility understands its data, governs its processes, can explain AI-supported decisions, can audit outcomes and can change course when needed. It means that the utility can benefit from innovation without becoming dependent on opaque systems it cannot understand or influence.

A sovereign utility can use partners without becoming helpless. It can use AI without losing accountability. It can use cloud without losing governance. It can automate parts of the operation without giving up human control where it matters. It can adopt new capabilities while keeping a stable core.

This balance is difficult, but it is essential. AI can strengthen utility sovereignty when it improves transparency, decision quality, productivity and resilience. It can weaken sovereignty when it hides logic, fragments architecture, removes internal understanding or creates dependency on systems that cannot be properly governed.

The strategic task for utility leaders is to ensure that AI becomes a source of strength, not a new form of dependency.

6 The platform becomes the intelligence layer

AI will not create sustainable value if it remains separate from the operational landscape. A chatbot that cannot access trusted data will be limited. A model that cannot support workflows will remain theoretical. A dashboard that detects risk but does not help trigger action will be incomplete.

The next generation of utility platforms must therefore become intelligence layers. They must connect meter data, customer data, asset data, grid data, water network data, GIS data, billing data, field data, satellite data, process data and reporting data. They must make this information usable not only for humans, but also for governed AI applications and agents.

This does not mean replacing the core with experimental technology. Utilities need stability. They need security. They need systems that work every day. But they also need innovation at a much higher speed than traditional utility software cycles often allow.

The answer is modularity. New capabilities must be able to enter the platform without damaging the core. Partner applications, customer-specific modules and specialized AI tools must be integrated into a coherent experience. Users should not be forced into fragmented tools and disconnected workflows. At the same time, the platform must not become a monolithic bottleneck where every innovation requires a major core release.

Innovation at the edges, stability at the core: this principle will define modern utility platforms in the AI era.

The platform is no longer only the system of record. It becomes the system of understanding. And increasingly, it becomes the system through which intelligence becomes action.

7 The role of specialized models

As AI matures in utilities, we will see a stronger role for specialized models. These models may be smaller than the largest general-purpose systems, but they can be more relevant, easier to test and better aligned with domain-specific outcomes.

A fraud model can be evaluated against inspection results. A leakage model can be evaluated against field confirmation. A transformer model can be evaluated against asset events and operational interventions. A topology model can be evaluated against validated network corrections. An AI maturity model can be evaluated against organizational progress.

This is important because utilities need evidence. They need to know whether a model works, where it fails, how it improves and under which conditions it should not be trusted. Generic intelligence is useful, but operational accountability requires measurable performance in a specific context.

This is also where the idea of country-specific and utility-specific models becomes important. Regulation, market design, infrastructure and customer behavior are not universal. A model that supports a European grid operator may not be suitable for a water utility in the Gulf or a distribution company in Africa without adaptation. The more operational the use case becomes, the more context matters.

In this direction, we will likely see a gradual move beyond the language model as the dominant concept. Utilities will need models that understand not only text, but the real-world systems in which they operate. One possible term for this next step is a World Model, or WM (in research the term LWM is used, but I don’t think the world will be so large what the model can understand): a model that understands real-world context and the consequences of actions. This concept should not be treated as a marketing slogan. It is a useful way to describe the direction AI must take in critical infrastructure: from language to context, from context to consequence, and from consequence to responsible action.

For now, the practical lesson is clear. Utilities should not only ask which AI model is most powerful. They should ask which intelligence is most grounded in their operational world.

8 AI maturity is a leadership discipline

Many organizations still treat AI transformation as a tool rollout. They buy licenses, create guidelines, run training sessions and announce pilots. These steps are useful, but they are not enough.

AI changes how work gets done. It changes how teams access knowledge, how they make decisions, how they automate tasks, how they collaborate and how they measure productivity. This means AI maturity cannot be assumed. It must be managed.

In every utility, some teams will move quickly. Some will discover strong use cases. Some will experiment without structure. Some will be blocked by data quality. Some will need training. Some will be skeptical, and sometimes for good reasons. Some processes will be ready for AI support; others will require redesign before AI can create value.

Leadership must make this visible. Utilities need to understand where AI is already being used, which teams are ready, which use cases create measurable value, which risks require governance and which capabilities must remain under human control. AI maturity should not be a one-time assessment. It should become a management discipline.

This is especially important because the greatest risk is not that employees use AI. The greater risk is that they use it without transparency, without guidance and without connection to the organization’s data and governance strategy. If official systems are too slow, people will find unofficial ones. If governance is too restrictive, innovation will move into the shadows. If leadership is passive, AI adoption will still happen — but without direction.

A serious AI strategy must therefore combine enablement and control. It must encourage useful experimentation while defining boundaries. It must build capability while protecting sensitive data. It must give teams tools that help them work better, while ensuring that critical decisions remain explainable and accountable.

AI maturity is not an IT project. It is a leadership responsibility.

9 The new leadership agenda for utilities

The next phase of utility leadership will be shaped by a simple but difficult question: can we become more intelligent without losing control?

This question touches every part of the organization. It touches data quality, cybersecurity, architecture, vendor strategy, regulation, operations, field work, customer service, reporting, skills and culture. It cannot be delegated entirely to innovation teams or IT departments. It belongs on the management agenda.

Leaders must ask whether their organization understands its data foundation, whether AI-supported decisions can be explained, whether model behavior can be audited, whether teams are building internal capability, whether the architecture supports modular innovation and whether AI is improving operational outcomes rather than only producing attractive demonstrations.

They must also decide where human judgment remains essential. AI can support decisions, but not every decision should be automated. In critical infrastructure, accountability cannot disappear behind a model. The goal is not to remove people from the system. The goal is to make people, teams and organizations more capable.

This is a powerful opportunity. Utilities have always been data-rich, but often insight-poor. Smart meters, sensors, GIS systems, customer systems, billing platforms, field reports and operational applications contain enormous value. AI can help unlock that value — but only if the utility has the discipline to connect data, govern intelligence and act on results.

The utilities that succeed will not necessarily be those with the most ambitious AI press releases. They will be those that turn AI into measurable operational capability.

10 Conclusion: the sovereign utility will win

The future utility will not be defined by the size of its AI model. It will be defined by the quality of its intelligence architecture.

The winning utility will combine trusted data, strong governance, modular platforms, specialized models, operational feedback loops, team-level AI maturity and clear accountability. It will use AI to improve decisions, but it will not give up responsibility for those decisions. It will adopt innovation, but it will not lose control of its core. It will work with partners, but it will not become dependent on systems it cannot understand. It will move fast, but not blindly.

AI will shape utilities. That is no longer in question. The real question is whether utilities will shape AI in return.

This requires leadership. It requires a clear view of sovereignty. It requires the courage to redesign work. It requires investment in data foundations and platforms. It requires practical experimentation, but also discipline. It requires understanding that AI is not just another software feature. It is becoming part of how infrastructure is operated.

Power, water and gas require trust. They require resilience. They require accountability. They require control.

That is why the future of AI in utilities must be more than magical. It must be grounded. It must be governed. It must be useful. It must be sovereign.

And above all, it must prove itself in the real world.

But utilities are not ordinary businesses. A utility does not sell convenience. It provides essential infrastructure. Electricity, water and gas are not optional services; they are part of the foundation of modern life. A wrong recommendation in a generic business process may be inconvenient. A wrong recommendation in a utility operation can be expensive, unsafe, disruptive or damaging to public trust.