Enterprise AI in action

How E.on is powering up responsible AI

The utility shows that hard-wiring strict limits and minding the wattage of artificial intelligence need not short-circuit the technology

collage illustration in grey, white and purple colours using energy themed graphic elements

When a customer contacts their energy supplier, something has usually gone wrong. A bill looks odd, a payment has failed or the lights have gone out. In such moments, speed matters, which tempts utility firms to deploy artificial intelligence (AI). Yet accuracy matters just as much, and relying on a probabilistic technology to resolve precise billing or supply problems carries real risks.

E.on, one of Europe's largest energy companies and a major supplier of gas and electricity to millions of British households and businesses, manages those risks through clear limits on what AI is allowed to do and defined decision routes when the technology is used. Fiona Humphreys, chief digital officer of E.on’s British operations, describes the firm’s responsibility in AI as helping customers and colleagues “without breaching the conditions we have as a supplier”, and without losing sight of “who carries the risk when a model is wrong”.

For utilities, insurers, banks and any firm where a wrong answer can harm a customer or breach a licence, there is a strategic question of what tasks to use AI for, but also how to govern that use. E.on’s approach to the technology rests on three principles: build systems with a sense of what’s to come, treat energy use as a cost that belongs on the ledger and design trust into every system that interacts with customers.

Build for the next model

Generals are famously accused of fighting the last war. In the realm of AI, many corporate boards appear to be doing much the same. When engineers at Samsung, an electronics giant, leaked proprietary code to ChatGPT a few months after the tool was released, a run of firms worldwide rushed to ban AI chatbots across their business. This instinct is understandable but insufficient.

When Ms Humphreys observed the early phase of the generative AI boom, she saw a “race between curiosity and chaos”. Rather than joining the ranks of the prohibitors, the energy giant decided to bring the technology “inside the fence” by launching E.on GPT, a generative AI platform. Doing so achieves two things. It cuts the risk of data leakage when staff use AI without the right licences, and it creates a controlled setting in which governance policies can be tested against reality.

E.on also knew it needed to prepare for what comes next. Ms Humphreys describes one of her team’s jobs as sending “postcards from the future”. They scan the horizon, from Silicon Valley startups to new energy entrants, to translate looming shifts in AI into decisions about architecture and governance. The aim is to prevent brittle systems by building in redundancy and safety margins from the start.

collage illustration in grey, white and purple colours with building and pylon themed graphic elements

One anticipatory step that most firms can take is to dismantle what security researchers call the “lethal trifecta”, when an AI system is simultaneously granted access to untrusted data (like incoming emails), access to sensitive secrets (like passwords or customer databases) and the ability to communicate with the outside world. A “future-gated” architecture separates these functions, for example, by using two separate AI models. One less trusted model would parse incoming data but have no access to secrets and a second, trusted model would hold the keys but only receive sanitised inputs from the first. The technical details differ by firm but the underlying point does not: separate what can be poisoned from what can do damage.

“AI is a great tool, but it shouldn't be your first choice for everything. People need to understand the energy and water cost behind every AI prompt.”

Fiona Humphreys

Chief digital officer, E.on UK

Chief digital officerE.on UK

Put energy use on the ledger

To the uninitiated, “the cloud” sounds weightless. But to a firm that sells kilowatt-hours, the data centres where AI is born and lives weigh heavily. E.on is “very conscious of the environmental impact of AI”, says Ms Humphreys. Other firms should be too, even if the emissions tied to AI often feel remote. Most of AI’s headline-grabbing energy use does not come from training models but from “inference”, the routine querying in everyday work. Jefferies, an investment firm, estimates that inference can account for up to 90% of total AI power consumption.

E.on’s response begins with a question that too few organisations ask: is AI the right tool for this task? Using a large language model (LLM) “just for the sake of it”, says Ms Humphreys, carries a cost, both financial and environmental. Her team applies that principle twice: first when choosing which projects should use AI at all, and again when nudging staff to treat LLM use as part of the firm’s resource use. To reinforce the point, E.on built a game-like tool that teaches staff to write tighter AI prompts. Academic research shows that such prompt design choices can more than halve the energy used during AI inference without affecting how well tasks are performed.

collage illustration in grey, white and purple colours with arrows and counter themed graphic elements

The logical next step for firms is to match model size to the task. Using a massive model for basic administration is akin to using a sledgehammer to crack a nut. Small language models, specialised alternatives to full-scale LLMs, are increasingly seen as the future of routine AI tasks. A paper by researchers at Nvidia, the world's leading AI chipmaker, and the Georgia Institute of Technology notes that a small language model can be up to 30 times cheaper and less energy demanding to run than an LLM that is up to 25 times bigger. This should encourage an architectural shift across firms to build a toolbox of small, specialised and energy-thrifty AI models for routine tasks, and reserve the most capable models for the few tasks that genuinely require them.

Design trust into customer contact

Enduring a thirty-minute flute solo while on hold is one of the annoyances in customer service that generative AI promises to end. Yet as firms rush to replace call centres with chatbots, new risks emerge. Air Canada, for example, found itself liable for the hallucinations of its customer-service chatbot after it invented a refund policy for a grieving passenger. And in a survey of over 5,700 customers by Gartner, an advisory firm, about two-thirds said that they would prefer that firms did not use AI for customer service. One of their worries is “deflection”, when firms create digital mazes that prevent users from reaching human support. 

E.on’s response is to treat customer-facing AI as an addition to its staff. In its contact centre, the energy giant has built two voice agents, Claire and David, to handle routine queries and triage calls, leaving more complex matters to human staff. The agents follow a “decision tree,” a predetermined series of yes-or-no steps that channels each query to a defined outcome, and are additionally trained to stay within scope. If a conversation drifts beyond their remit, or the system appears unsure, the agents pass the call to a human adviser. “You need to have a clean path to say: this is outside the remit, and the human will help,” Ms Humphreys argues. 

The broader point is instructive. An AI system that never admits uncertainty creates one kind of problem: customers learn to distrust both it and the firm deploying it. But if customer-facing AI acknowledges its limits, maintains a record of what it attempted and escalates to a human when appropriate, customers might appreciate it for what it is worth.

“We treat the agent almost like a human: what's its capability, what are the boundaries, and when does the decision tree say 'bring a human colleague in'? The moment an agent starts to get confused, it escalates.”

Fiona Humphreys

Chief digital officer, E.on UK

Chief digital officerE.on UK

Regulation as co-design

Energy suppliers in Britain operate under a licence. Ofgem, the regulator, can revoke it. That makes using AI responsibly at E.on a condition of doing business. “We are very conscious that ideas coming out of [AI] systems need to be congruent with our supplier licence,” says Ms Humphreys. If an AI agent advises a customer incorrectly, E.on must show how the answer was generated, which data informed it and why human oversight did not catch the error. “You have to be able to trace that answer all the way back to its core data source,” she explains.

That level of scrutiny can sound like a brake, but the rules in regulated environments can also help firms design systems that are safer and easier to defend. In July 2025 Ofgem announced it was exploring an “AI technical sandbox”, a controlled environment where energy companies could test AI tools against the regulatory framework before deploying them. The sandbox aims to support “safe, secure, sustainable and fair” AI, with an eye on familiar failure modes: bias, opacity and security weaknesses. E.on has acted as a “sounding board” on what good practice might look like, says Ms Humphreys.

In sectors from finance to healthcare, regulators are increasingly edging from after-the-fact enforcement towards earlier engagement, because AI governance cannot be bolted on to systems that are already live. For firms, that shift could be an opportunity to treat regulation as a form of co-design.

collage illustration in grey, white and purple colours with brain and persons head themed graphic elements

Responsible AI in action

By building governance into AI’s architecture, weighing the technology’s energy costs and teaching it to know its limits, E.on avoids costly blunders. This lesson stretches far beyond the energy grid. In any industry where mistakes matter, a company that knows what its AI cannot do will outlast one that races blindly ahead. This principle of responsible leadership anchors the third pillar of our Enterprise AI in action article series. The instalment also includes an article on how DHL, the world’s biggest logistics firm, guides its AI projects to keep order and an article on how to make responsible AI use a daily habit, drawing on conversations with experts from the UN and Cambridge University.

Sources

  1. Forbes. Samsung Bans ChatGPT And Other Chatbots For Employees After Sensitive Code Leak. May 2023. Available at: https://www.forbes.com

  2. The Guardian. AI drive brings Microsoft’s ‘green moonshot’ down to earth in west London. June 2024 (last modified April 2025). Available at: https://www.theguardian.com

  3. Riccardo Rubei; Aicha Moussaid; Claudio di Sipio; Davide di Ruscio. Prompt engineering and its implications on the energy consumption of Large Language Models. January 2025. Available at: https://dl.acm.org

  4. Peter Belcak; Greg Heinrich; Shizhe Diao; Yonggan Fu; Xin Dong; Saurav Muralidharan; Yingyan Celine Lin; Pavlo Molchanov. Small Language Models are the Future of Agentic AI. June 2025. Available at: https://arxiv.org/pdf/2506.02153

  5. BBC Travel. Air Canada chatbot misinformation: what travellers should know. February 2024. Available at: https://www.bbc.co.uk

  6. Gartner. Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn’t Use AI For Customer Service. July 2024. Available at: https://www.gartner.com

  7. Ofgem. AI technical sandbox. January 2026. Available at: https://www.ofgem.gov.uk/consultation/ai-technical-sandbox

More from the series