Winning strategies for risk mitigation for utilities in the GenAI era

Utilities can take steps to capitalise on generative AI while mitigating concerns around data accuracy and security, writes Shriram Ramanathan, AI strategist and chief of staff at Bidgely.

Few organisations are as risk averse as energy utilities, evidenced by the industry’s historically conservative adoption of new technologies.

The usefulness of artificial intelligence (AI) to streamline operational efficiencies, however, has been hard to ignore. Today, nearly three-quarters of energy and utility companies have implemented or are exploring using AI in their operations, according to IBM’s recent global study. 

As the generative AI (GenAI) hype cycle peaks, utility decision makers are weighing the pros and cons of integrating another new – and potentially risky – technology into their workflows.

The benefits of GenAI are clear: improve asset management; reduce operating costs; increase employee productivity; and enhance customer engagement.

That said, adopting GenAI can come with its own set of risks and challenges, as with any other technological advancement. Fortunately, it doesn’t have to be an either-or scenario.

Have you read?
Generative AI has ‘turbocharged the power of the human’ says Octopus Energy founder Greg Jackson
AI is the master key to unblocking our power grids

Prioritise clean and ‘non-sensitive’ data

One of the main concerns for GenAI is that the underlying data used for training GenAI algorithms itself could be wrong. ‘Dirty data’, i.e. datasets with errors or inconsistencies, can unwillingly train the GenAI model to learn incorrect patterns, resulting in ineffective outputs based on unwanted biases or inaccurate predictions.

In addition to generating wrong answers, training with dirty data also increases the risk of exposing sensitive customer information.

When information is given proper human oversight, analysts can more easily differentiate between confidential information such as home addresses, phone numbers or social security numbers, and non-confidential information. Blockers can then be put in place to prevent information leakage.

However, it becomes harder to parse out confidential information from non-confidential information since the GPT model extracts ‘knowledge’ rather than stores information in its raw form.

Utilities can clean their data by scrubbing for all PII and tax-related information from the dataset. They can also make any vendors handling the data equal stakeholders to be bound by service level agreements around data sanitisation.

Then, by applying a ‘trust and verify’ model to its GenAI programme, utilities are able to balance the powerful capabilities of GenAI with the validity of its outputs. This evaluation is critical to ensuring utilities make decisions based on reliable data.  

Understand hallucination risks for B2B, B2C, and B2B2C

GenAI for business-to-business (B2B) and business-to-business-to-consumer (B2B2C) is very different from business-to-consumer (B2C).

GenAI models for B2C are fairly low on the risk scale because these models are typically based solely on publicly available data. This not only reduces the risk of inaccurate results, but there is not much concern for exposing a person’s confidential information.

Consumer-facing GenAI is also trained on the internet’s exorbitant archive of information. Businesses, on the other hand, work within limited data sets and therefore do not have sufficient enough context to generate correct answers. This goes back to the idea of training on incorrect data and ending up with nonsensical or unverifiable answers, known ‘hallucinations’.

When working with more limited data sets, utilities will want to carefully curate its source data and set proper parameters for each prompt around what information is allowed to be shared or not. Periodically conducting quality assurance tests or ground truth verifications also greatly reduce the likelihood of incorrect responses.

Start with a strong foundation of traditional AI

Rapid proliferation of enterprise-level GenAI has spurred dozens of new technology providers to offer GenAI services to utility companies, allowing utilities to buy utility-specific GenAI solutions from vendors rather than build these from scratch.

Unfortunately, the majority of these vendors have jumped directly to GenAI and do not have strong backgrounds in traditional AI and machine learning. For a technology that feeds off long-term learning, these solutions would still be prone to these common risks.

And, many of the solutions are not completely plug-and-play, requiring a great deal of time and ongoing hand-holding to successfully integrate and optimize these tools.

The good news: there are companies that made the journey to GenAI after spending years using traditional Al to analyse utility meter data. The advantage is having billions of data points to help identify errors and established protocols for separating confidential information from non-confidential information – eliminating the two largest risks of GenAI.

The biggest risk of all – falling behind

With any emerging technology, there is a journey to proficiency and optimisation. Organisations embracing GenAI as early adopters will achieve greater results and improvements from lessons learned faster than organisations that delay. While it is impossible to eliminate all risk, starting small but early will ultimately lead utilities to scale fast later.

As the energy landscape continues to grow more complex in the face of distributed energy resources, evolving customer expectations and a shift in the supply and demand paradigm, this advantage will become even more critical for utilities to fulfill their main mission – powering the world as we know it.