Utilities are increasingly using conversational AI to handle routine customer-service tasks such as billing questions, outage updates, service requests, payment guidance, and account navigation. For electricity, gas, and water providers, that can be valuable because contact centers often experience sharp demand spikes during storms, billing cycles, rate changes, or service disruptions. But a utility chatbot is not just another website widget. Utilities operate essential services and often serve customers with disabilities, limited English proficiency, low digital confidence, intermittent connectivity, or urgent safety needs. A chatbot that works well for a retail store may be unacceptable when a customer is reporting a gas smell, a downed power line, a medical vulnerability, or loss of water service. The most useful goal is therefore not “replace the contact center with AI.” It is to build an inclusive digital service channel that resolves low-risk tasks efficiently, recognizes when automation is inappropriate, and hands people to a human or emergency channel quickly when necessary.
What Makes a Utility Chatbot Inclusive
An inclusive chatbot is a conversational interface designed so that as many customers as reasonably possible can use it successfully, regardless of disability, language, device, digital literacy, or communication preference. It should also avoid creating a second-class experience for people who cannot or do not want to use the bot. In practice, inclusive Chatbots for utilities should be built around several principles: Accessible interaction by keyboard and assistive technologies; Clear, plain-language responses; Multilingual support where the customer base requires it; Low-friction human escalation; Strong privacy and authentication controls; Safe treatment of emergency and high-risk situations; Transparent disclosure when the user is interacting with AI; Reliable fallback when the model is uncertain.
Why Utilities Are a Strong Use Case for Conversational AI
Utility companies receive large volumes of repetitive questions. Many do not require a specialist to answer. Common requests include: “When is my bill due?”; “Why is my bill higher this month?”; “How do I update my mailing address?”; “Is there an outage in my area?”; “How do I start or stop service?”; “Where can I find my usage history?”; “How do I enroll in paperless billing?”; “What payment methods are available?”. These are suitable automation candidates because the requested information can often be retrieved from a controlled system or a well-maintained knowledge base. Automating them can reduce queue times and free human agents for disputes, hardship cases, safety questions, complex account changes, and vulnerable-customer support.
Where Chatbots Can Improve the Customer Experience
Billing explanations. A well-designed bot can help customers understand billing components, locate previous bills, compare recent usage, and explain common reasons for a change. It should distinguish between general education and an account-specific determination. If a customer disputes a meter reading, alleges billing error, or needs a formal review, the bot should route the case into the utility’s established resolution process rather than inventing an answer. Outage information. During a widespread outage, customers often want the same basic information: whether the outage is known, whether a crew has been assigned, and whether an estimated restoration time exists. A conversational interface can surface information from the utility’s outage-management system without requiring the customer to navigate multiple pages. It should never fabricate an estimated restoration time when none has been issued.
Service start, stop, and transfer. Moving-related requests can often be guided through a structured workflow. The bot can collect non-sensitive preliminary details, explain documentation requirements, and transfer the customer into an authenticated process for the actual account change. Usage and conservation guidance. Customers may ask why electricity, gas, or water use has increased. The bot can explain seasonal patterns, common household causes, conservation programs, and available usage tools. It should not make unsafe technical recommendations about electrical systems, gas appliances, plumbing, or medical equipment. Those cases require qualified professionals or appropriate utility teams. Payment and assistance programs. Chatbots can direct customers to official payment options, payment plans, energy-assistance programs, or hardship resources. They should be particularly careful with financial information and should never ask a customer to send card details, bank credentials, or passwords in an ordinary chat message.
Accessibility should be treated as a product requirement rather than a separate compliance screen at launch. W3C: Web Content Accessibility Guidelines (WCAG) 2.2 provides the normative accessibility standard, and W3C: What’s New in WCAG 2.2 highlights additions such as accessible authentication, consistent help, target size, and reduced redundant entry. Those ideas map directly onto utility chat experiences used by customers with visual, motor, cognitive, or authentication-related barriers.
Accessibility, Language, and Digital Inclusion
The Web Content Accessibility Guidelines (WCAG) 2.2 provide a useful baseline for customer-facing digital services. W3C’s guidance covers issues such as keyboard accessibility, readable content, predictable navigation, accessible authentication, status messages, and consistent access to help. For a utility chatbot, practical accessibility questions include: Can the entire conversation be operated without a mouse?; Does a screen reader announce new bot messages?; Is focus moved predictably when the chat opens?; Are buttons large enough and clearly labeled?; Can users zoom text without breaking the interface?; Are color and icons supplemented by text?; Is the human-help option consistently located?; Can authentication be completed without inaccessible cognitive tests?. A chatbot can technically “exist on an accessible website” while the chat experience itself remains unusable. Test the actual end-to-end interaction with assistive technologies and with people who have disabilities.
Multilingual support is more than translation. Utilities often serve linguistically diverse communities. Generative AI can make it easier to support more languages, but machine translation alone does not guarantee an accurate or culturally appropriate service. Important controls include: Use approved terminology for billing, safety, and regulatory concepts.; Test high-volume languages with fluent human reviewers.; Preserve account numbers, dates, units, and monetary values accurately.; Identify when a mixed-language or dialect request is not understood.; Provide a multilingual human escalation path.; Do not automatically translate emergency instructions without validation. The bot should also recognize that language preference may change during the conversation. A customer may begin in English and switch to another language when explaining a complex problem.
Low-bandwidth and low-digital-literacy customers. Inclusion also means designing for older devices, slow connections, and users who are not comfortable with complex digital interfaces. A utility chatbot should avoid unnecessary animations, oversized media, and interfaces that require precise gestures. Simple text, clear choices, short steps, and the ability to repeat or rephrase instructions can make the service more usable. When a customer repeatedly fails a step, the system should offer another route rather than looping the same instruction indefinitely.
Human Escalation and Emergency Safeguards
A chatbot becomes frustrating when it acts as a barrier between the customer and a person. Good automation should recognize when human judgment is more appropriate. Escalation triggers can include: Repeated misunderstanding; Customer explicitly asks for a human; Billing dispute; Disconnection or shutoff concern; Financial hardship; Accessibility accommodation request; Safety concern; Potential fraud or account takeover; Medical-equipment dependency; Complaint requiring regulatory handling. When escalation occurs, pass useful context to the agent with the customer’s permission so the person does not have to repeat the entire conversation. Emergency situations must bypass ordinary chatbot flows. This is one of the most important design rules for utilities. A chatbot should recognize emergency terms and immediately direct customers to the correct emergency channel. Examples include: Smell of gas; Suspected gas leak; Downed or sparking power line; Electrical fire; Flooding near electrical equipment; Contaminated drinking water alert; Immediate danger to life or property. The chatbot should not attempt an extended troubleshooting conversation in these scenarios. It should provide concise, pre-approved safety instructions and the appropriate emergency number or emergency-services direction.
Risk management should be explicit when a utility adds generative AI. NIST: AI Risk Management Framework offers a general structure for governing, mapping, measuring, and managing AI risk, while NIST: Generative AI Profile addresses risks that are particularly relevant to generative systems, including confabulation, privacy, content provenance, and human oversight. For a utility, this supports a design in which billing amounts, outage status, payment instructions, and account-specific answers come from approved systems rather than model improvisation.
Generative AI Needs Grounding, Guardrails, and Authentication
Traditional rule-based bots answer from predefined decision trees. Generative AI can interpret more natural language and produce more flexible responses, but it can also generate plausible-sounding information that is wrong. NIST’s AI Risk Management Framework and its Generative AI Profile emphasize structured risk management across the AI lifecycle. NIST also announced work in 2026 on a profile specifically addressing trustworthy AI in critical infrastructure, which is relevant to utility operators. For customer service, important generative-AI risks include: Hallucinated policy or outage information; Inconsistent responses to similar questions; Disclosure of sensitive information; Prompt-injection attacks; Bias across languages or customer groups; Overconfidence when the model lacks data; Unsafe technical advice. Use retrieval and approved systems instead of letting the model guess. For high-value utility use cases, the model should retrieve information from authoritative sources rather than rely on general training data. Examples: Outage status from the outage-management system; Account balance from the billing system; Rate information from an approved tariff database; Service rules from the current knowledge base; Payment-assistance programs from approved program records. If the authoritative system does not return an answer, the bot should say that the information is unavailable and offer the next appropriate step. Authentication should match the risk of the task. Not every question requires identity verification. A customer asking for the outage map should not be forced through the same authentication as someone changing a mailing address or accessing account history.
| Task | Typical authentication need |
|---|---|
| General rate explanation | None |
| Public outage status | None or location input |
| Account balance | Authenticated |
| Change service address | Strong authentication |
| Payment method change | Strong authentication and secure payment flow |
| Password reset | Dedicated identity-recovery process |
Do not collect more personal information in the chat transcript than is necessary.
Privacy, Payments, and Customer Trust
Utility accounts can contain names, addresses, phone numbers, usage patterns, payment information, and other sensitive data. Conversational systems need clear rules for what is stored and who can access it. Good governance includes: Data minimization; Retention limits; Encryption in transit and at rest; Role-based access; Audit logging; Redaction of payment credentials and secrets; Vendor data-use restrictions; Testing environments that do not expose real customer data unnecessarily. Do not send customers to unsafe payment experiences. A chatbot can guide a customer toward payment, but the actual transaction should occur through an authenticated, secure payment system. Avoid asking users to type full card numbers, banking credentials, one-time passwords, or security codes directly into a general conversational interface unless that interface is specifically designed and certified for the payment process. The bot should clearly identify the utility’s official domain and warn customers about common impersonation scams when appropriate.
Chatbots Can Support Employees as Well as Customers
Conversational AI does not have to face customers directly. An internal assistant can help contact-center agents search policies, find troubleshooting scripts, locate current rate information, and summarize account history. This “agent assist” model can be lower risk because a trained employee reviews the answer before it reaches the customer. It can also reduce average handling time without removing human judgment. Historical chatbot forecasts should not be confused with current evidence. Older industry articles often repeated aggressive forecasts about chatbot adoption. For example, one widely circulated 2019 forecast projected a chatbot market value of $5,638.64 million by 2023. Another article cited a projection of 800 million users. Those figures are historical forecasts, not current 2026 measurements. They are useful as examples of how rapidly the market was expected to grow, but utility leaders should base investment decisions on current operational evidence from their own customer journeys rather than old market projections.
How Utilities Should Measure Chatbot Performance
A chatbot should not be judged only by the number of conversations it handles. Useful metrics include: Successful task completion rate; Containment rate for genuinely suitable tasks; Human escalation rate; Repeat-contact rate; Customer satisfaction; Accessibility failures; Language-specific success rates; Incorrect-answer rate; Safety-trigger accuracy; Average time to resolution; Complaint rate. A high containment rate can actually be a bad sign if customers are being prevented from reaching human help. Test for different customer groups. Before wide deployment, test with scenarios representing real customers:
Screen-reader user; Keyboard-only user; Older adult with low digital confidence; Customer using a small phone; Customer with limited English proficiency; Customer in financial hardship; Customer reporting an outage; Customer reporting an emergency; Customer who repeatedly changes the question. Testing should include both expected flows and adversarial or confusing inputs.
A Safer Rollout Model for Utility Chatbots
Phase 1: low-risk informational questions. Start with public information such as office hours, general billing definitions, and outage links. Phase 2: authenticated account lookup. Add controlled retrieval of balances, usage, or service status. Phase 3: guided transactions. Introduce service requests, payment-plan guidance, or account updates with strong validation. Phase 4: generative assistance. Use generative AI for interpretation and summarization while grounding responses in approved data sources. At every stage, monitor failures and maintain an easy path to human support.
Conclusion
Inclusive conversational AI can improve utility service when it handles the right tasks: repetitive, low-risk requests with reliable underlying data. It can reduce wait times, extend service hours, support more languages, and help customers navigate complex account information. But utilities should not optimize for automation at any cost. Accessibility, privacy, emergency handling, human escalation, authentication, and factual reliability are essential. The most successful chatbot is not the one that prevents the most calls; it is the one that helps the customer reach a correct and safe outcome through the most appropriate channel.