AI in citizen service: how it works for water utilities

By Bernardo Gontijo, founder of NeftPublished on Updated on 6 min read

Summary

AI can guide residents, look up authorized information and pass a conversation to a person. For water and other utilities, results depend on clear rules, current sources and tracking what was actually resolved. Managers should start with clearly defined requests and also test failures, exceptions and situations where no team member is available.

In video · 1 minute

How AI and people serve customers together

A water outage, many questions and a team with the context to help. Watch an illustrative example of citizen service.

Audio and on-screen text in Portuguese. Portuguese, English and Spanish captions are available in the video controls. Illustrative demonstration; actions and lookups depend on the integrations authorized for each operation.

Open video in a new page

How does AI citizen service work in practice?

Service starts with a specific request, such as understanding a charge or reporting a water outage. AI identifies the topic, responds with information approved by the utility and, when an authorized integration exists, queries the relevant system. If information is missing, there is a risk or the resident asks for a person, the conversation should move to the team.

For a water utility manager, the key is organizing this journey. Putting a bot at the entrance is not enough: define what it can say, which actions require confirmation and who handles exceptions. Explaining document requirements is different from changing account details, negotiating a debt or promising a reconnection.

An example: from a water outage to the right next step

Consider a fictional situation: a resident contacts the service because she has no water. AI asks what it needs to identify the area and checks an authorized operational source. If there is a current outage notice, it explains the confirmed facts and provides the recorded estimate, when available, without inventing a time to reassure her.

If there is no known incident, the response must make that clear. Depending on the operation's rules and integrations, the next step may be collecting information for review, opening a request or transferring the case. Confirmation that a request was opened should appear only after the responsible system returns its record. Understanding the complaint does not mean a work order was created.

The interaction ends with a clear outcome: information provided, a request recorded, a transfer or a pending follow-up. The resident needs to know what happens next. Management must distinguish a closed conversation from a resolved problem. This matters especially during peaks, when many people contact the service center about the same event.

Which requests make a good starting point

Start with common, documented topics with clear boundaries. Guidance on official channels, required documents or the stages of a request generally needs less autonomy than changing data. Even a simple question depends on accurate content: someone must be responsible for updating hours, procedures and notices.

Bill copies, reference-number lookups and request status checks may require identity verification and access to the customer management system. Before offering these features, confirm which integrations exist, which data may be accessed and how to prevent one person from receiving information about another service account. A demo with fictional data is not proof of a completed integration.

Debt arrangements, disputed charges, account-holder changes and urgent situations need their own assessment. AI can help organize the account of what happened, but operational decisions must follow approved rules. For energy and gas, adapt the design to the utility's procedures and emergency channels; a water service workflow does not automatically fit every essential service.

Human handoff must be part of the design

Define the path to a person before launch. Set the triggers: an explicit request from the resident, an out-of-scope topic, conflicting information, a failed lookup or repetition without progress. The handoff should include the reason for contact, what has already been explained and what remains unresolved.

Also define what happens when no agent is available. Explain the actual situation and the alternative provided by the operation, such as a callback or appointment. Do not promise immediate human service if the team is unavailable. Managers should test this scenario outside business hours, as well as the ideal journey where everything works.

How this workflow works in Neft

In Neft's AI customer service solution, the virtual assistant speaks through the browser and shares the team's queue. The operation defines when it handles calls and which topics and actions are allowed. When a person takes over, the interaction keeps its reference number and includes a conversation summary.

There are important limits to consider: telephone and WhatsApp are not part of this module. Entry requests microphone and camera access, but AI communicates by voice and does not analyze images. Scheduled appointments remain with the team. Cover these points in the demo to avoid confusing browser-based voice service with an automated telephone contact center.

The water and utilities page describes the sector use case and CORSAN's public video service case study. That case demonstrates video use in the sector; it is not evidence of an AI deployment at the company. The examples in this article are illustrative and do not describe customers' internal processes or results.

What to check before scaling the operation

Choose one service for the first test and write observable criteria. For a reference-number lookup, for example, the answer must match the authorized record. If the lookup fails, AI should acknowledge the limitation and explain how to proceed. Also test incomplete requests, topic changes and requests that must be declined or routed elsewhere.

Compare results with the current operation using the same request types and time period. Track confirmed resolutions, transfers, abandonment, repeat contacts for the same reason and time to resolution. Review a sample of conversations to find convincing answers that resolved nothing. Do not focus solely on reducing average handling time: a quick closure may move work to another channel.

Assign responsibility for reviewing the knowledge base, fixing errors and stopping problematic actions. Before using real data, have the responsible teams validate access controls, purpose, retention and vendor processing. Starting with fictional scenarios lets you assess the workflow without exposing residents' information. Our article on AI reliability explores how to distinguish a plausible answer from a verified result.

Bring a real operational request to the demo

A good demo should cover the start of contact, an allowed lookup, a failure and the handoff to the team. Bring the service name, rules and expected outcome, using fictional data. Also ask to see the final record and how managers identify cases that still need a follow-up.

Explore Neft AI customer service and request a demo. The conversation helps assess which requests suit automation, which integrations need verification and where people should remain involved. The goal is to give residents a reliable next step and the operation a way to track the result.

Frequently asked questions

Can AI promise a reconnection deadline?
It should only provide a deadline confirmed by the authorized operational source and within the utility's rules. Without confirmation, it must explain the limitation and route the request onward.
Does AI replace every service channel?
No. The design should preserve alternatives that suit the audience and services. In Neft, this module provides voice service in the browser; it does not include telephone or WhatsApp.
How can you tell whether the test worked?
Check the result in the responsible system as well as the conversation. Assess errors, transfers and repeat contacts for the same reason before expanding the scope.

Sources

Videos, studies and documents used in this article. Article prepared with AI assistance and reviewed by Neft. Public information checked on November 16, 2026; fictional examples and editorial recommendations, with no customer results.

  1. Reference page

    Neft AI customer service (opens in a new tab)

    Neft · accessed on

    Source for the product features, channels and limits described in this article.

  2. Reference page

    Water and utilities (opens in a new tab)

    Neft · accessed on

    Sector context and a public reference to CORSAN's video service case, without attributing an AI deployment to the customer.

Want to evaluate AI for citizen service?

Request a demo and assess the workflow with fictional data from your operation.