Water Utility Staff Are Already Using Generative AI; Nearly Half Of Utilities Have No Policy For It

When I run an AI session with a utility’s leadership team, I start with two questions. Who in the room has used ChatGPT, Copilot, or a similar tool for work in the past month? Most of the hands go up. Next - who can tell me what the utility’s policy says about it? Usually one or two hands stay up, and one of them belongs to the IT director.

The industry numbers say the same thing. AWWA added a generative AI section to its 2026 State of the Water Industry survey of 2,171 water professionals. Just over half (56%) expect some positive impact from the technology. Nearly as many (49%) work at organizations with no established policy governing AI use: 26% have no formal policies or procedures and 23% are still writing them. Another 15% didn’t know.

So the technology is already inside the utility, on personal accounts and free tiers, and the rules haven’t caught up. Not surprising for a regulated industry, and certainly less so for water which we know to be notoriously slow on policies and procedures for the use of new technology. This post covers where utility staff are getting value from generative AI today, where it should stay out, what a use policy should say, and what to pilot first.

If you’re evaluating machine learning or A.I. products for operations (pipe failure prediction or chemical dosing, for example), my article on what to buy when every vendor claims AI covers that side. This one is about the general-purpose tools your staff already have open on their desktop.

Where utility staff are using AI today

Generative AI is good at working with text: drafting it, summarizing it, searching it, and reformatting it. A utility produces and consumes a lot of text, and most of the early wins are in the office and the training room.

Use	Where it’s running	What it needs
Drafting and summarizing	DC Water (Microsoft Copilot)	An enterprise license and staff training
Answers from the utility’s own documents	Denver Water (plant procedure manuals)	Current, digitized SOPs and O&M manuals
Capturing what senior operators know	WRF pilot, DC Water training tool	Interview time with the operator before they retire
Plan and submittal review	Denver Water (early stage)	Written standards the tool can check against

Drafting and summarizing. Board memos, council reports, grant narratives, job descriptions, customer letters, and the summary of a 300-page consultant deliverable. DC Water rolled out Microsoft Copilot with an eight-week, role-based training program, and the case study its training partner published reports that 77% of surveyed users open it five or more times a week. One employee described asking it to summarize a 200-page report and build a presentation from it, a job that used to take two days. That’s a vendor-published study, so take the numbers with a (large) grain of salt, but the pattern matches what I see at other utilities.

Answers from the utility’s own documents. Every plant has binders of standard operating procedures and O&M manuals that nobody can search at 2 a.m. Denver Water licensed Copilot for the whole organization with its data secured internally, and built an agent on the procedural manuals for its Northwest Water Treatment Plant so staff can ask for an operating instruction and get it. The same approach works on a consent decree, a permit, or a proposed rule: load the document and ask how it affects your operation.

Capturing what senior operators know. The Water Research Foundation built a chatbot called “What Would Jerry Do?: Chlorine” from hours of interviews with Jerry Kemp, an 82-year-old operator with more than four decades of experience who rural Georgia utilities kept calling for advice after he retired. DC Water has an internal training tool that organizes videos of its experts, SCADA operators among them, into searchable content and quizzes. For a utility facing a wave of retirements, this is the use I’d put near the top of the list; the interviews are valuable even if the chatbot never gets built.

Plan and submittal review. Denver Water is also testing an agent that gives developers instant feedback on construction plans submitted for approval. It’s early, and a licensed engineer still makes the call, but a first-pass check against written standards is the kind of repetitive reading these tools handle well.

Where it stays out

Generative AI should not make or execute operational decisions on a treatment or distribution system.

CISA, the NSA, the FBI, and partner agencies in six other countries published joint guidance on AI in operational technology in December 2025, and it’s blunt on this point. Because these models can fabricate plausible answers, the guidance says large language models “almost certainly should not be used to make safety decisions for OT environments.” It places them in the business layers of the network, working on data that has been exported from the control system, and it recommends push-based architectures where data leaves the OT network one way and the AI tool has no persistent access back in.

In practice that gives you a pretty clear rule to follow; a language model can read last month’s SCADA export and draft the compliance report. It can explain an alarm code from the manual. But the control network and its setpoints stay off limits. If a vendor proposes otherwise, the vendor cybersecurity review applies in full.

What the AI use policy should say

A workable policy fits on one or two pages. It should answer five questions.

Which tools are approved? An enterprise license (Microsoft Copilot, ChatGPT Enterprise, or a comparable product) comes with contract terms about how your data is stored and whether it’s used to train the vendor’s models. A personal account on a free tier generally doesn’t give the utility those protections. The practical way to get staff off personal accounts is to give them a sanctioned tool that’s as good or better.

What never goes into a prompt? Customer account and billing data, personnel records, network diagrams, SCADA configurations, and anything from your risk and resilience assessment or emergency response plan. If the information would be withheld from a public records request for security reasons, it stays out of any tool the utility hasn’t cleared for it.

Who is responsible for the output? The employee who uses it, the same as if they’d written it. When Cascade PBS and KNKX obtained thousands of pages of ChatGPT logs from Washington city officials, they found the tool had fabricated airport passenger data for one city’s comprehensive plan update and made repeated errors analyzing housing cost percentages for another. A number in a board report or a consumer confidence report has to be checked against the source by a person.

Are prompts and outputs public records? For a public utility, plan on yes. The same reporting exists because those chat logs were disclosable. Washington’s Municipal Research and Services Center concluded that outputs used in government work are public records and prompts most likely are too, and Seattle’s policy requires employees to be able to retrieve their prompts and outputs on request. Every state’s law is different, so have your records officer and counsel write this section.

Who decides on new uses? Select a person or a small group that keeps an inventory of how the utility uses AI and reviews new requests. Without one, every department buys its own tool and nobody knows what data went where.

You don’t have to write it from scratch. The GovAI Coalition, started by the City of San Jose, publishes free policy templates for public agencies, along with a vendor fact sheet and an incident response plan, and they’re aligned with the NIST AI Risk Management Framework.

What to pilot first

Pick a first project where the work is mostly text, the source documents belong to the utility, a wrong answer is easy to catch, and nothing physical happens as a result. Good candidates:

A procedures assistant for one plant. Load the current SOPs and O&M manuals for a single facility and let operators query them. You’ll learn quickly which manuals are out of date, which is useful on its own.

A knowledge-capture project with one senior operator. Record structured interviews with the person everyone calls when something goes wrong. Transcribe them and make them searchable.

Regulatory and board document support. Summaries of proposed rules and first drafts of staff reports, each reviewed by the person who would’ve written it by hand.

Whichever you choose, set the measure of success before you start: hours saved on a specific task, or time for a new operator to find a procedure. Budget for training alongside the licenses. In the DC Water study, 93% of surveyed users said training was essential to using the tool well, and PPIC’s review of California water agencies reached the same conclusion; a group of Orange County agencies has since launched an AI course built for water professionals.

Give the pilot ninety days. At the end you’ll have a measured result and a group of staff who know what the tool is good for at your utility.

Frequently asked questions

Can water utility employees use ChatGPT for work?

That depends on the utility’s policy, and nearly half of utilities in AWWA’s 2026 survey don’t have one. Where staff are using it without a policy, the main exposures are sensitive information entered into a personal account and unverified output ending up in an official document. An approved enterprise tool and a short written policy address both.

Are AI prompts public records for a public utility?

In many states they likely are. Washington’s local government advisory body concluded that outputs used in government business are public records and that prompts most likely are as well, and reporters have already obtained thousands of pages of city officials’ ChatGPT logs through records requests. Check with your records officer and counsel on your state’s law and your retention schedule.

Should generative AI be connected to SCADA?

No. Joint guidance from CISA, the NSA, the FBI, and international partners says large language models almost certainly should not be used to make safety decisions in operational technology. A language model can work on data exported from the control system; it shouldn’t have a path back into it.

What is a good first generative AI project for a water utility?

One that works on the utility’s own documents and keeps a person reviewing the result: a searchable assistant for one plant’s procedures, a knowledge-capture project with a senior operator, or drafting support for regulatory summaries and board reports.

---

HydroKnowledge helps water utilities adopt AI in a way that fits their staff and their risk tolerance. Get in touch if you want help writing the use policy or choosing and scoping a first pilot.