Where AI Helps in Municipal Operations, and Where It Doesn't

Software vendors have used the word "AI" loosely for years, and most public works directors have learned to ignore it. That was a reasonable habit. The language models available now are more useful than what came before, though, and when they work from the data your department already records, they can save staff a few hours a week on routine reading, sorting and writing. They do not replace anyone, and their output needs a person to check it. Below are seven uses we see in municipal operations today, followed by the limits.

1. Citizen request triage

Most cities take requests through two or three channels: a citizen portal, phone calls entered by staff, and a third-party intake tool such as SeeClickFix. Requests arrive in plain language ("There's water bubbling up at the corner of 4th and Maple"). Someone has to classify each one, set a priority and send it to the right division. In most departments that someone is a supervisor reading them one at a time.

A triage model reads the description and the photo, then suggests a category (water main break), a division (Water) and a priority (P1). It drafts a work order from that. The supervisor reviews the suggestion and confirms it in two clicks or corrects it. For common request types this cuts the handling time from a few minutes to a few seconds. The supervisor still makes the call, and unusual requests still need a careful read.

2. Plain-language data questions

Take a question like "Which assets cost us the most this quarter?" Answering it usually means a report request, a pivot table or part of an analyst's morning. An assistant connected to your operational data can answer it directly and link the work orders behind the number. The way it is set up matters. The assistant should query a defined schema with read-only access, so it cannot change records and is not guessing at table names.

When it is set up properly, this is one of the more useful AI features in a CMMS. A reporting task that took 30 minutes takes about 30 seconds, and directors who would never open a BI tool will use it. Anyone relying on the answer for a budget or council decision should still open the linked work orders and confirm the figures.

3. Asset history summaries

Each asset in your registry builds up a history of work orders, parts used, downtime and repeat failures. Reading through it takes time. A summary pulls out the pattern, for example: "Lift Station #3 has had 7 work orders in the last 90 days, 5 related to the discharge pump, and mean time between failures is down 18% from last year." That kind of summary gives a capital planning meeting something to work from, where a 200-row spreadsheet export would not. The summary is only as accurate as the work orders it reads, so it helps to spot-check it against the records.

4. Risk scoring and PM interval review

Manufacturers set PM intervals conservatively because they have to assume the worst operating conditions. After an asset has eight or ten PM cycles on record, a model can point out where the data suggests the interval is too frequent (inspections that never find a defect) or too far apart (failures between scheduled visits). The same history can feed a simple risk score that ranks assets by failure frequency and impact. The supervisor decides whether to change an interval or move an asset up the list. The model only produces the list of candidates. More on PM strategy here.

5. Voice to work order

Field crews do not like typing on a phone, and most are happy to talk into one. Speech-to-text combined with a structured prompt can turn a 30-second voice memo into a work order with the title, description, category, asset link and location filled in. The technician looks it over and saves it. In our experience this does more for mobile adoption than most other features, because it removes the part of the app crews complain about most. Paired with offline mode, it fits the way crews work at the curb.

6. Duplicate and cluster detection

Say five residents report water in the street within 30 minutes, all on the same block. Without help, that becomes five tickets, and a supervisor eventually notices they are the same problem. A duplicate detection model sees the cluster, flags it as a possible main break and suggests grouping the requests under one work order. Each resident still gets their own status updates. The supervisor confirms the grouping, since two reports on one block are sometimes two separate problems.

7. Drafting reports and procedures

Council updates, monthly division reports and regulator submissions are mostly the same data arranged differently each period. A drafting tool can pull the numbers, write a first draft of the narrative in your department's usual style and hand the director something to edit. The same approach works for turning equipment manuals into draft standard operating procedures. It saves time. Someone who knows the subject still has to read the draft closely, check the figures and fix what the model got wrong.

Limits worth stating plainly

AI tools in operations today should not be used as:

• A substitute for a dispatcher's judgment during a complicated emergency.
• A substitute for a foreman's knowledge of which crews work well together.
• A source for safety-critical decisions without a person reviewing them.
• A reason to put off cleaning up your asset registry. Poor data produces poor suggestions.

Models also make mistakes with confidence. They will occasionally misread a photo, put a request in the wrong category or state a number that is not in the data. Build review into the workflow and keep a log of what the model suggested and what staff changed.

What to ask vendors

Nearly all CMMS vendors now mention AI. These three questions help sort out which features will be useful:

1. "Can you show me the AI feature working on my data, instead of a prepared demo?"
2. "Which step in the workflow does it remove? I mean a step that goes away, not one it helps with."
3. "How is my data isolated and rate-limited, and how do you keep it out of third-party model training?"

If a vendor cannot give specific answers to all three, treat the AI claims with caution.

What comes next

Over the next couple of years, expect these tools to take on more of the workflow itself, such as drafting crew schedules, flagging unusual spending and preparing first drafts of capital plans. How useful they are will depend on the data underneath. Departments that build a clean asset registry, structured work orders and a habit of attaching photos and notes will get better results. Departments still running on spreadsheets will have little for the tools to work with.

---

WorkmanIQ includes these AI features in the base platform. See what the AI does →