The pitch arrives at every agency eventually: you have a backlog, AI can help. Sometimes that is true. But there is a precondition nobody puts on the slide. AI amplifies whatever process it sits on. If the workflow underneath is sound, automation makes it faster. If the workflow is broken, you get the broken version at scale.
The Dashboard Says 94 Percent
Here is what amplification looks like, followed all the way through. Picture a cross-connection control program whose dashboard reads 94 percent compliant. It looks like a program in good shape. But the status field flips to Compliant the moment a tester uploads a test report — not when anyone reviews it. Around 300 assemblies are sitting in that state right now: report submitted, never looked at. Some of those reports show a failed relief valve on an RPZ. A few were filed under tester certification numbers that expired in March. The utility clerk who used to work the review queue retired, and the queue went with her. Nobody knows the real compliance rate, and the dashboard is not lying exactly — it is faithfully reporting a status that stopped meaning anything.
Now put an AI assistant on top, trained to answer questions from the system. A customer asks whether their assembly is current; it says yes, citing the report on file. The director asks for the compliance number before a council meeting; it says 94 percent, in a confident, complete sentence, and the number goes into the packet. The error existed before the AI arrived. The AI just delivers it to more people, faster, in a tone that discourages anyone from double-checking. That is the failure mode — not dramatic, quiet: more output, produced faster, with less scrutiny per item.
The Honest Diagnostic
Before any AI initiative, an agency should be able to answer four questions:
- Are our records accurate enough that summarizing them is useful — or would we be summarizing the 300 unreviewed reports?
- Does each status mean one thing, or does it mean whatever the last person who set it intended?
- When work routes around the official system — into spreadsheets, inboxes, a binder on someone's desk — do we know where and why?
- When a decision is challenged, can we reconstruct who decided, on what basis, from the record alone?
If the answers are no, those are the projects. They are just not AI projects yet.
Fix the Substrate First
The unglamorous sequence works:
- Get the records clean and in one system — including the duplicates, the vacant services, and the assembly that was replaced during a remodel but still carries the old serial number
- Define statuses that match how the work actually flows, exceptions included. Submitted is not Reviewed. Reviewed is not Compliant.
- Make review and approval explicit, with names and timestamps, so the queue cannot quietly die when one person retires
- Then automate the steps that are now boring
Notice that steps one through three pay off with no AI at all. The clerk's replacement can find the real backlog on day one. The director's council number is defensible. That is the test of a real foundation: it pays for itself before the interesting part starts.
A Multiplier, Not a Workaround
AI is not a workaround for process problems. It multiplies whatever is already there. Agencies that fix the workflow first get compounding returns from automation. Agencies that skip ahead get their existing problems, delivered faster, with an invoice.