By Sagar Shankaran, Founder of CallSphere
How a 4,200-connection water system uses Claude Cowork or ChatGPT Work to draft its Consumer Confidence Report, with a source cited for every number in it.
Key takeaways
That objection is correct and should not be waved away. Plenty of compliance coordinators fed a chat box a pile of lab results in 2024, asked for a Consumer Confidence Report, and got back a beautifully formatted document with an MCL that did not match 40 CFR Part 141, a range that averaged the wrong quarters, and mandatory health effects language that had been paraphrased — itself a violation, since that language must appear verbatim. Everybody agreed the machine was not ready. That was a fair verdict on a chat box.
What shipped in 2026 is not a chat box. Anthropic released Claude Cowork on 12 January 2026 and expanded it to mobile and web in July; OpenAI released ChatGPT Work on GPT-5.6 on 9 July 2026. Both do the same fundamentally different thing: you give a goal instead of a question, they connect to your files and applications, break the job into steps, work for hours on their own, and hand back finished work. The difference that matters to a water system is that the output is a draft document with a source listed for every number in it, not a paragraph of confident prose you have to fact-check from scratch.
Both were aimed at non-technical staff, which in a 4,200-connection district means the compliance coordinator who is also the deputy clerk, and who has never written a line of code and never will.
The report has to reach every customer by 1 July and the certification has to reach the primacy agency by 1 October. That sounds like a small annual chore until you watch it get built. Somebody exports twelve months of results out of WaterTrax or the lab's portal, or types them off PDF lab reports. Somebody cross-checks those against the monthly operating reports and against what the state has in SDWIS, because the two do not always agree. Somebody builds the detected-contaminant table: contaminant, unit, MCL, MCLG, range, average or running annual average depending on the rule, likely source, and the right footnote for anything waived or on a reduced schedule.
Then the required-language pieces. The lead paragraph. The Cryptosporidium and radon statements if they apply. The immunocompromised persons statement. The Spanish notice — "Este informe contiene información muy importante sobre su agua potable" — where the population requires it. In 2026 the lead service line inventory status has to be reflected honestly, and PFAS results from the fifth Unregulated Contaminant Monitoring Rule round are on many systems' tables for the first time, which means customers who never thought about their water are going to read a number and call.
Call it two weeks of a coordinator's June spread across four other jobs, plus the chief operator's evenings checking figures, plus layout and the board packet — in the same month as fiscal-year budget adoption and the start of peak irrigation season.
The instruction now looks like this, typed in plain English: build this year's Consumer Confidence Report draft, using the lab results in this folder, last year's approved report as the format, the monthly operating reports in this drive, and the state's template; fill every table cell from a source you can name; leave any cell you cannot source blank and list it separately; do not paraphrase required language.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent in your browser — 60 seconds, no signup.
Then it works. It opens the lab PDFs, reads the results, matches them to the right rule, builds the table, carries the language blocks over verbatim from the state template, and hands back a draft plus a list: here are the eleven numbers I could not tie to a source document.
flowchart TD
A["Compliance coordinator states the goal, points at the folders"] --> B["Agent reads 12 months of lab results and the monthly operating reports"]
B --> C["Fills the detected-contaminant table: MCL, MCLG, range, average, likely source"]
C --> D["Carries required lead, PFAS and health-effects language over verbatim"]
D --> E["Returns the draft plus a source citation for every number"]
E --> F{"Chief operator checks each figure against the lab report"}
F -->|Figure does not tie out| B
F -->|Ties out| G["GM signs, mail house prints, certification filed by 1 October"]
The coordinator drops the year's lab PDFs into one folder, points at the drive with the monthly operating reports, attaches last year's approved report and the state's current template, and states the goal. Then she goes and does the billing adjustments she has been putting off.
At 11:15 she has a draft. Total trihalomethanes shows the locational running annual average by site, correctly, with the third-quarter number that is always highest because warm water at the far end of the distribution system makes more disinfection byproducts. Lead and copper show the 90th percentile from the last round with the correct sample count. The PFAS row carries the detection and the advisory language exactly as written. And the exceptions list says: no source document for the September nitrate sample, the source water assessment summary is from 2019, and the population served figure disagrees between last year's report and the latest monthly operating report.
Those three exceptions are the real product. Each would otherwise have been quietly wrong in the final report, and the population figure is exactly the kind of discrepancy a state reviewer circles.
Assumptions: the coordinator is loaded at $38 an hour, the chief operator at $52, the report takes 34 coordinator hours and 9 operator hours today, and a drafted packet cuts the coordinator to 9 hours of review and the operator to 5 of verification. Assume $600 a year for the seat that does this and several other jobs.
| Task | Hours today | Hours with a drafted packet |
| Pulling and reconciling results | 16 | 2 |
| Building the detected-contaminant table | 11 | 3 |
| Required language and footnotes | 7 | 4 |
| Chief operator verification | 9 | 5 |
| Labor cost | $1,762 | $602 |
About $1,160 saved on one document, against a $600 seat that also drafts the monthly operating report summary, the board packet and the annual water loss audit worksheet. The hours are not really the point, though. A late or wrong report is a monitoring and reporting violation that goes on the compliance record and into next year's report as a violation you have to disclose to every customer. That is the cost this actually avoids.
The habit to break is handing out task lists. Managers who have run a utility for twenty years assign work in steps: pull the results, build the table, check it, send it to layout. Handing over a goal means describing the finished thing and the rules it has to satisfy, then leaving it alone for two hours. It feels like less control and it is more, because you get back a draft with an argument for every number instead of a folder of half-finished spreadsheets.
Three habits make the difference. Name the source of truth explicitly — the lab report is authoritative, not the spreadsheet somebody retyped it into. Require a citation per figure and reject the draft if any cell cannot point at a document. And keep the approval gate where the regulation already puts it: the ORC verifies the numbers, the general manager signs. The signature is a legal act and it never moves.
Still reading? Stop comparing — try CallSphere live.
CallSphere ships complete AI voice agents per industry — 14 tools for healthcare, 10 agents for real estate, 4 specialists for salons. See how it actually handles a call before you book a demo.
Judgment calls about what counts as a violation, what triggers Tier 1, Tier 2 or Tier 3 public notice, and how to describe an exceedance to customers stays human — with the person who will be at the microphone if a councilmember asks about it.
Anything handwritten is a risk. Field sheets, chlorine residual logs on clipboards, and the operator's marginal note that a sample came from the alternate site because the primary was behind a locked gate get read the way a new hire reads them, without context. Verify anything that came off paper.
And tone matters more than people expect. A report disclosing a PFAS detection needs a paragraph a homeowner in your town will believe, written by someone who knows whether this is the year the district is asking for a rate increase. That paragraph is yours.
Do not let it. Preparing the certification form, the mailing list and the proof-of-delivery documentation is fine. The submission itself is a signed statement about the accuracy of your report to your primacy agency, and the person whose name is on it should click the button.
Mostly no — reading numbers off scanned lab reports got dramatically better through 2025 and 2026. But it is exactly where you verify hardest. Spot-check every result near a limit and every result on a report that arrived as a fax or a scan.
Yes, and that is the bigger win. The monthly operating report, the discharge monitoring report on the wastewater side, the quarterly disinfection byproduct summary and the annual water loss audit are the same shape of job: pull numbers from known sources, put them in a fixed form, flag what does not tie out. Start with the annual one because you have time to check it.
Then this belongs to the clerk and the ORC and it gets done at night. Those systems get the most back, and they are the ones most likely to carry a monitoring and reporting violation because nobody had four uninterrupted hours in June.
Take last year's approved report, ask for a rebuild from the same source files, and compare it line by line against what you actually mailed. You know the right answers already, so any drift shows up immediately and you learn what to verify hardest before it matters. One afternoon settles the 2024 objection either way.
One adjacent note: the week the report mails, the office phone rings with people asking what haloacetic acids are, whether they should stop drinking the water, and whether their house has a lead service line. CallSphere builds AI voice and chat agents that answer the utility's phone and website chat around the clock, respond from the language you approved in the report, look up whether an address is in the lead service line inventory, and pass anything sensitive to staff with the address and callback captured. It does not write your report and it does not certify anything. It absorbs the week of calls the report creates.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
The monthly IEEE 1366 reliability close takes 64 hours across three people. What goal-driven agents change, the arithmetic, and what stays with the engineer.
Payouts, oversells, returns aging and chargeback deadlines, prepared between 9pm and 6am. What a DTC ops manager reviews instead of rebuilds each morning.
How water systems use a cheap-model first pass and a strong-model escalation so nobody drives out at 2 a.m. for a SCADA alarm that already cleared itself.
How pest control service managers hand the monthly food-account trend packet to a 2026 work agent as a goal - and what has to change about assigning work.
The phased plan, insurance estimate, predetermination narrative and financing page, finished before the patient leaves. What the owner has to change to get it.
A monthly drone flight over the standpipe, lagoon berms and booster station catches the deficiencies a sanitary survey would otherwise find three years late.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI