Find the board meeting before the pilot starts

A vendor and the utility's staff run a pilot, the tool does what it promised, and then the work stops. The production contract, the multi-year term and the connection to the utility's own systems all need approval, and nobody put that approval on the plan. Months can pass before a board hears it, and some pilots never get a hearing.

A new book describes the same gap in companies outside water. In The Enterprise Brain, which came out this year, Ragy Thomas, Sravan Vadigepalli and Chandhu Nair argue that AI pilots stall because companies bolt AI onto the org charts, approval chains and systems they already have. Data sits in separate systems, the rules differ from team to team, and nothing gives the AI a shared picture of how the business runs. So each new project rebuilds the same connections, data checks and approvals from scratch. I recommend it. It is a blueprint for putting AI to work in a large organization. It gets into the detail of what success takes, and it gives you a roadmap for thinking through each part of the job.

I believe a pilot with a utility is decided on its first day. A utility's approval chain is public: the board meets on a published schedule and votes in the open, so the end of a pilot is a date you can find before it starts. If the vendor and the utility agree at kickoff what "ready for production" means, and which board meeting will hear it, the pilot has a date to work toward. If that is left open, a pilot can run well for months and end nowhere.

A pilot that worked in weeks

Ragy Thomas tells the story of a CIO whose team built a procurement agent, in a Forbes column by Sandy Carter (11 August 2026). The pilot worked within weeks. Fifteen months later, the agent was still out of production. What held it up was everything around the model: 14 system connections, reconciling vendor data, approval workflows and audit requirements. As Thomas told Carter, "The model was the easiest part. The constraint was architecture."

In a YouTube interview, Thomas says that 95% of AI pilots are not hitting production.

Water already has a program built around this problem. Isle Utilities, a water technology consultancy, runs a Trial Reservoirs Initiative. Before a trial starts, the vendor and the end user sign a Trial and Purchase Agreement with "pre-agreed" success criteria, KPIs, volumes and price, so a trial that passes "leads directly to a purchase order" (Isle Executive Exchange, episode 87). Isle reports a 72.5% trial success rate against a 20 to 25% industry average. Other Isle pages say 75% and 73%, and I have not seen the figures checked independently.

Why a utility pilot waits

A utility has an approval chain by design, because it spends public money. Staff can test a tool on their own time and budget. A production contract, a multi-year term, or a connection into the CMMS or the ERP usually needs a vote, and the vote happens at a scheduled board or council meeting with a published agenda. At East Valley Water District in California, for example, the general manager can sign contracts up to $100,000, and anything at or above that needs board authorization (purchasing policy, revised 22 April 2026).

The board record of Columbia Power and Water Systems in Tennessee shows how this sets the pace, even for a program with years of work behind it. On 18 March 2026, the CPWS board approved a Master Bond Resolution for its Long-Term Water Supply Program, 5-0, and sent it on to the City Council. The utility board was one step and the council was the next. On 10 June 2026, the board approved, 5-0, the ninth guaranteed maximum price amendment with its construction manager, with a not-to-exceed budget of $390,391,412. Each step had to reach a meeting date before it could move.

That is a water supply program, far larger than any software pilot, and a software contract that needs board approval at that utility waits for the same agenda.

So a pilot runs on two clocks: how fast the tool works, and the board's calendar (the next meeting, the agenda deadline before it, and the start of the next budget year). Vendors tend to plan around the first, and the board sets the end date with the second.

To see how often the board clock ticks, we counted meeting dates in the same board minutes and agendas we read for the AI count below. Across the 210 utilities whose 2025 meeting files carry a full date, the median was 22 meeting dates in the year, close to two a month, and the middle half ranged from 11 to 32. Utility boards and districts alone had a median of 19; city and county governments, which add council and committee meetings, had 24. Either way, the meeting that could hear a production contract is a date you can look up before kickoff.

AI pilots already reach the board

Utility boards are already voting on AI pilots. At the Los Angeles County Sanitation Districts, the District No. 2 board minutes of 8 April 2026 record a decision "to extend the pilot testing of ML and/or AI tools for sewerage collection system management for a one-year period," and the board confirmed "the extension of a purchase order to Jacobs Engineering Group, Inc., in the amount of $150,000." Minutes from 12 March 2025 describe the earlier pilot: a purchase order of about $100,000, awarded in September 2023, to test Jacobs' "AquaDNA" software. At Orange County Water District, the board packet of 18 February 2026 describes testing an "AI Coach" that analyzes data from the GWRS RO plant "and makes recommendations to plant operators." A staff status report in the 2 July 2025 packet describes a 12-month trial of Pani Energy's AI/ML service at the same plant. In Omaha, a slide in the Metropolitan Utilities District's 6 May 2026 board documents lists an "AI Risk Model Pilot," with main break prediction results and cost among its topics.

Of the utility boards we read, the share whose minutes or agendas mention AI near a water topic rose from 4% in 2023 (10 utilities) to 11% in 2025 (29) and 11% in 2026 so far (30). Boards with AI and the word "pilot" in the same document went from 7 to 21. Some of those mentions are AI policy items or data-center water demand, and the boards we read are a sample, so I would not read a national rate into them. Within this sample, AI pilots are reaching public agendas, and each one has a date.

Measure how well

The book's website describes a move from asking "how fast?" to asking "how well?". I would apply that to the pilot's exit test: name a decision the utility made better, and the record that proves it, such as the right main on the replacement list, a work order closed against the correct asset, a capital plan line that staff could trace to its source.

The book also names three stages:

• Lab: "Can we build it?"
• Factory: "Can we scale it?"
• System: "This is how we work."

In a utility, each stage lands on an approval. A Lab pilot on one plant or one billing queue can often run on staff time. Factory needs money and connections, so it needs a vote. System means the work is in the standard operating procedures and in the adopted budget for a fiscal year.

The book also uses an idea from economics called the J-Curve. Picture the letter J on a chart of results over time: a short dip down, then a long climb. When a company first puts AI to work, results often get worse before they get better. Staff spend time learning the tool, old workflows break, and data problems nobody had noticed come to the surface. Costs go up while output stays flat or drops, and that is when boards start to lose patience.

The authors place the Lab stage at the bottom of the J. The climb starts once the company builds a shared layer of knowledge, rules and approved actions that every AI project can reuse, which they call the Enterprise Brain. Factory and System are the climb.

Economists have measured the dip. Erik Brynjolfsson, Daniel Rock and Chad Syverson studied general purpose technologies, the kind that change how whole industries work, AI among them. They found that each one first needs investment in new processes, skills and data, and that investment does not show up as output when it is made. So measured productivity falls first and rises later.

A utility pilot sits at the bottom of the J. Staff time, data cleanup and security reviews all come first, and the payoff arrives after the board approves production. A board that expects the dip can judge the pilot by its exit test.

What to agree at kickoff

Here is how I would set up a pilot with a utility, in order.

1. Find the meeting. Look up the utility's board or council calendar and the start of its fiscal year. Pick the meeting where a production decision could be heard, then work back from that meeting's agenda deadline. That date is the end of the pilot.
2. Write the exit test. One or two sentences: what the tool must show, how it is measured, and which record proves it. Both sides agree to it at kickoff, and the utility's project lead can take it to the board as written.
3. List what production needs. Every item from the CIO story has a water version: system connections, data reconciliation, approvals and audit. Name each one with an owner. If a CMMS or ERP connection needs a security review, start that review in week one.
4. Start where no access is needed. A pilot that runs on public records (board minutes, capital plans, RFPs, permits, SRF loans) can begin on day one with no connection to a utility system. The utility has results to look at while the IT and security work runs in parallel.

Pilots / Harrow Bend Water Works / Pilot planACME · workspace

Pilot plan: asset records pilot

Harrow Bend Water Works · kickoff 11 Jan 2027 · starts on public records

Pilot ends at the board meeting

Board of Directors, Tue 9 Mar 2027

Agenda deadlineTue 23 Feb 2027Production decisionHeard at the 9 Mar meetingFiscal year starts1 Jul 2027

Board calendar 2027 · p. 1

Exit test

By the 9 Mar meeting, staff trace each of the 12 largest lines in the draft capital plan to a dated public document, and the project lead presents the result.

Measured bySample check by utility staffRecord that proves itSource page for each lineAgreedBoth sides, at kickoff

Day-one data, no system access needed

RecordWindowStatus

Board agendas and minuteslast 24 monthsLoaded

Adopted capital plancurrent editionLoaded

Open bids and RFPsopen nowReading

Discharge permitscurrentQueued

State loan listscurrent yearQueued

Needed for production

ItemOwnerWhen

CMMS connection and security reviewUtility ITWeek 1

Board approvalProject lead9 Mar

Audit trailVendorNot started

For utility staff

Your board calendar and your budget calendar are both public. Share them with a vendor at kickoff, along with the agenda deadline and the signing limit that sends a contract to the board. A vendor who knows the earliest meeting at which a production contract can be heard will plan the pilot around it, and the board will see a result it can vote on.

For vendors

Before the next pilot starts, write down the board meeting where it ends. On Monday, open the utility's board calendar and find the meeting where your production contract could be heard. If you cannot name that meeting yet, agree on it at kickoff.

At AquaIntel we read utility board minutes and agendas every day, which is how we counted the meetings and the AI pilots above. The meeting where a pilot ends is in the same record.

Sources

• Forbes, Sandy Carter, 11 Aug 2026, column on The Enterprise Brain with comments from Ragy Thomas and Sravan Vadigepalli (CIO procurement agent: pilot in weeks, still out of production after 15 months; 14 system connections, vendor data reconciliation, approval workflows, audit).
• YouTube, "What Does UnifyApps Actually Do?" (Ragy Thomas): "95% of AI pilots are not hitting production." Speaker's claim; underlying source unknown.
• The Enterprise Brain: Rewiring Your Business for the AI-Native Era (Ragy Thomas, Sravan Vadigepalli, Chandhu Nair, 2026). ("bolting AI onto old org charts, approval chains, and workflows"); book site theenterprisebrain.ai ("how fast?" to "how well?"); Lab, Factory and System; the J-Curve.
• Isle Utilities case studies, read Oct 2026: Samoa Water Authority, Rennes Métropole, Beaurade WWTP, Yorkshire Water; and Isle Executive Exchange, episode 87 (Trial Reservoirs Initiative; Paper Excellence trial).
• Columbia Power and Water Systems, Board of Public Utilities agendas and minutes: cpws.com (18 Mar 2026 Master Bond Resolution, 5-0, sent to City Council; 10 Jun 2026 GMP Amendment 9 with Brasfield & Gorrie, 5-0, not-to-exceed $390,391,412: cpws.com).
• Los Angeles County Sanitation Districts, District No. 2 board minutes, 8 Apr 2026, p. 3 (pilot extension, Jacobs purchase order, $150,000), posted on the District No. 2 minutes page
• Los Angeles County Sanitation Districts, District No. 2 board minutes, 12 Mar 2025, p. 3 (AquaDNA pilot purchase order of about $100,000, awarded September 2023), posted on the District No. 2 minutes page
• Orange County Water District, board packet, 18 Feb 2026, p. 188 ("AI Coach" at the GWRS RO plant)
• Orange County Water District, board packet, 2 Jul 2025, p. 66 (R&D status report: 12-month trial of Pani Energy's AI/ML service)
• Metropolitan Utilities District (Omaha), board documents, 6 May 2026, p. 93 ("AI Risk Model Pilot" slide, 2026 Infrastructure Replacement Program Update)
• AquaIntel reading of public board minutes and agendas, 25 Oct 2026: 317 utilities with dated meeting documents in 2023 to 2026 (225 in 2023, 240 in 2024, 260 in 2025, 278 in 2026 to October). AI within about 300 characters of a water term: 10, 18, 29, 30 utilities; AI and "pilot" in the same document: 7, 10, 15, 21. Shares are of the boards read.
• Brynjolfsson, Rock and Syverson, "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies", American Economic Journal: Macroeconomics 13(1), 2021 (NBER w25148).
• Ragy Thomas, Forbes Technology Council posts: "What's Really Killing Enterprise AI In Production?" (16 Jul 2026) and "Why Enterprise AI ROI Is An Architecture Problem" (24 Sep 2026): fragmented data, disconnected systems and inconsistent governance; without shared context each project starts from scratch. The July post also tells the CIO procurement-agent story, before the Forbes column did.