The Twenty-Three Wells That Did Not Get the Update

The software engineer at Baker Hughes in Florence opened the final deployment manifest on 27 January 2026 and ran a query against the asset list Expand Energy had pushed three weeks earlier. The count came back at 5,782 wells flagged for Leucipa rollout across the Marcellus, Utica, and Haynesville formations. Twenty-three wells that had been in the scoping spreadsheet in November were missing. They were still producing. They were still Expand wells. They simply did not fit the data schema the team had locked in December to ship the multi-year AI-powered production platform on time.

The missing wells were not a bug. They were the cost of shipping at scale.

Expand Energy held a working interest in approximately 6,600 wells as of December 2025, operating 5,800 of them. The January 2026 Baker Hughes announcement committed to deploying Leucipa, an AI-powered automated field production solution, across thousands of gas wells in the Marcellus, Utica, and Haynesville shales. The press statement made no mention of wells left out. Expand Energy would also pilot Lucy, a generative AI production assistant providing real-time analysis through a conversational interface. The commercial announcement landed well. The equity analysts noted operational agility and margin opportunity. Nobody asked about coverage.

The engineer knew why the 23 wells were excluded, because she had written the parser that failed on them. Leucipa ingests production telemetry from SCADA systems, pressure and temperature sensors, and flow meters installed at the wellhead. The data arrives in 15-minute intervals in a format specified by the sensor vendor. Across Expand's three operating regions, Baker Hughes had mapped six distinct SCADA vendors, four flow-meter protocols, and 11 variations in how timestamped pressure readings were serialized. The December schema handled 10 of those variations. The 11th, used by a legacy Honeywell deployment still running at 23 Marcellus wells in Greene County, Pennsylvania, embedded the timestamp inside a proprietary binary header that required a parsing library Honeywell had deprecated in 2021. The library still ran under a commercial license that cost $48,000 per deployment and required manual approval from a Honeywell vice-president who had retired in May 2024.

The team had three options in December. Write a bespoke parser for the Honeywell format and absorb six weeks of delay while legal negotiated the deprecated license. Require Expand to upgrade the SCADA units at those 23 wells before Leucipa went live, adding $340,000 in capital spend and another four-week delay for the field work. Or ship without them and flag them for phase two. The third option kept the January announcement date intact. It also kept the 23 wells on a manual monitoring cadence while the other 5,782 got real-time AI optimization.

The trade was never discussed outside the engineering standups. It did not appear in the slide deck the product team showed Expand's COO in the December readiness review. The rig count across the three shales had been rising through 2025, and Expand completed 272 gross wells as operator during the year. Every well the AI touched earlier meant more data for the model to learn on, which meant tighter optimization loops on artificial lift, faster detection of compressor-station anomalies, and better gas-lift allocation when Haynesville wells started to load. Shipping in January with 5,782 wells beat shipping in March with 5,805.

The decision was correct in a way that will not survive the first board-level post-mortem if one of those 23 wells experiences an unplanned shut-in that Leucipa would have flagged 40 minutes earlier. The data is there. The sensors work. The telemetry reaches an on-premise historian every 15 minutes. It simply does not reach the AWS environment where Leucipa runs, because the message bus cannot parse the header. Leucipa was delivered as software-as-a-service hosted on Amazon Web Services, which meant the parsing happens at ingest, not at the wellhead. A fix deployed to AWS in March will not rewrite the Honeywell units installed in 2019.

The engineer who built the schema knew the Honeywell edge case in October. She flagged it in the risk register in November. The risk was accepted in the 16 December sprint review when the product manager ran the coverage numbers and concluded that 99.6% wellhead coverage was sufficient for a phase-one launch. The decision was sound. It was also the kind of decision that only makes sense until it does not.

The upstream oil and gas industry is already seeing value from AI deployment, with examples including SLB and Vår Energi building collaborative well-planning workflows that reduce cycle times from discovery to first oil from months to days on the Norwegian Continental Shelf. McKinsey estimated in August 2026 that AI could unlock about $65 billion in annual recurring value in the near term using today's technology, $125 billion as proven technologies are adopted more broadly, and $230 billion at full potential. Those estimates are net of more than $30 billion in annual AI implementation costs. The figures assume adoption at scale. They do not footnote the wells left behind because the data format was wrong.

The velocity matters in a way the coverage percentage does not capture. Expand produced 7,183 million cubic feet equivalent of natural gas per day in 2025, with 42% from Haynesville, 36% from Northeast Appalachia, and 22% from Southwest Appalachia. A 1% improvement in uptime across that production base, sustained over a year, is worth more than the cost of upgrading every legacy SCADA unit Expand operates. But uptime improvements accrue as the model learns, and the model learns from the wells it can see. The 23 wells in Greene County will contribute to that learning in phase two, which the internal roadmap schedules for Q2 2026. Between now and then, they produce into a market where the algorithm cannot help.

The Honeywell units are not the only edge case deferred. Leucipa's artificial-lift optimization module, which adjusts electric submersible pump speed and gas-lift injection rates in real time, was scoped in November to handle wells with dual-string completions common in the deeper Haynesville intervals. The module shipped in January with single-string logic only. Dual-string support is scheduled for the June release. In the Utica, 14 wells use a plunger-lift system instead of ESPs. Plunger logic was descoped in the December sprint because fewer than 0.3% of Expand's wells use the technology and the optimization gain, modeled at 2.1% incremental production, did not justify the four-week engineering cost. Those wells will stay on timer-based plunger cycles while the ESP wells get dynamic optimization.

None of this is malpractice. It is the ordinary arithmetic of shipping software into hardware-constrained environments where the installed base was laid down over 15 years by contractors who no longer exist, using equipment selected for reasons that made sense in 2014 and were never revisited. The platform works. It will deliver value. The fact that it does not work everywhere, yet, is a trade every vendor makes and few operators see until the井 井 coverage map is audited.

The announcement in Florence on 29 January was ceremonial. Josh Viets, Expand Energy's executive vice president and chief operating officer, said the collaboration marked a significant step forward in digital transformation, and that deploying AI-powered systems across the nationwide network of wells would enhance asset value through more efficient and reliable energy delivery. The equity release mentioned thousands of wells, three shale plays, real-time visibility, and a pilot of the conversational AI assistant. It did not mention the Honeywell units, the dual-string wells, or the plunger cycles. The coverage figure, 99.6%, did not make it into the press release because it raised a question nobody wanted to answer in public.

The software engineer in Florence committed the final manifest to the production repository at 9:14 AM on 27 January. The deployment began that afternoon. By 31 January, 5,782 wells were streaming telemetry into Leucipa's optimization engine. The algorithm started learning. The 23 wells in Greene County kept producing the way they had in December, which is to say manually, which is to say at a level of efficiency the AI could improve but cannot measure because the data never arrives.

The cost of shipping is not the code you write. It is the wells you leave behind while you write it.

Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.