An eight-to-14-person team is the fix one utility AI vendor prescribes for stalled pilots

Add the roster up and it comes to eight people at the floor and 14 at the ceiling ^[12][13]. That is a standing function rather than a project, and two of the eight minimum seats, the director and the data governance specialist, exist to coordinate and audit rather than to build ^[12]. A utility funding the floor is buying supervision before it buys capacity, which is defensible if the diagnosis is right and expensive if it is not.

The two failure figures in Gupta's case count different things. The Gartner research he cites concerns enterprise projects that data readiness issues stall or kill before commercial deployment, more than 60% of them ^[5]. The MIT NANDA figure, up to 95%, concerns generative AI pilots that fail to deliver measurable impact ^[6]. Those are 35 percentage points apart and describe different failures ^[19]: a pilot that never ships is not the same problem as one that ships and underdelivers. Gupta presents the MIT number as showing something similar to the Gartner one ^[6]. A center of excellence built to unblock data pipelines answers the first failure; one built to choose better use cases answers the second.

The strongest support in this record for reading the problem as organizational is the BCG split Gupta quotes: 70% of an AI transformation's value from workforce change, talent and process redesign, 20% from technology infrastructure, 10% from the algorithms ^[16]. That assigns seven times as much value to people and process as to the models ^[17]. It also arrives secondhand in a post by a vendor CEO ^[1], with no method attached, so it carries the weight of a plausible prior rather than a measurement.

On the evidence supplied, a center of excellence reads as a reorganization sold as a remedy: nothing here compares deployment rates at utilities that have a CoE against those that do not ^[20]. What the piece does supply that a buyer can use this quarter is a procurement test. It asks what platform IP and operational models the internal team owns outright on day one, whether the platform integrates with the existing cloud data environment without proprietary wrappers or custom connectors, and how raw meter data becomes measured workflows for grid planners, field technicians and customer service staff ^[15].

Those questions are not neutral. Gupta's company sells AI energy analytics to utilities ^[1], and his test is easier to pass for a platform supplier than for an integrator whose business model favors continuous engagement, which is the risk he attributes to CoE builds handed to firms such as Accenture, Deloitte, EY and Wipro ^[9][10]. The provenance does not make the questions wrong. A utility that cannot name what it owns on day one, or point to a 12-month schedule for taking over model governance, has learned something about its own position regardless of who drafted the question ^[15].

The tradeoff Gupta names is size, and he leaves it unresolved. Decentralized decision-making is the diagnosis he borrows from MIT ^[6], centralization is the cure ^[8], and he warns that too many members slow centralized decisions and create the bureaucratic drag the CoE was meant to remove ^[14]. Somewhere between eight and 14 seats, range stops buying speed ^[13][14], and this record does not say where. Utilities standing a CoE up this quarter will locate that point by watching which decisions stop moving, and the contract terms they sign now determine whether they can still change vendors when they find it ^[15].