Concentric AI's CTO blames uncoordinated team agents for AI's stalled profit impact

The board-deck version of the McKinsey result is that staff like the tools and the programme is working. It leaves out the enterprise line. The share of respondents reporting any EBIT impact was unchanged from the year before ^[2]. The spread between the two answers is 43 points ^[1], and they measure different things: one asks about a person's own work, the other about company profit. Shashanka's point about incentives is plainer. Individual and enterprise performance differ, he wrote, and at most organizations only the individual kind is watched ^[7].

His account of the cause starts with how agents get built. Agentic tools let employees set up an autonomous system in plain English without expert help, so companies end up with many systems deployed independently for local goals ^[8]. He cites W. Edwards Deming's observation that improving each part of an organization independently does not necessarily improve the whole and can degrade it ^[9]. His name for the accumulated cost is complexity debt. "Each team's locally rational choices become the enterprise's collectively irrational problem," he wrote ^[13].

Shashanka separates this from agent sprawl. A set of well-governed agents, each working correctly alone, can still be wrong as a group, because the failure sits in the interactions between systems ^[12]. In his supply-chain case the agents pass quantities to one another, but the reason behind a spike, such as a promotion or a one-time bulk order, does not cross the interface ^[19].

BCG's July report reaches a similar diagnosis in different words. "The core problem," BCG analysts said, "is not technology. It is execution." ^[6]

The argument comes from an interested party. Shashanka is cofounder, chief scientist and CTO of Concentric AI, and his book Coherence is forthcoming ^[4]. His supply-chain case is a constructed example, though the failure it describes is old: the bullwhip effect is an established systems failure even in fully integrated chains ^[15]. In my view the diagnosis is better supported than the remedy. The surveys he cites measure reported gains, adoption and governance ^{[1][5][10][11]}; none compares EBIT impact at companies that coordinate their agents with those that do not ^[2].

The timing splits into this quarter and this decade. Gartner analysts expect large enterprises to run more than 150,000 agents by 2028, and only 13% of organizations say theirs are adequately governed ^[11]. The first of Shashanka's four steps is nearer. He orders them as inventory, constraint, containment and a judgment layer on top. "The order matters because you can't constrain or contain what you can't see," he wrote ^[17]. The inventory lists every autonomous deployment, maps which system's output feeds which, and sorts each into buckets that include continue under monitoring and needs an owner this quarter ^[18].