Organizational Coherence Decides Whether AI Pays Off Or Not At Scale

Madhu Shashanka is the Cofounder, Chief Scientist and CTO at Concentric AI and author of the forthcoming book, Coherence.

gettyWhy is it that individual team dashboards show successful AI deployments and wins, but the impact barely shows up at the organizational level? McKinsey’s 2026 “state of AI” report shared that 80% of respondents said that AI improved their individual productivity. But the share reporting any enterprise-level EBIT impact stayed flat at 37%, unchanged year over year. And “AI high performers”—companies who can attribute 5% or more of their EBIT to AI and see significant value—accounted for only 6% of the organizations.

More companies are scaling AI, but business outcomes haven’t caught up. Individual performance differs from enterprise performance, and only the former is watched at most organizations.

In July 2026, BCG researchers reported that about nine in 10 CEOs see benefits from AI in targeted areas, but most can’t scale those gains. “The core problem,” BCG analysts said, “is not technology. It is execution.” The same report points out that although two-thirds run pilots, only about one-quarter embedded AI in a real transformation.​

Agentic AI has made it easier for employees to set up an individual autonomous system using plain English and without expert help. The end result is many systems, deployed independently, for specific local goals. W. Edwards Deming observed decades ago that improving each part of an organization independently doesn’t necessarily improve the whole and can often degrade it.

There’s already a name for part of this: agent sprawl. According to SAP LeanIX’s 2026 survey, 98% are deploying agents or planning to. But fewer than half can inventory all the agents they run. Gartner analysts expect 150,000-plus agents at a large enterprise by 2028, but only 13% of organizations say they’re adequately governed.

Sprawl is only part of it. There’s a deeper problem that I call organizational coherence. Incoherence lives in the interactions between systems and not the overall number of them. You can have a portfolio of well-governed agents, each one working fine by itself, but they can be wrong together as a collective.

Organizational coherence is the integrity of the link between local actions and enterprise intent. In practice, it means the capacity to see what your systems are doing, to judge and monitor whether they’re doing it well and to correct them when they aren’t.

What happens when capable systems multiply without organizational coherence? You incur a cost that keeps compounding: complexity debt. It can be thought of as similar to software technical debt but at organizational scale. Each team’s locally rational choices become the enterprise’s collectively irrational problem. What makes it particularly insidious is that it sits silent, unmeasured and surfaces only once it’s severe.

Consider three supply-chain agents deployed by different teams. A forecasting agent updates the forecast from live sales, a replenishment agent places orders to meet the forecast and a production agent schedules output to incoming orders. Each one is locally right. Although agents exchange quantities, reasons behind a sales spike such as a promotion or one-time bulk order doesn’t cross the interfaces between them. A temporary spike is read as a new baseline, leading to higher orders and ramped-up production. The inflated orders look like even greater demand for the next cycle. Agentic speed greatly exacerbates the problem before a human review catches it.

This is a classic example of the bullwhip effect, an established systems failure even in fully integrated chains, because each node optimizes locally while no agent holds the global view. Each agent treats the other’s output as ground truth and none knows whether the demand signal is real. This shows how agents can be collectively wrong while each one is optimized for a rational local goal.

There are concrete actions enterprises can take now and that involves four steps: inventorying, constraining, containing and a judgment layer on top. The order matters because you can’t constrain or contain what you can’t see.​

The foundational step is visibility. Inventory all the autonomous deployments and map the handoffs between them. Identify which system’s output feeds which. Triage them into three buckets: continue under monitoring, needs an owner this quarter or pause now.​

The next move is to set constraints. Decide what your systems aren’t allowed to do and what you won’t automate at all. For instance, decide in advance which actions require a human to approve, such as any agent action that moves money, changes a price or deletes records. Setting up constraints helps prevent incoherence from arising in the first place.

Next, put monitored boundaries between systems that must not contaminate one another. That helps ensure that bad outputs stay local instead of cascading through the enterprise.

The final layer is judgment. Weigh the breadth, depth and reversibility of each deployment before you decide to scale. A narrow but deep dependency can create more complexity debt, and risk, than a broad but shallow one. The calls you need to make here are specific to your enterprise, and no vendor can sell them to you.

Enterprises need to create a new structure for coherence to ensure they can judge and correct what’s been built as they scale AI. The cost of skipping this work might gain deployment speed initially but comes with complexity debt. Only coherent enterprises can confidently move fast, while the incoherent spend years undoing debt that’s compounded.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?