Why AI Alone Doesn’t Move EBIT

Eighty percent of employees say AI has made them more productive. Only 37 percent of organizations can point to any earnings impact from it, and just 6 percent can point to a meaningful one. That gap has held steady for two years running. It isn’t a sign that AI doesn’t work. It’s a sign that most companies have adopted a tool without changing how they make decisions.

We see the same pattern inside industrial pricing and commercial organizations. A team stands up an AI model to flag margin leakage or pricing opportunities. The dashboard looks sharp. The model surfaces hundreds of recommendations. Six months later, discounting behavior hasn’t moved, the sales team still prices the way it always has, and the tool quietly becomes another login nobody opens. The model wasn’t wrong. It just wasn’t enough.

Two things separate the companies that convert AI into EBIT from the ones that don’t, and both are less about the model than what surrounds it.

The first is whether domain expertise is built into the AI itself, not applied afterward. A pricing model that doesn’t understand which customer relationships carry strategic weight, which discounts reflect a competitive threat rather than poor discipline, or which accounts are mid-negotiation on a contract will generate technically correct output that no one trusts enough to act on. That distrust is rational. A sales VP who has been burned once by a recommendation that ignored real business context won’t open the tool a second time. Building that context into the model, rather than asking a person to manually filter every output through it, is what turns a list of anomalies into a list of decisions worth making.

The second is whether there’s a team responsible for turning a validated recommendation into a changed behavior. Flagging that a customer is underpriced doesn’t reprice the account. Someone has to walk the rep through the trade case, reset the deal desk guardrail, or coach the account team through the renewal conversation. That is a workflow and change management problem, and it is the layer most AI deployments skip entirely, because it is slower and harder to automate than the analysis itself. It is also where the value actually gets captured.

This is precisely the premium McKinsey points to in its research on human skills. AI hasn’t reduced the need for judgment. It has concentrated it into fewer, higher-stakes moments: deciding whether a recommendation is right, and deciding what to do about it. Organizations that treat those moments as an afterthought will keep generating productivity anecdotes without earnings impact. Organizations that build judgment into both the model and the team around it are the ones showing up in that 6 percent.

None of this is an argument against AI. It’s an argument for building it the way the data says the winners already are: as a continuous capability that carries business context in its logic, paired with people accountable for making sure what it finds actually changes how the business runs. Technology alone was never going to close a judgment gap. Only expertise, embedded in both places, will.