Why Your AI Tools Aren't Working

You bought the tools. Your team has access. Three months later, the work looks the same.

The number that exposes the real problem

McKinsey's 2025 and 2026 State of AI surveys show only 37–39% of organizations report any enterprise-level EBIT impact from AI. The share of firms classified as "high performers" has held flat at roughly 6% across both years, despite rising adoption and rising investment. If the tools were the bottleneck, that share would grow as tools improve. It hasn't moved.

The explanation the research offers is structural. Brynjolfsson and Hitt's firm-level work establishes that technology investment pays off when paired with complementary changes in decision rights, skills, and job design. Layered onto unchanged processes, the same investment produces near-zero returns. This pattern showed up first with enterprise software in the 1990s. It is showing up again now.

The counterargument deserves a fair hearing. EBIT is a lagging, aggregated measure. It absorbs AI-generated gains slowly because accounting systems were not built to isolate the contribution of a specific process change to operating income. Governance constraints, data quality problems, and compliance guardrails suppress reported impact independently of whether integration was done well or badly. Under this reading, the 37–39% figure understates real value, and blaming workflow design is too convenient.

This defense holds at the individual-firm level. It breaks down when you look at two consecutive years of flat high-performer share. A measurement lag that persists unchanged across two years of active investment and learning is not a measurement lag.

What the redesign-first cases actually show

HR tech vendors, banks, healthcare networks, and SaaS firms that redesigned onboarding end-to-end before embedding AI saw large reductions in cycle time, error rates, and early attrition. Firms that added AI tools to their existing onboarding flows saw marginal change. These are operational metrics, not EBIT. Accounting lags do not explain why the two groups diverged at the process level.

The HBS AI Institute's founder survey adds a behavioral dimension. How founders delegate to AI and what they believe about it shapes whether AI changes actual work patterns. This is not a technology finding. It is a job design finding. The binding constraint is not what the tool does. It is what the workflow permits the tool to do.

I find the "just add AI" consulting pitch genuinely irritating, and I'll say why. Vendors selling AI-powered onboarding tools, AI-powered CRM overlays, AI-powered meeting summaries — they all share the same sales motion: drop this into your existing stack and watch the numbers move. The pitch works because it asks nothing of the buyer. No process change, no role redesign, no decision about what the AI is actually deciding. The McKinsey data is a two-year argument against this sales motion. The vendors know it and keep pitching anyway.

Where to start if you've already bought the tools

Pick one process where outcomes are measurable and the cost of failure is visible. Onboarding is the clearest example because early attrition is trackable and the process has a defined start and end. Map the current flow from intake to completion. Identify which steps require human judgment and which steps are purely information transfer, document extraction, or routing. Redesign the flow so AI handles the latter category as an embedded component, not an optional add-on sitting beside the old steps.

Then run the old process and the new one in parallel long enough to collect operational data. Cycle time, error rate, completion rate at each stage. Not EBIT. Operational metrics move faster and tell you whether the integration is working before the accounting systems catch up.

Bresnahan, Brynjolfsson, and Hitt's work on workplace organization shows this pattern across decades of IT investment: the firms that captured returns from technology were the ones that changed job design alongside the technology, not before it and not after it, but as part of the same decision. The sequence matters because the technology shapes what redesign is possible, and the redesign shapes what the technology is allowed to do.

The 6% of firms reporting large enterprise-level impact from AI are not using better tools. They redesigned something first.