← Back to all tipsNovember 17, 2026

Everyone has adopted AI. Most haven't made money from it.

Adoption numbers for AI are, by now, almost universal. Stanford HAI puts organizational AI use at 88%. McKinsey's latest read has 44% of companies saying AI is "scaling" across the enterprise, up from 38% a year earlier. Global corporate AI investment hit $581.7B in 2025 — up 130% year over year. By any adoption metric, this is settled.

The number that doesn't move with it

Only 37% of organizations attribute any EBIT impact to their AI use, and more than 80% still report no measurable financial effect at the enterprise level. Adoption and return have decoupled. Companies keep buying and deploying AI at a faster rate every year, and the share of them who can point to money it made or saved isn't keeping pace.

Why swapping models doesn't close it

The default response to disappointing AI results is to swap something at the model layer — a newer frontier model, a different vendor, a fine-tuned variant. That's treating this as a model-choice problem, and the data doesn't support that framing. 63% of organizations say they lack, or aren't sure they have, the data-management practices AI actually needs. No model, however capable, produces a reliable financial return on top of historical data that's incomplete, inconsistently labeled, or scattered across systems that don't agree with each other. The model was rarely the bottleneck.

What actually determines the payoff

The groundwork that decides whether an AI project pays off is unglamorous and happens before any model gets trained: historical data that's actually clean and joined across sources, a success metric defined in business terms before the project starts (not a model accuracy number chosen after), and a discovery phase that establishes whether the data can support the question being asked at all. Skipping straight to model selection is skipping the step most likely to determine the outcome.

Where this shows up in how we work

It's why Predictions starts with "prepare the data" as its own explicit phase, not a footnote before training — collecting, cleaning and joining historical data from a customer's actual sources, before any model gets near it. It's the same reason Development engagements open with "discover and plan": requirements, constraints and a roadmap the customer signs off on, before a single pipeline gets built. Neither is process for its own sake. It's the part of the work that determines whether the rest of it is worth doing.

The AI adoption curve isn't the story anymore. Whether that adoption turns into a number on an income statement is — and that number is decided earlier than most teams start looking.