AI in the enterprise: what works and what breaks, sector by sector (with data and sources)

Companies are adopting AI at scale, yet almost none turn it into profit. In 2025, 88% of organizations were using AI in at least one function, up from 78% the year before. Even so, only about 6% were capturing meaningful financial value, according to McKinsey. An MIT study put it more bluntly: roughly 95% of generative-AI pilots produced no measurable impact on the bottom line.

The gap between those who harvest value and those who only spend is almost never the AI model. It is the implementation method. This piece gathers what worked and what broke across five sectors, every number sourced, and shows the pattern behind both sides.

The paradox: everyone adopts, almost no one harvests

Adoption is nearly universal, but returns stay rare. In McKinsey's survey (1,993 respondents, 105 countries), 39% of companies report some EBIT impact from AI, and most of them say it is less than 5% of EBIT. The group actually extracting significant value is around 6%.

MIT called it the "GenAI Divide". The root cause is not technological, it is organizational: the inability to integrate AI into workflows, structure and culture. In the same study, tools bought from specialized vendors succeeded about 67% of the time, versus one-third of that for internal builds without a method.

Sector by sector: proven wins and failures

The table below sums up both sides of each sector. The detail and the lesson for each follow.

Sector	Where it worked (sourced)	Where it broke (sourced)
Finance	GenAI could add $200–340B/yr to banking, ~9–15% of operating profit (McKinsey). Klarna: two-thirds of support and ~$40M impact.	Klarna itself walked it back in 2025 and started rehiring humans for premium support.
Healthcare	1,451 AI-enabled devices cleared by the FDA through 2025. AI clinical scribes saved ~15,791 hours of documentation at Kaiser Permanente.	IBM Watson/MD Anderson: $62M and "unsafe and incorrect" recommendations. The Epic Sepsis Model missed two-thirds of cases (JAMA).
Retail and support	Klarna resolved tickets in ~2 min vs 11 min before and cut repeat contacts by 25%.	Air Canada was held liable in a 2024 tribunal for wrong information its own chatbot gave.
Industry	WEF Lighthouse factories: +40% labor productivity, −41% defects, −44% cycle time.	Tesla: Musk admitted "excessive automation was a mistake" on the Model 3 line.
Legal and software	Contract review at 94% accuracy vs 85% for lawyers. GitHub Copilot: +55% speed on a controlled task.	Experienced devs were 19% slower with AI (METR). Legal tools hallucinate 17–34% (Stanford). Sanctions for fake cases in Mata v. Avianca.

Finance: real scale, but keep a human in the loop

McKinsey estimates generative AI could add $200 to $340 billion a year to banking, mostly through productivity. The most-cited case is Klarna: its assistant had 2.3 million conversations in the first month, handled two-thirds of support, and represented about $40 million in profit impact in 2024.

The counterpoint came from the same Klarna. In 2025 the CEO admitted the company had cut too deep on people and reopened premium-support roles. The lesson is not "AI has no place in support". It is that quality needs a human in the loop for sensitive cases. Automate the repetitive volume, keep people where the decision matters.

Healthcare: diagnosis advances, clinical decisions need caution

Through 2025 the FDA had cleared 1,451 AI-enabled medical devices, 295 in 2025 alone, many in imaging and cardiology. On the administrative side, AI clinical scribes saved around 15,791 hours of documentation across 2.5 million encounters at Kaiser Permanente, with a meaningful drop in burnout.

The other side is serious. IBM Watson for Oncology at MD Anderson consumed about $62 million and was shut down after giving recommendations rated "unsafe and incorrect". The widely deployed Epic Sepsis Model missed two-thirds of sepsis cases in external validation published in JAMA. In healthcare, AI as support is one thing; AI deciding without oversight is another.

Retail and support: automate the volume, not the accountability

In support, the speed gains are real: Klarna cut resolution time from 11 to about 2 minutes and reduced repeat contacts by a quarter. But the legal risk is concrete. In 2024, a Canadian tribunal held Air Canada liable for wrong information its chatbot gave a customer. The company argued the bot was a "separate entity"; the tribunal rejected that. The company answers for everything the bot says.

Industry: AI is only as good as the data and legacy beneath it

The World Economic Forum's Global Lighthouse factories show the best numbers on the upside: on average +40% labor productivity, −41% defects and −44% cycle time in the latest wave. The classic counterpoint is Tesla, where Musk himself admitted that over-automating the Model 3 was a mistake because robots failed at fine-dexterity tasks.

The industry lesson is the most transferable of all: the bottleneck is rarely the algorithm. It is data quality and integration with the legacy systems underneath. That is exactly why, before any AI project, a diagnosis of what already exists pays off.

Legal and software: it accelerates, but never skips verification

On the upside, a controlled NDA review showed AI at 94% accuracy versus 85% for experienced lawyers, and in seconds. In development, a GitHub study found 55% more speed with Copilot on a specific task.

But context changes everything. A 2025 METR experiment measured experienced devs working in codebases they knew well and found 19% slowdown with AI, even though the devs believed they had sped up. Dedicated legal tools hallucinate 17% to 34% of queries, per Stanford. And Mata v. Avianca ended with lawyers sanctioned for citing fake cases invented by ChatGPT. The practical rule: every AI output is a draft until it is verified.

The pattern: what separates the 6% who capture value

Notice that none of the failures above were caused by "bad AI". Watson, Epic Sepsis, Air Canada, Tesla and Mata v. Avianca are failures of implementation and governance, not of the model. What the 6% who capture value do differently fits in five points:

They start from a process with clear ROI, not from the trendy tool.
They fix data and legacy systems before plugging AI on top.
They keep a human in the loop for decisions that matter.
They require verification of every sensitive output: citation, number, code.
They measure impact on results, not just usage volume.

Where Reche comes in

That is the difference Reche operates. RecheOS delivers AI with method: quality gates, a human in the loop and clean data, instead of "buy a tool and hope". If your company wants to adopt AI but does not know where to start, the ROI diagnosis defines scope and priority before you spend. If the bottleneck is an old system, the legacy code diagnosis attacks the very cause that sinks most projects.