Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

95% of AI Pilots Deliver No Profit: Europe Gets Stuck in Approvals, Russia in Deployment

MIT analyzed more than 300 corporate AI deployments and found that 95% of pilots added nothing to net profit (the P&L line in the financial statement). In Europe, projects stall at the start: lawyers spend months clearing them under the EU AI Act. In Russia, AI spending grew 95% in a year, but prototypes do not make it into real operations. The takeaway: rebuild the process first, or a borrower dossier that AI assembles in a second will still wait three days for manual review.

AI-processed from Habr AI; edited by Hamidun News
95% of AI Pilots Deliver No Profit: Europe Gets Stuck in Approvals, Russia in Deployment
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1In-house AI built from scratch reaches a result in 33% of cases; a product bought from a vendor does so in 67%
Europe71% of companies lack engineers, 52.5% cite uncertainty around the EU AI Act
Russiaof 12 public AI agents, zero are ready autonomous solutions

Companies have poured tens of billions of dollars into AI, yet the money is barely visible in their financial statements. MIT researchers analyzed more than three hundred corporate deployments and found that 95% of pilots produce no measurable effect on net profit, that is, on the P&L line (the profit and loss statement). Other surveys, meanwhile, cheerfully report that 64% of product managers see product metrics grow thanks to AI.

There is no contradiction here. The numbers simply lump together three different things: an employee who writes emails faster with an LLM, a polished demo in an isolated sandbox, and money that actually reached the P&L. Here is an example: a neural network assembles a borrower dossier in a second, but the file then sits for three days in a queue for manual review in a legacy accounting system.

The work got faster; the profit did not change. The author of an analysis on Habr compared statistics for Europe and Russia: the two regions approach AI from opposite ends and run into the same wall. If you are launching an AI pilot, this analysis can tell you where it is most likely to get stuck.

First, Let's Align the Samples

According to Eurostat, 20% of European enterprises with ten or more employees use AI, from 5.2% in Romania to 42% in Denmark. Comparing that with surveys of large Russian businesses would be misleading, so we take companies with 250 or more employees. There, the adoption rate is 55%. One caveat: Eurostat counts any AI, including classic scoring and text recognition. In Russia, according to Yakov and Partners and Yandex, 71% of large companies use generative AI, and including classic models the share reaches 88%.

Europe Stalls at the Entrance, Russia at the Exit

European business is not held back by money. The main blockers according to Eurostat are a shortage of engineering and product skills (71% of companies), legal uncertainty around the EU AI Act (52.5%) and strict data privacy restrictions (48.8%). A pilot spends months going the rounds of lawyers and compliance, while the model has not yet seen a single live request.

In Russia it is the opposite. According to a FinExpertiza study based on Rosstat data, spending on AI by business and the public sector grew 95% in a year, almost doubling, to 176.1 billion rubles.

Almost 80% of the money went to software and development, 92% of budgets landed in Moscow (largely because the headquarters of federal corporations are registered there), and one sector, finance and insurance, accounts for 84% of spending. 78% of executives report a "visible economic effect." But a checked box on a questionnaire and growth in operating profit are different things.

A market audit by Yakov and Partners found zero ready autonomous solutions among twelve public Russian AI agents: only builders and narrow assistants. As a result, large banks and corporations have to build agents from scratch in-house.

Why the Result Is the Same

The causes are described in the MIT Project NANDA study (The GenAI Divide), based on a sample of large American companies. The authors used a horizon of six months after a pilot ended and saw two patterns:

  • In-house solutions reach production-level results in 33% of cases, while deploying a specialized product from an external vendor does so in 67%.
  • More than half of budgets go to sales and marketing, although the real savings come from the boring back office: document processing, compliance, risk scoring and automation of routine operations.

The author, who personally took AI initiatives at a bank to the pilot stage and tracked more than ten startups, adds an engineering detail. On a clean test dataset, any model shows impressive numbers. In production, it has to pass bank information security, fit within strict response-latency SLAs, and have the cost of a type II error calculated.

If an antifraud or risk model cuts off legitimate customers with false positives, all the savings burn away. One team stopped a voice antifraud pilot three weeks after it went live on real traffic, because the economics and detection quality did not add up. In the author's view, this kind of emergency brake saves the budget from a "zombie project" that eats resources for years for the sake of a nice slide.

Forecast and What to Do

The author makes a verifiable bet. By 31 December 2027, the share of large companies in Europe and Russia using AI in products and operations will grow to 75–85%. The share of projects with a proven positive impact on P&L will not exceed 15%, meaning it will grow by at most a factor of three from the current 5%. He puts the probability at 70% and points to the history of ERP and cloud platforms, where roughly one in five projects got a return. The forecast will turn out wrong if audits at the level of McKinsey, Gartner or Yakov and Partners record growth in operating profit at more than a quarter of companies.

It is important to remember the limitations: this is an author's column, not new research. The data comes from sources with different methodologies, surveys about "effect" are subjective, and the MIT figures apply to American business, so extending them to Europe and Russia remains the author's assumption.

The practical takeaway for product managers and founders: count not the number of pilots, but what survived contact with the customer, the security team and the P&L line. Before launch, check whether the process has been rebuilt around the model. Before writing your own, see whether a ready-made product exists. Look for the effect in the back office, not only in the storefront, and set a stop criterion in advance for the third or fourth week of live traffic.

Story

Key Questions Answered

What are the key details on First, Let's Align the Samples?

According to Eurostat, 20% of European enterprises with ten or more employees use AI, from 5.2% in Romania to 42% in Denmark. Comparing that with surveys of large Russian businesses would be misleading, so we take companies with 250 or more employees. There, the adoption rate is 55%. One caveat: Eurostat counts any AI, including classic scoring and text recognition. In Russia, according to Yakov and Partners and Yandex, 71% of large companies use generative AI, and including classic models the share reaches 88%.

What are the key details on Europe Stalls at the Entrance, Russia at the Exit?

European business is not held back by money. The main blockers according to Eurostat are a shortage of engineering and product skills (71% of companies), legal uncertainty around the EU AI Act (52.5%) and strict data privacy restrictions (48.8%). A pilot spends months going the rounds of lawyers and compliance, while the model has not yet seen a single live request. In Russia it is the opposite.

Why the Result Is the Same?

The causes are described in the MIT Project NANDA study (The GenAI Divide), based on a sample of large American companies. The authors used a horizon of six months after a pilot ended and saw two patterns: In-house solutions reach production-level results in 33% of cases, while deploying a specialized product from an external vendor does so in 67%.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…