Everyone is using AI. Almost nobody is making money from it.
McKinsey's State of AI 2025 is the largest annual survey of AI adoption. 1,993 respondents across industries and geographies. The headline finding: only 109 firms — 5.64% — report that AI contributes meaningfully to their EBIT (earnings before interest and taxes). 61% report zero measurable financial impact.
I build AI products for clients. I have watched companies spend six figures on AI initiatives that produced nothing measurable. I have also watched companies spend $20,000 on a focused LLM integration that paid for itself in the first month. The difference is not the technology. It is the approach.
The Data
McKinsey State of AI 2025 (n=1,993):
- 5.64% report meaningful EBIT contribution from AI
- 61% report zero measurable financial impact
- 72% of organizations have adopted AI in at least one business function
- The gap between adoption (72%) and value (5.64%) is the story
Boston Consulting Group (2024):
- Only 26% of companies have moved AI pilots to production
- 74% are stuck in "pilot purgatory" — running experiments that never scale
Gartner (2025 forecast):
- Through 2025, at least 30% of generative AI projects will be abandoned after proof of concept
- The primary reason: inability to demonstrate business value
Three research firms. Same conclusion. AI adoption is widespread. AI value is not.
Why 94% Fail
They start with technology, not problems
The pattern I see repeatedly: a company buys an AI tool (or hires an AI team), runs a pilot, builds a demo, shows it to the board, and declares the pilot a success. Six months later, nobody is using it. The demo was impressive. The business value was zero.
The companies in the 5.64% do the opposite. They start with a specific business problem: "Our patient intake calls convert at 12%. How do we get to 20%?" Then they work backward: the AI implementation is the last decision, not the first.
RiseMD works this way. The problem was measurable: dental practices spending money on marketing with no visibility into which campaigns drove actual patient revenue. The AI layer (call grading, search positioning, attribution analytics) was built to solve that specific measurement problem. Result: $3.2M in production from $160K spend. 20X ROI. Not because the AI was sophisticated. Because the problem was specific and the success metric was defined before a line of code was written.
They build custom when they should integrate
Most companies do not need a custom-trained model. They need a commercial LLM (OpenAI, Anthropic, Google) integrated into their existing product with proper engineering around it.
Custom model training makes sense when you have proprietary data that no commercial model can access and when the performance improvement justifies the cost. For most business applications — document Q&A, content generation, data extraction, customer support — a well-engineered integration with a commercial model outperforms a custom model at a fraction of the cost.
The 94% who see no ROI often overspend on custom AI when an integration would have been faster, cheaper, and more effective.
They skip the engineering
A Jupyter notebook is not a product. A demo that works on 10 examples is not production software. The gap between "it works in the demo" and "it works reliably for 10,000 users per day" is where most AI projects die.
Production AI requires: error handling (what happens when the model returns garbage?), cost controls (what happens when a runaway loop burns through your API budget?), latency optimization (users will not wait 15 seconds for a response), monitoring (how do you know when quality degrades?), and fallback strategies (what happens when the API is down?).
These are software engineering problems, not AI problems. The companies that solve them are companies with strong engineering teams. The companies that fail are companies that treat AI as a separate initiative disconnected from their engineering organization.
They cannot measure success
"We implemented AI" is not a result. "AI reduced our customer support ticket volume by 34% and saved $180,000 per quarter" is a result.
The 5.64% who see EBIT impact measure AI the same way they measure any other business initiative: revenue generated, cost reduced, time saved, errors prevented. The 94% who see no impact often cannot tell you what they expected AI to achieve in the first place.
What the 5.64% Do Differently
Based on what I see across our client work and the McKinsey data:
They start small. One use case. One measurable outcome. Not "transform the business with AI." More like "automate the extraction of line items from invoices so the accounting team saves 12 hours per week."
They integrate, not build. Commercial LLMs via API, not custom models trained from scratch. RAG pipelines with their own data, not general-purpose chatbots.
They measure from day one. The success metric is defined before the project starts. Baseline measured. Target set. Progress tracked weekly.
They treat AI as engineering. The AI project runs in the same sprints, with the same code review process, the same CI/CD pipeline, and the same deployment standards as every other feature. It is not a separate "innovation lab" project.
They ship to production. 74% of companies are stuck in pilot mode (BCG). The 5.64% ship to production, get real user feedback, and iterate.
The EltexSoft Approach
We build AI as features inside larger products, not as standalone experiments. Every AI feature we ship has a defined business metric, a production deployment, monitoring, and cost controls.
Snapwire used ML-powered image tagging and quality scoring across millions of photos for Fortune 500 brands. The metric was matching quality: did the right photographer get matched to the right brand?
RiseMD uses AI for search positioning and call grading. The metric is patient revenue attribution: does the practice know which marketing dollar produced which patient?
Woodies Clothing uses AI for product recommendations and demand forecasting. The metric is conversion rate and inventory efficiency.
None of these started as "let's implement AI." All of them started as "here's a business problem" and AI was the tool that solved it.
If you are in the 94%, the fix is not more AI. It is better engineering around the AI you already have. Start with a problem. Define a metric. Build to production. Measure the result.
Last updated January 18, 2026