Technical Context: Where the Study Punctures the Hype
I reviewed the NBER paper w34836 and immediately noticed the main takeaway: the authors don't deny that AI accelerates specific tasks. Instead, they highlight a much more uncomfortable reality for management—at the company level, this effect has barely materialized.
The sample includes nearly 6,000 executives from the US, UK, Germany, and Australia. Around 70% of companies already use AI, but over 80% haven't seen measurable impacts on employment or productivity, and 89% noticed no change in sales per employee.
Another figure caught my attention: top executives use AI themselves for about 1.5 hours a week on average, and a quarter don't use it at all. To me, this is a strong signal that in many companies, decisions about artificial intelligence implementation are made by people who don't live inside the new toolkit and don't understand exactly where the economic impact occurs.
However, the study doesn't dismiss local victories. In customer support, content drafting, and data processing, growth can be noticeable. Yet, there is a massive gap between «an employee finished a task faster» and «the company became more productive,» filled with integrations, processes, KPIs, and managerial discipline.
Impact on Business and Automation: Who Wins and Who Loses Time
I've been telling clients an uncomfortable truth for a long time: buying LLM access does not equal results. If a company stops at a chatbot, a few prompts, and a presentation for the board of directors, it is almost guaranteed to end up in that 80% with no measurable effect.
The winners aren't those who «experiment with AI,» but those who build AI architecture around a specific operational bottleneck. When I design AI solutions for businesses, I don't start with the model. I start with the unit economics of the process: where hours are lost, where the cost of an error increases, where there's a repeatable loop, and where quality control can be embedded.
Companies that measure productivity only through top-level metrics and expect an instant leap are the ones losing out. If AI reduced response time, improved ticket classification quality, or decreased document defect rates, this still needs to be properly translated to the P&L via routing, SLAs, CRM, ERP, and team regulations.
This is exactly why AI automation without proper integration rarely yields results. In our experience at Nahornyi AI Lab, a measurable effect appears where we connect the model to data, roles, risk limits, and the decision-making system, rather than leaving it as an isolated «smart button.»
Strategic Perspective: The Productivity Paradox Doesn't Debunk AI, It Exposes Weak Implementation
I don't read this study as a death sentence for AI. I read it as a diagnosis for the market. Businesses have widely reached the stage where demos exist, but AI solution architectures do not.
This looks very much like the early stages of ERP, CRM, and RPA. First, companies bought the tool, then they were disappointed, and real returns went to those who redesigned the process, shifted responsibilities, and built end-to-end integration of artificial intelligence into the operational loop.
My forecast is simple: over the next 12–24 months, the gap between «we use AI» and «we profit from AI» will become even more pronounced. The market will split into two groups—some will continue counting the number of licenses, while others will start calculating the cost of resolving a case, cycle speed, and margin per employee.
In Nahornyi AI Lab projects, I see one consistent pattern. As soon as we remove the abstract goal of «doing AI automation» and replace it with a combination of process, metric, integration, and an owner, the effect becomes financial rather than philosophical. This is exactly what most surveyed companies are currently missing.
This analysis was prepared by Vadym Nahornyi — Lead Expert at Nahornyi AI Lab on AI architecture, AI automation, and practical AI implementation in business processes.
I invite you to discuss your case with me and the Nahornyi AI Lab team. If you need a working architecture, AI integration, and clear project economics rather than just another pilot for a report, contact me — I will help assemble a solution tailored to your actual operational environment.