AI is 280 times cheaper in about 2 years, and adoption has tripled in about 1 year. Yet, approximately 95% of organizations see no impact on their bottom line—Causes and Countermeasures for No Increase in Sales or Profits After AI Implementation

FULLFACT Inc. (Headquarters: Tokyo, Representative Director: Ayato Adachi), which provides AI implementation support, has released "Causes and Countermeasures for No Increase in Sales or Profits After AI Implementation 2026," a re-analysis of domestic and international public statistics, official documents, and international organization reports.

This report cross-references various indicators of AI's capabilities, cost, and adoption with financial performance indicators such as profit and loss impact and value creation. It organizes, based on primary sources from Japan and abroad, why this discrepancy arises and what conditions differentiate successful outcomes.

Key Points

Capabilities have soared in one year: In challenging benchmarks introduced in 2023, AI scores rose by MMMU+18.8, GPQA+48.9, and SWE-bench+67.3 points in one year. For SWE-bench, which measures bug fixing in actual code, the percentage of solved problems jumped from 4.4% in 2023 to 71.7% in 2024.

The same performance became 280 times cheaper in about 2 years: The cost of queries to achieve GPT-3.5 level performance fell from $20.00 per million tokens in November 2022 to $0.07 in October 2024. This is a drop of over 280 times in about two years (approximately 23 months). Hardware costs are falling at an annual rate of 30%, and energy efficiency is improving by 40% annually.

Adoption is unstoppable: Individual generative AI usage experience in Japan grew by approximately three times, from 9.1% in FY2023 to 26.7% in FY2024. A McKinsey survey shows that 88% of organizations routinely use AI in at least one business function. Capabilities, cost, and adoption are all moving in a positive direction.

Yet, the bottom line remains stagnant: An MIT NANDA survey found that despite companies investing $30-40 billion in generative AI, approximately 95% of organizations have gained no measurable impact on their profit and loss. Even with improvements in capabilities, cost, and adoption, financial results are not following. This is the core of the paradox.

Only a few companies effectively boosted profits: A McKinsey survey showed that only 39% of organizations reported any impact on company-wide operating profit, and most of those stated that AI's contribution to profit was less than 5%. Only about 6% of high-performing companies achieved a significant impact on company-wide profits with AI. A BCG survey also found that only 26% of companies are actually creating value from AI, with the remaining 74% yet to demonstrate concrete value.

The dividing line is not intelligence but implementation: The lack of results despite falling prices and rising performance is not an issue of capability. In the US, 57% of AI-using companies limit AI to three or fewer business functions, and only 4% of user companies deploy it comprehensively. The design of implementation—which tasks to use it for, what to measure, who confirms, and what data to input—is what separates success from failure.

Background

The numbers surrounding AI are moving simultaneously in two contradictory directions. Capabilities have significantly redrawn challenging benchmarks in one year, the cost of achieving the same performance has fallen by over 280 times in about two years (approximately 23 months), and the utilization rate has tripled in one year. All indicators showing improvement are looking up.

However, financial indicators are not moving. An MIT NANDA survey found that approximately 95% of organizations investing in generative AI have not gained a measurable profit and loss impact, and a McKinsey survey found that only about 6% of high-performing companies achieved a significant impact on company-wide profits. Despite capabilities, cost, and adoption all being in place, results are lagging. This is the AI productivity paradox.

FULLFACT believes that the cause of this discrepancy lies not in a lack of capability but in implementation. In the US, the majority of companies using AI limit it to three or fewer functions, and only 4% of user companies have deployed it company-wide. Which tasks to use it for, what to measure, who confirms, and what data to input. The presence or absence of this design is what differentiates companies that achieve results from those that don't, even when facing the same favorable conditions.

What to check first for AI implementation

Choose tasks that yield results: Instead of increasing users, focus deeply on one task that directly impacts profit and loss. Start with tasks where business frequency, data handled, confirmers, and performance indicators are visible.

Decide on metrics in advance: Before implementation, define what success means in terms of points moved. Measure by business processing time, changes in sales/costs, rather than user count or number of contracted tools.

Draw a line for acceptable data: Decide for each task what information can and cannot be provided to AI. As usage expands, the absence of this demarcation will undermine both results and trust.

Assign confirmation and accountability: For each task, decide who confirms the output, who corrects errors, and who updates operations. A vacuum in confirmation can confuse the quality of PoC output with actual production results.

Expand scope quarterly: Record the procedures and failure logs for tasks that yielded results, and continue to expand.