Solow Paradox: The productivity puzzle of AI can be cracked

This phenomenon of extensive use of advanced technology without an expected increase in productivity is called the ‘Solow Paradox.’ 

This particular paradox was resolved in the 1990s, led by industries like technology, banking and retail showing productivity gains thanks to their adoption of information technology (IT), supported by process redesigns and re-imagined product and service offerings.

Today, the adoption of artificial intelligence (AI) may be suffering from its own version of the Solow Paradox. BCG’s AI Radar (January 2025), which surveyed 1,800 plus C-suite global executives, found that only 25% see value gains from AI. 

What may be amiss? Let’s consider the case of banking, an industry with large investments in data, AI and technology, to  grasp why some banks vouch for AI benefits while others do not.

Also Read: Productivity puzzle: Solow’s paradox has come to haunt AI adoption

Whether an organization captures value from AI tends to depend on four interconnected factors. When these factors form a mutually reinforcing virtuous cycle, it offers the organization market dominance, higher margins and strategic moats. 

Focus on decision quality: Real-time AI driven predictions, run on terabytes of data, do not improve decision making on their own. This is a function of decision-making process discipline; AI only plays a supportive role. A high-quality decision is one whose post-facto assessment shows that a better decision could not have been taken, given the available information and trade-offs. Firms often assume that good business outcomes are a sign of good decision quality. However, bad decisions can be profitable and good ones may lead to losses. 

Companies with well-defined processes for decisions tend to consistently make better calls and exhibit better overall performance than others. Such companies usually capture more value from AI initiatives.

Let’s see how. Say, a similar predictive credit risk score is developed by three banks. Bank A decides to share this score with the underwriter, who has full authority to overwrite the score for loan approvals. AI is unlikely to improve Bank A’s performance  meaningfully. Bank B goes to the other extreme and makes risk scores the key determinant of loan decisions across its entire asset portfolio without any human intervention. Since no model has high predictive power across all borrower segments, Bank B may exhibit a higher delinquency rate. 

Bank C, known for its decision quality focus, uses the score for differential underwriting. It identifies segments where the score will beat human assessments and lets it drive such loan decisions, while asking human underwriters to take credit calls where the data is either limited or misleading, using the model score only for support. Bank C is likely to benefit the most from AI.

Also Read: Sovereign AI is best developed as an enabler of soft power rather than hard

Go for big AI bets: Organizations that track their decision quality are aware of the shortcomings and tend to use AI to solve their biggest decision-making problems, be it at a strategic or transactional level. Such organizations place bigger bets on AI, give it their full corporate commitment and capture most value from AI. Some other companies, while being equally excited about AI, are often unaware of the areas where their decision making is weak. They tend to run AI pilot initiatives in less significant business areas. Even if such trials are successful, they do not justify the investment and are rarely scalable.

Don’t automate sub-optimality: Generative AI and Agentic AI create a strong case for automating complex but repetitive operations. Processes need to be checked for optimality and the absence of decision bugs or ambiguity before being automated using AI. However, if AI is used to automate process flows that are suboptimal or have decision bugs, then automation will cause bad decisions to be taken faster. 

Also Read: Reasoning AI: Cracking the IIT test is a giant leap for machine-kind

Don’t let AI magnify core weaknesses: AI initiatives work best when the organization has been investing in data governance and has high-quality data on which models are developed and run. However, at least some businesses appear to be chasing GenAI on account of the belief that it does not require a lot of high-quality data and employees skilled in all aspects of AI. 

While pilot trials may show success, scaling up requires high-quality data plus significantly upskilled employees. After all, AI models and utilities customized for an organization will not automatically update themselves.

Also Read: Free-AI: How well India adopts AI will shape its future

The path best taken: Merely buying cloud computing capabilities and initiating an array of traditional AI and GenAI pilots is not enough. Using AI to improve operating metrics will require organizations to work on the four factors outlined above. However, success will only buy time and not ensure long-term sustainability unless the core business model is adapted to the AI era.

The electric bulb did not come from a continuous improvement of candles, to quote Oren Harari. If 19th century candle makers had access to electricity, most would have used it to melt wax faster to make more candles. It is unlikely that they would have invented the light bulb.

However, GenAI offers organizations the power to tap civilizational knowledge and reduce their institutional knowledge gaps. Can today’s organizations be candle makers that also invent the light bulb of the future? It has never happened before, but then, there was no AI till recently.

The author is a risk management and AI consultant, and a member of the visiting faculty, IIM Ahmedabad and IIM Calcutta.