Are We Approaching Diminishing Returns from Generative AI?

Why increasingly capable AI may not necessarily deliver greater business value.

Generative AI has been one of the most remarkable technological developments in recent history. In just a few years, it has transformed how we research, write, learn, analyse data and develop ideas. More recently, it has demonstrated the ability to build software, create dashboards and automate tasks across different platforms.

At FYT Consulting, we have experienced these benefits ourselves. AI has allowed us to accomplish tasks that previously required specialised skills, accelerate our work and explore ideas more comprehensively.

But the AI industry has promised something much bigger: a future where autonomous AI agents manage entire business processes, companies operate with a fraction of their workforce, and perhaps even a billion-dollar multinational can be run by one person.

For employees, this raises concerns about job security. For business owners, it promises extraordinary productivity and profitability.

Yet after several years of working with GenAI, we are beginning to see a more complicated picture.

What if AI continues to become more capable, but the additional business value becomes progressively harder to realise?

The productivity gains are real, but uneven

Research has demonstrated meaningful productivity improvements from GenAI. A study involving more than 5,000 customer service agents found that AI increased productivity by approximately 14% on average, with less experienced workers benefiting considerably more.

However, the results are not universal. A 2025 METR study found that experienced software developers working on familiar projects actually took 19% longer when using AI, despite believing they were more productive. Subsequent research suggested that newer tools were improving outcomes, although the extent was difficult to measure reliably.

The lesson is not that AI doesn't work. Rather, its value depends on the task, the user and how it is deployed. And this raises several questions about the economics and sustainability of AI adoption.

1. Cheaper AI does not necessarily mean lower business costs

AI providers typically charge based on tokens, a measure of the information processed and generated. While token prices are generally published, the total processing required for a complex task can be difficult to predict, particularly when autonomous agents repeatedly search, reason, execute and correct themselves.

There is good news. According to Stanford's 2025 AI Index, the cost of running AI at a comparable level of performance fell dramatically between 2022 and 2024.

But token costs are only part of the equation.

Businesses must also account for integration, data preparation, maintenance, security, supervision and correcting mistakes. An AI agent might successfully perform 95% of a workflow, but the remaining failures could require expensive human intervention or create significant business risks.

The question is not simply whether AI can do the work, but whether it can do so more reliably and economically than the alternatives.

2. Is the information feeding AI becoming less reliable?

Much of GenAI's capability comes from learning patterns across vast collections of information, including internet content. But the internet itself is changing.

AI-generated articles, reviews, images and commentary are multiplying. Some are excellent; others are misleading, repetitive or simply wrong. As generating content becomes cheaper, producing misinformation at scale also becomes easier.

A 2024 study published in Nature demonstrated that repeatedly training AI models on AI-generated data without appropriate safeguards can lead to model collapse, where information and diversity are progressively lost.

Meanwhile, content owners are increasingly restricting AI access. In 2025, Cloudflare introduced measures allowing websites to block AI crawlers by default and publishers to charge for access.

This raises an interesting possibility. What happens if credible information becomes harder or more expensive to access, while misinformation remains freely available?

It does not necessarily mean AI outputs are already deteriorating. Developers have ways to manage these challenges, and the evidence of an industry-wide decline is not established.

But it does challenge the assumption that AI will always have increasingly comprehensive access to reliable knowledge. For users, a convincing AI-generated answer should never be confused with a verified one.

3. Autonomous AI is more complicated than impressive demonstrations suggest

One of the biggest promises of AI is autonomy: agents capable of researching customers, updating CRM records, preparing proposals, arranging meetings and managing entire workflows.

Much of this is already technically possible. Making it work consistently is another matter.

Business processes involve exceptions, changing conditions, incomplete information and competing priorities. An AI agent might perform individual tasks correctly but misunderstand the broader context, miss an important exception or make an error that affects subsequent steps.

A 2026 study of AI agents operating in Salesforce demonstrated the difficulty of completing complex enterprise tasks reliably, with even stronger approaches achieving only partial success under the study's test conditions.

These findings do not mean autonomous AI cannot work. They highlight the gap between demonstrating a capability and deploying it reliably across an organisation.

There are also practical questions about security, system compatibility, accountability and how an AI agent responds when circumstances change. The more interconnected and autonomous the workflow, the more important it becomes to manage exceptions and consequences.

4. More productivity does not always translate into more value

AI has dramatically reduced the effort required to produce reports, presentations, software and analysis. But producing more work does not necessarily mean producing better work.

Research by BetterUp Labs and Stanford's Social Media Lab identified the growing problem of workslop: polished AI-generated outputs that create additional work for colleagues who must interpret, verify or correct them.

We see a similar challenge in data analytics.

AI can help someone with little statistical knowledge build a regression model or generate an attractive dashboard. But can that person determine whether the analysis addresses the right problem, whether the assumptions are reasonable or whether the findings support the proposed decision?

Without those capabilities, organisations risk generating more analysis without gaining better understanding.

This is particularly important because the additional value from AI may diminish as we attempt to automate increasingly complex work. Reducing a two-hour task to 20 minutes may deliver obvious benefits. Eliminating the final 20 minutes could require extensive integration, supervision and safeguards that cost more than the time saved.

The technology may continue improving, while the economic returns become progressively harder to realise.

5. Can the economics sustain the AI industry's ambitions?

AI companies and investors are committing enormous resources to infrastructure, computing capacity and increasingly sophisticated models, anticipating substantial future revenues and productivity gains.

But those returns are not guaranteed.

A 2026 NBER study involving nearly 750 corporate executives found positive productivity effects from AI, but also suggested that perceived improvements were greater than measured gains.

There is also an interesting commercial tension. AI providers are investing heavily in developing more capable systems, while competition and technological progress are reducing the cost of comparable AI

capabilities.

This benefits customers, but raises questions about the profitability and sustainability of some providers.

We may eventually see consolidation, changes in pricing or companies exiting the market. A broader financial correction remains possible, although its timing and scale are uncertain.

For organisations, the practical risk is becoming overly dependent on a particular AI provider or ecosystem whose pricing, capabilities or commercial viability may change.

AI adoption should therefore consider not just today's capabilities and prices, but the long-term cost, flexibility and sustainability of the solution.

So, are we approaching diminishing returns?

Possibly. But we need to distinguish between three things.

• Technological capability: AI will likely continue becoming more capable, reliable and affordable.

• Individual productivity: AI can make people faster and more capable, although the benefits vary considerably.

• Organisational returns: The value created depends on whether those improvements outweigh the full costs, risks and complexity of deployment.

These do not necessarily move together.

The evidence does not yet establish that GenAI as a whole has reached diminishing returns. However, it is sufficient to challenge the assumption that increasingly capable AI will automatically produce proportionately greater business value.

And that distinction has important implications for how organisations should approach AI.

Beyond AI adoption: Building better decision-making capabilities

At FYT, we believe the real opportunity is not simply to adopt more AI, but to determine the right approach for each problem.

Our Decision System brings together three complementary capabilities:

• Human Judgement: Understanding the problem, recognising context, challenging assumptions and deciding what matters.

• Evidence: Using data and analytics to test hypotheses, understand relationships and evaluate alternatives.

• Automation: Using technology, including AI, to perform suitable tasks more efficiently and consistently.

Some tasks are best handled through straightforward automation. Others benefit enormously from GenAI. Decisions involving uncertainty, competing objectives or significant consequences may still require considerable human judgement, supported by credible evidence.

The same distinction applies to individuals.

Those who use AI to replace their thinking risk becoming dependent on technology they cannot properly evaluate. Those who use AI to strengthen and challenge their thinking may become considerably more effective.

GenAI is not a passing fad. It represents a remarkable advancement that will continue to transform how we live and work. But neither should we assume that every ambitious promise from the AI industry will materialise.

Like spreadsheets, databases and the internet, AI is becoming an essential part of the modern professional toolkit. It is potentially more transformative than its predecessors, but remains a means to an end.

The organisations that succeed may not be those that deploy the most AI, but those that understand when to use it, when not to, and how to combine it with sound judgement and credible evidence.

Because in an age when generating answers is becoming increasingly easy, knowing which questions to ask, what evidence to trust and how to make informed decisions may become our most valuable capability.

FYT Consulting — Building Better Decision-Making Capabilities in the Age of Data and AI.