The growing gap between promise and practice

Artificial intelligence has often been touted as the decade's defining technology. Venture capital has poured billions into AI startups, corporations have announced ambitious AI transformations, and governments have invested heavily in national AI strategies.

Adoption of AI has been rapid since its widespread emergence in AI tools in late 2022. A vast majority of organisations have now started using GenAI in some form, be it through off-the-shelf tools designed to generate content, write code or summarise information.

The costs of running generative AI do not currently make sense, as several. Reports have suggested that many leading generative AI providers continue to face high operating costs, particularly related to computing infrastructure, model training and inference. Some analysts have questioned whether current revenue growth can sustainably offset these expenses in the long term.

Hence, a growing number of reports and many industry surveys appear to suggest that many AI initiatives never make it beyond the pilot stage. Some analysts have even begun describing this phenomenon as “pilot purgatory” — a situation where projects generate excitement, funding and presentations, but never become a part of everyday operations.

Rather than indicating that AI itself lacks values, these trends may point toward deeper organisational and human challenges.

Several widely cited studies have suggested unusually high failure or abandonment rates for AI initiatives: research frequently linked to the RAND Corporation reportedly suggested that more than 80% of AI projects fail, and roughly twice the failure rate of traditional software projects.

Analysts at Gartner predicted that at least 30% of generative AI projects could be abandoned after proof-of-concept stages because of unclear business value, poor data quality, rising costs or inadequate risk controls.

Other industry reports have claimed that 70% to 95% of AI pilots never reach full production. These figures vary significantly depending on methodology, making definitive conclusions difficult. However, they collectively suggest a recurring pattern: many organisations appear enthusiastic about starting AI projects but struggle to scale them.

The biggest and most major challenge for organisations is in justifying the substantial investment in GenAI for productivity enhancement, which can be difficult to directly translate into financial benefit, according to Gartner.

The cost of everything that happens from when you put a prompt in to generate an output from an AI model is rising rapidly, thanks to the token-heavy generations that are necessary for “reasoning” models to generate their output, and with reasoning being the only way to get “better” outputs, they're here to stay.

Meaning every interaction with a generative AI model consumes computing resources. More advanced ‘reasoning’ models often require significantly more processing power and, therefore, higher operational costs. As these systems become more sophisticated, the infrastructure required to support them may also become more expensive.

An AI model may perform exactly as intended while the business case surrounding it fails to generate sufficient returns. In this sense, some observers argue that the challenge facing the industry is not necessarily technological capability, but commercial viability.

Some observers argue that the recent generative AI boom created an atmosphere where companies felt compelled to launch AI initiatives simply because competitors were doing so.

In this environment, executives may have faced pressure from investors, boards and shareholders to demonstrate AI adoption, even before identifying clear business problems the technology could solve.

As a result, some projects may have begun with the question- “How can we use AI?” rather than “What problem are we trying to solve that might need AI?” Several analysts suggest that this reversal could be one reason why projects struggle to show measurable returns.

One recurring problem across reports is that many AI failures may actually be management failures. Some analysts argue that leadership teams sometimes tend to underestimate the training requirements, workflow redesign, employee adoption and long-term maintenance costs.

Here, AI becomes the mirror that reflects already existing organisational weaknesses.

A pilot may reveal fragmented databases, unclear accountability structures, poor communication between departments, or even unrealistic expectations about automation. What the public sees as an AI failure may instead be a broader operational problem.

These organisational challenges may help explain why measurable returns remain difficult to achieve.

Now, less than 1% of executives report a 20% or higher ROI from AI, and 53% report mostly 1-5% ROI around 60% of organisations see little to no value, and roughly 30% of generative AI projects are being abandoned after proof-of-concept. Companies are spending heavily but are struggling to show it's working.

While AI demonstrations can appear impressive, production deployment often has requirements like cloud infrastructure, cybersecurity safeguards, compliance systems, human oversight, continuous retraining and integration with existing software.

Some organisations initially assume that AI would immediately help reduce costs, but several reports have suggested that companies have discovered that AI often creates new expenses before delivering measurable results.

Like, Data centers are being built, but revenue isn't there to cover the power bills. Usage-based pricing is starting to emerge because flat-free models are losing them money, like if a developer uses too many tokens, the company loses money.

Simultaneously, it may be worth noting that many transformative technologies initially struggled to demonstrate immediate profitability. The internet itself experienced periods of intense speculation and investment before sustainable business models emerged. Some AI advocates argue that the industry may currently be experiencing a similar adjustment period.

It may be a bit early to conclude that artificial intelligence is failing. However, growing evidence appears to suggest that many AI initiatives are failing to cross the gap between the promises made and the performance in practice.

The recurring pattern across the industry reports seems to indicate that technology alone is rarely enough. Data quality, governance, employee trust and organisational culture may ultimately determine whether an AI project succeeds or quietly disappears before launch.

If these assessments are accurate, the future of AI may depend less on building smarter models and more on building environments capable of using them responsibly and effectively.

Sources:

. . .