Every technological innovation moves through a lifecycle. After an initial phase of rapid growth and adoption – fuelled by excitement, by the promise the technology holds, and at times by claims its creators oversell—comes a phase of reflection. This is the moment when we step back and ask whether the innovation has truly lived up to the expectations it set.
Information technology has seen many innovations that stopped just short of solving every problem they were heralded to solve – the internet, Java, the dotcoms, smartphones, and now AI. Each promised more than it ultimately delivered, even as it reshaped the world in lasting ways.
Like every innovation before it, AI is now entering its Reflection Phase. Until recently, it was widely regarded as a technology capable of solving most – if not all – of our problems on its own. Its creators, the media, and consumers alike – indeed almost anyone with the remotest connection to technology – have been captivated by it. AI may well be the most widely used acronym in the world today. That fascination drove remarkably fast proliferation and adoption, which is genuinely good news.
While recent disclosures on cost from Microsoft, Nvidia, and Uber have triggered this Reflection Phase, the underlying reasons are many and have been evolving for some time. The ROI gap that MIT highlighted a couple of years ago does not appear to have been meaningfully bridged. Nvidia’s report is perhaps the biggest surprise: not long ago, Jensen Huang suggested that engineers should consume tokens worth at least 50% of their compensation. It calls to mind what Adam Alter wrote early in his book Irresistible – that “people producing tech products were following the cardinal rule of drug dealing: never get high on your own supply.” Alter made the point in an entirely different context, yet the principle now seems to apply to Nvidia.
The Main Drivers of Reflection
Absence of clear goals and outcomes. AI is a significant innovation with immense potential, but it remains a resource—and a resource cannot be an objective in itself. While many companies and individuals are embracing AI, only 25% of AI-adoption projects have clearly defined outcomes and metrics. All too often, those metrics measure AI adoption rather than business impact. As a result, the remaining projects either fail or fall short of expectations. Companies are now ripping out “useless” AI add-ons after concluding they add no business value, and 78% of enterprise AI projects remain stuck in development.
Adverse impact on customer experience. Most of us have had less-than-pleasant encounters with chatbots, an early precursor to today’s AI. Too much AI can degrade the customer experience—not only through potential errors, but through over-automation. Research indicates that consumer trust is four times lower for a company that generates its brand content or services with AI rather than with people.
Missing foundations. Every AI model carries the disclaimer that it “can be wrong,” and the problem worsens when models are built on weak foundations—inaccurate or incomplete data, ill-defined workflows, and the like.Three in four executives say bad data has cost their organization $500,000 or more, and more than one-third (37%) report damages exceeding $1 million.
Employee backlash. Beyond the unintentional, inefficient use of AI that drives up costs and produces poor outcomes, cases of active sabotage by employees are beginning to emerge. Inefficient and ineffective use of AI is, in some instances, becoming deliberate.
Absence of security and governance. Partly because they receive no structured training on how to use AI, employees often adopt or deploy unapproved, unsanctioned tools, creating operational risk. Data-security exposure grows as employees upload private information to public platforms.
Accountability and liability. Liability cases stemming from faulty AI implementation are mounting—among them the class-action suits against UnitedHealthcare and McDonald’s, and other companies sued over AI breaches involving third-party vendors.
The cost of tokens. Tokens can be consumed very quickly, especially when engineers lack structured training on what to use AI for and how to use it well. The latest in a series of token shocks involves a corporation that burned through USD 500 million worth of tokens in a single month.
Further risks arise from the technology itself – still-maturing models, infrastructure reliability, and conventional issues such as code defects and inadequate checks and balances.
So does all this mean AI is dangerous and should be abandoned? Not at all. If we extend the lifecycle of technology adoption, every Reflection Phase is followed by the amendments needed to get to the next phase: Mindful Adoption. What would Mindful Adoption of AI look like? Stay tuned.
