The Soaring Cost of 'Making AI Work'
The initial excitement of building an AI model is often replaced by a 'rude awakening' when it comes to making it operational, according to Gartner. The firm's finding that the cost to achieve positive returns from AI has increased 18 times in just two
years highlights a critical challenge. This isn't about the initial development, which is becoming cheaper, but about 'operationalising' the technology. Operationalising means moving an AI system from a controlled lab environment into the messy reality of live business operations. It involves continuous data pipelines, integration with legacy systems, constant monitoring, maintenance, and retraining. Gartner predicts that by 2028, at least half of all generative AI projects will run over budget because companies underestimate these complexities. The day-to-day cost of running AI, known as inference, is now expected to account for the majority of a model's lifetime costs, fundamentally shifting the economic equation for businesses.
A Dangerous Governance Gap
The second half of the story is even more concerning. While companies are eagerly deploying more autonomous 'agentic' AI, the oversight required to manage it is critically absent. A recent Gartner survey found that while 71% of organisations plan to deploy agentic AI within the next year, only 21% currently have the necessary governance capabilities in place. The headline's 14% figure appears to be a related, and equally alarming, statistic from Gartner's wider research into CIO preparedness, which found only 13% of CIOs report their enterprise has realised significant value from AI so far. This massive gap between ambition and readiness means that for many organisations, AI is creating risk faster than it is creating value. Without clear rules, accountability, and safety checks, companies are flying blind, exposing themselves to spiralling costs, service disruptions, and other unforeseen consequences.
Why AI Projects Spiral Out of Control
The connection between skyrocketing costs and poor governance is direct. A lack of strategic oversight is a primary driver of budget overruns. Many projects fail because they move from a successful pilot to production without a clear plan for scale. The infrastructure needed for 100 simultaneous users can be ten times that for 10 users, a non-linear leap that many budgets don't account for. Other common culprits include poor data quality, which is the single most cited reason for AI failure, and the immense difficulty of integrating new AI tools with decades-old legacy systems. Furthermore, a lack of clear exit criteria means low-value projects are often allowed to continue draining resources. Gartner found that only 31% of CIOs have rules for stopping a pilot from becoming a full-blown project, and a mere 17% have criteria for shutting down a live AI initiative that isn't delivering.
Building a Bridge to Value
To close the gap, experts advise a fundamental shift in thinking. AI adoption must be treated as a systems engineering discipline, not just a technology experiment. This starts with establishing a centralised governance authority to coordinate accountability, technology choices, and economic decisions across the business. This body would set the strategy, define risk tolerance, and ensure that AI initiatives are aligned with real business outcomes. Companies also need to become ruthless about measuring value. This includes defining success—and failure—before a project even begins. Rather than just adopting AI for automation's sake, leaders should focus on redesigning workflows and business models to truly leverage the technology's capabilities. Ultimately, the solution is organisational before it is technical. Putting the brakes on a car allows it to go fast safely; the same principle applies to AI.
















