The hidden cost of AI done wrong: why implementation isn't enough 

Businesses across every sector are pouring money into AI. Global spending is climbing, with worldwide AI spending forecast to total $2.59 trillion in 2026, a 47% year-on-year increase, according to Gartner, Inc., and boardrooms are awash with enthusiasm for efficiency gains and the promise of transformative productivity benefits. Yet for all that investment, a quiet and expensive problem is hiding in plain sight: a significant proportion of AI implementations are failing because the people using it aren’t equipped to work with it effectively. 

 

Until businesses address this, they’ll continue wasting significant resources on platforms that underdeliver, outputs that mislead, and decisions that require costly corrections. 

 

Research shows that 23% of AI responses contain inaccurate information, and 31% of automated decisions require human correction. In industries like insurance, where decisions directly affect customers at their most vulnerable moments, those error rates can translate into real consequences: delayed claims, unfair outcomes, and eroded trust, both with customers and the regulator. 

 

It would be tempting, when faced with figures like these, to question the value of AI altogether. But that misses the point entirely. 

 

AI is extraordinarily good at processing vast datasets, surfacing patterns and generating recommendations at a speed no human could match. What it lacks is human empathy and contextual understanding. A system might flag a claim as suspicious based on statistical patterns. At the same time, a seasoned handler would immediately recognise that the circumstances are entirely consistent with the customer’s history and unique circumstances. 

 

This is where so many implementations go wrong. Organisations deploy sophisticated tools and, consciously or not, treat their outputs as gospel. Staff, unsure how the technology works or lacking training, follow the recommendation rather than interrogate it. 

 

The hidden costs of poor implementation 

I hear of organisations spending hundreds of thousands of pounds on sophisticated AI platforms, only to find them largely unused months after launch. Staff, uncertain about how to interpret outputs or uncomfortable with tools they don’t understand, default to the manual processes they’re most comfortable with. The investment sits idle. And AI gets the blame for a failure that was fundamentally one of implementation, not capability. 

 

Poor interpretation can be just as costly as non-use. When employees don’t understand that AI can inherit biases from its training data, struggle with unfamiliar cases, or generate outputs that at first appear correct but aren’t, they’re poorly equipped to apply the critical scrutiny those outputs require. 

 

A 2025 McKinsey report found that more than a fifth of employees have received little to no support when working with AI. That figure should give any business leader pause, when it suggests a substantial portion of the workforce is being asked to engage with technology without the tools or training to do so effectively. 

 

Interpretation is a skill that needs to be developed 

The solution isn’t to slow down AI adoption, but instead to invest as seriously in human capability as in technological capability. 

 

At Avantia Group, when we developed Holmes, our proprietary AI tool built to support the claims process, we were explicit from the outset that it was designed to assist handlers, not replace them. Holmes can analyse vast amounts of data in seconds, flag potential fraud with six times the accuracy of previous methods, and complete payment calculations with 98% accuracy. But every final decision remains with a human claims expert who has the judgement, experience, and contextual understanding to know when to trust the output and when to challenge it. 

 

That human oversight is the source of the system’s real value, because every intervention teaches it how to perform better. An AI system without informed human oversight isn’t a smarter operation; it’s an unsupervised one who’s performance is static. This principle doesn’t change as technology becomes more sophisticated; if anything, the need for highly skilled human oversight grows. 

 

Building the confidence to challenge 

For businesses looking to address this gap, the starting point is straightforward; employees at every level need to understand how to properly use AI tools, how those tools work, and their limitations. 

 

Staff should understand concepts like training data and model bias; they should know that AI is built to find patterns but can struggle with context, nuance, and situations it hasn’t encountered before. Crucially, they need the confidence to treat an AI recommendation as one input in a broader decision-making process. 

 

At Avantia, we’ve rolled out AI training across all departments, covering fundamentals and the practicalities of working with generative models. We’ve also embedded data scientists directly into operational teams, because the gap between technical capability and practical application closes fastest when those with deep AI expertise work alongside those with domain expertise. The data scientist learns what the business needs; the claims handler learns what the technology can and cannot do. 

 

Getting implementation right 

AI will continue to advance, and the pressure on businesses to adopt it will only intensify. But the organisations that extract lasting value from these tools will be those who invest in building a workforce equipped to work with AI critically and confidently. 

 

---

 

Dan Huddart is Chief Technology Officer at Avantia Group 

 

Main image courtesy of iStockPhoto.com and ismagilov