AI value gap linked to operating models, PwC Cyprus analysis says

Nicosia, Cyprus. Organisations leading in artificial intelligence are widening gaps in cost, growth and margins by redesigning their operating models, according to an analysis by PwC Cyprus Senior Manager Yiannis Stavrianos. The analysis identifies three failures that limit AI returns, beginning with insufficient workflow redesign.

Technology and workflow redesign

The analysis cited audited financial impacts including about $60 million in compressed customer-operations costs at Klarna, about $2 billion in annual benefits at JPMorgan, and GitHub Copilot’s use by approximately 90 per cent of Fortune 100 companies.

It said the leaders are not using fundamentally different models but are rebuilding operating models around them.

PwC’s 2026 Digital Trends in Operations survey of 767 US operations leaders found that 89 per cent said their technology investments had not fully delivered expected results. The figure indicates that most organisations have invested in technology but have not yet received the anticipated returns.

PwC’s 2026 AI Business Predictions estimates that technology accounts for about 20 per cent of the value of an AI initiative, while the remaining 80 per cent comes from redesigning workflows around the technology’s capabilities.

Procurement approach

The analysis said a common approach is to issue a request for proposals, assess vendors, select a platform, sign a contract and deploy it across an unchanged workflow. While the technology is installed, key performance indicators do not move, it said, and budgets may then be reassigned after pilots fail to produce expected returns.

Klarna’s savings did not result solely from purchasing an OpenAI subscription, the analysis said. They followed a redesign of the customer service operation, including new role definitions, escalation criteria, performance metrics and training for staff taking on more complex work.

PwC’s 2026 findings showed that 27 per cent of operations leaders had fully embedded an AI strategy across business units, while 37 per cent were comfortable assigning AI agents to execute complete end-to-end processes. The analysis described these as measures of organisational rather than technological readiness.

Compounding errors in AI agents

The second issue identified is the compounding effect of errors in agentic AI systems operating autonomously across multiple steps.

A task that is 95 per cent reliable at each individual step would produce about 60 per cent reliability across a 10-step chain and about 36 per cent reliability across a 20-step chain, the analysis said.

Without human checkpoints to reset the error rate, longer autonomous chains will eventually fail, it said. An agent that makes an error may not recognise it and can generate an incorrect invoice, recommendation or customer email before continuing to the next task.