SBS Insurance Services claims to have successfully used a new AI system that automates digital claims processing in insurance. Developed by UK firm UnlikelyAI, the system automated 40% of SBS’s digital claims with 99% accuracy during a pilot programme. Each automated decision was fully auditable. The pilot highlighted an approach to enterprise AI that steers […]
SBS Insurance Services claims to have successfully used a new AI system that automates digital claims processing in insurance.
Developed by UK firm UnlikelyAI, the system automated 40% of SBS’s digital claims with 99% accuracy during a pilot programme. Each automated decision was fully auditable.
The pilot highlighted an approach to enterprise AI that steers away from traditional large language models (LLMs), which have drawn scrutiny for their lack of reliability and transparency.
UnlikelyAI instead combines neural networks with symbolic reasoning, a rules-based method that the firm claims to deliver consistent, explainable outcomes.
“We built UnlikelyAI to address the fundamental limitations of large language models, especially in high-stakes industries where trust, precision and transparency are essential,” said William Tunstall-Pedoe, founder and CEO of UnlikelyAI.
“This deployment – our first to be made public – demonstrates that AI doesn’t have to be a black box; it can be transparent, explainable and genuinely accountable.”
SBS said it chose UnlikelyAI’s platform to enhance efficiency while maintaining compliance and human oversight where necessary.
“We’ve always prioritised excellent customer services and using technology as an enabler for that,” said Bazil Crowley, director of innovation at SBS Insurance Services.
“UnlikelyAI’s neurosymbolic approach stood out as a game-changer for our operations, with this partnership allowing us to deliver a faster, more reliable service to our customers whilst maintaining our human touch where it matters most.”
AI in insurance claims processing: PwC
UnlikelyAI’s recent research shows some of the key shortcomings of LLMs in regulated environments. According to OpenAI itself, even advanced models like GPT-4.5 can produce incorrect answers in up to 37% of factual queries.
Internal tests also showed significant variability in claim outcomes when identical inputs were run through different LLMs. The results ranged from 8% to 46% coverage depending on the model, with occasional inconsistencies even within the same model.
By contrast, UnlikelyAI reports that its rule-based system reduced categorisation errors by 99% while enabling traceable decision-making.
