UK government backs Onteric to build the world’s first AI agents that can prove every decision

The potential for how AI, and especially agentic AI, will change the world continue to dominate the news. But the reality is that 88% of agentic AI pilots aren’t even making it out of testing. And almost all businesses are a long way off realising the productivity gains their boards are pressurising them for.

Why such a high fail rate, aside from the obvious issues like misunderstood incumbent process and lack of data readiness?

Gen AI is very powerful, but that doesn’t mean it’s ‘successful’. If you kick off an agentic agent where the large language model (LLM) is as high as 95% correct, in just a 20-step workflow, due to compounding errors, that erodes to just 36% success. In other words, it will be wrong twice as often as it is right.

Secondly, and less well understood, is that LLMs are stochastic. They best guess the next answer but are built so that the answer is inconsistent. That means the agent can never be dependable. And never reconstructed – a requirement cited in the Financial Conduct Authority’s Mills review.

And the final side effect of the stochastic nature of LLMs is that they can’t be audited. You can ask them what they think they did, but that’s very different from an actual trail of action. It’s a probable, not a provable, answer. The Mills Review also says that the answer “the AI said so” won’t cut it with regulators.

Onteric has secured highly competitive government funding under the AI Champions: Frontier AI initiative to build a solution to these issues. Our Neuro-symbolic AI is built, not retrofitted, to financial services, combining the power of generative AI with formal logic and rigour. This means our agents have:

1. Strict adherence to rules, so you can control, audit and govern output
2. Transparent reasoning and assumption
3. Learn from changes and remember corrections
4. No need for guardrails, because the rules can’t be breached.

Agentic processes follow three stages: perception, reasoning, and action. With a team of PhD AI and data experts, Onteric has spent the last three years building the solution to the perception layer. Why? A perfectly applied rule on a misread figure gives you a confident, auditable, wrong answer.

Their solution is live with regulated clients and proven to deliver 3.5x ROI while increasing conversion, and even more driving hours of time saving per mortgage case. Anywhere there is manual reading and analysis of documents, emails or websites Onteric can free up your team to focus on what drives revenue.

“What previously required significant staff time can now be completed in seconds,” said Frances McCann, chief executive officer of Scotwest Credit Union.

The grant enables Onteric to extend that same architecture to deliver a full agentic solution because an answer without proof is just a guess.

Jen Lothian, chief executive officer and co-founder of Onteric, said: “Imagine how you could transform your business and remove all the heavy-lifting, morale-sucking admin if you could actually trust an AI agent to be right. Plus, you can audit it just to double check (and keep compliance happy).”

If you want to learn more, Jen Lothian will be speaking with Andy Oldham, chief executive officer of Mojo Mortgages in The Engine Room on Wednesday 14 October at Open Banking Expo UK & Europe, on why AI pilots are failing and how to actually get a positive ROI. They will be using key findings from Mojo’s live use case with Onteric.