Case Study - Reducing hallucination rates on complex analytical tasks

How to get ahead of the AI shockwave

The past few years has seen a deluge of agentic development proclaiming to upend everything from protein folding to self driving cars, and everything in-between, leading to pretty much everyone demanding their organisation somehow exploit this new magic.

In reality, 95% of these efforts fall by the wayside, busywork disguised as game changing in an effort to meet the impossible expectations driven by hype rather than need.

But, and there is a but - that 5%. That 5% is where genuine, industry changing value is found. Think chat agents, think material engineering – think big data.

It’s that last part where AdTecher comes in, in this case to help solve a dire noise/signal ratio that PushGroup were struggling with.

Building an integrated and harmonised AI data layer

PushGroup are an industry leading performance marketing agency group based in London who are at the forefront of finding real world application for LLMs in this space. Much like that 95%, what they found were endless wrappers, all suffering from the unacceptably high hallucination rate of around 20%.

The consensus was that this is essentially ‘The Great Filter’ for LLM’s in MarTech, with most initiatives ending up as pipedreams.

Integrated Solution & Value Addition

By introducing a dedicated data layer & tailored RAG architecture we were able to reduce the hallucination rate from around 20 percent to under 1 percent.

This created the foundation for material improvements in time to correct underperforming campaigns, uplift in ROAS compared to control campaigns, and improved status of volatile accounts. With this foundation in place the prospect of a decisive competitive advantage starts to emerge.

30-11-2025