A conversation between Helen Murphy and George Deskas, as part of AI Beyond Blind Trust series.
Most organisations don’t struggle to collect information, they struggle to use it. Too much time is still spent manually pulling key details from documents, slowing teams down and creating unnecessary friction.
AI information extraction helps change that. It takes repetitive, manual work and turns it into something far more efficient and scalable.
But success isn’t just about the technology. It comes down to how it’s applied, governed and trusted in day-to-day use.
Helen Murphy (Head of Marketing) spoke with George Deskas (Data Science Manager) about what good looks like in practice, and how organisations can start seeing value quickly, without overcomplicating things.
Let’s unpack information extraction
Helen: One of the first questions people ask is, how do we know the output is right?
George: That’s exactly the right concern. Trust doesn’t come from switching AI on overnight. It comes from evaluation. In practice, we run AI extraction alongside manual processes initially and review the outputs heavily. As accuracy and consistency are proven, reliance on manual review reduces. Governance is central to that journey – it’s not a tick‑box exercise.
Helen: What happens when the documents are really messy?
George:
That’s most of the time. Scanned documents, poor-quality images and inconsistent layouts are the norm. The key is being honest at the start about how messy the data really is and doing the preparation work properly. Models are only as good as the data you feed into them.
Helen: People often describe AI as a black box.
George:
It can feel that way if you don’t understand it. But the more teams learn how extraction works, what it’s good at and where it struggles, the more confidence they have. Education builds trust. Once people realise it isn’t magic, they’re much more comfortable relying on it.
Helen: Who should actually own an information extraction solution?
George:
That’s one of the hardest questions, and it’s rarely just a technology answer. Building the AI is often the simplest part. The bigger challenge is deciding who owns it once it’s live –how it’s governed, how performance is monitored, and how it stays aligned to business outcomes. In practice, it works best as a shared capability between technology and the business, not something owned in isolation.
Helen: How quickly do organisations usually see value?
George:
When the use case is well defined, value comes quickly. The mistake is trying to do everything at once. Start with a narrow, high‑impact problem, prove it works, then scale. That’s how you build confidence without unnecessary risk.
Helen: Is this mainly for financial services?
George:
Not at all. Any process where a human has to read unstructured information to extract specific details can benefit – HR, customer service, marketing, operations, research. The problem isn’t regulation. It’s friction.
Helen: What kinds of use cases are best suited to information extraction?
George:
Usually, the strongest use cases are the ones where the output is clear. If you need to extract names, addresses, dates, values or other defined fields from documents, information extraction can be very effective. The more subjective the task becomes, the more careful you need to be about expectations and design. That’s why the best starting points are often high-volume processes with structured outcomes and obvious operational value.
Helen: What does it take organisationally to make this work?
Extracting what really matters
Information extraction isn’t about chasing AI trends. It’s about removing friction — in a way people can understand and trust.
When that happens, the real value becomes clear. Teams spend less time extracting information, and more time using it to make better decisions. And that’s when adoption follows naturally.
If you’re starting to explore this space, it can help to see what this looks like in practice. You can read more about how organisations are using information extraction to unlock value from unstructured data, here.
You can also explore the wider Beyond Blind Trust series, which looks at how organisations are building confidence and trust in AI, here.
And if it would be useful to talk it through, you can speak to one of our AI experts, we’d be happy to share what’s working in practice and where to start.