AI-Powered Geopolitical Risk Tool Boosts Efficiency and Reduces Risk

A proof-of-concept generative AI tool that QuantSpark built for a public sector client processed over one million pieces of open-source content within hours, a task that would otherwise have absorbed roughly 9.5 years of continuous individual manual effort. Test scenarios confirmed the tool reproduced the same geopolitical insights as the client's existing manual data-gathering method, and it did so while cutting the operational risk and analytical-talent bottleneck that the old process created. That is the headline. Everything below explains how it was built and why it mattered.

The problem

The client wanted to put data and technology at the centre of its operations, but its existing approach to gathering geopolitical intelligence stood in the way. Analysis depended on laborious manual review of open-source material, a process that was slow, carried real operational risk, and tied up skilled analytical staff in repetitive work rather than higher-value judgement. That over-utilisation of talent was hindering the effectiveness of the whole department.

The brief was specific: give the client rapid, comprehensive geopolitical information in a format non-specialists could use, and do it via a proof-of-concept that would demonstrate whether generative AI could analyse international relations from open-source data at scale.

The approach

QuantSpark ran the engagement as a design-thinking, product-led build. Close collaboration with stakeholders identified where a specific part of the client's operations stood to gain most, and scoped the proof-of-concept around it rather than trying to solve everything at once. An interdisciplinary team, spanning machine learning, analytics, cloud engineering, software development, product, graphic design, UI/UX, commercial strategy and strategic insight, delivered the working proof-of-concept within a few weeks.

Two workstreams ran in parallel. The first was a data pipeline: unstructured open-source content sourced, then routed through an LLM-based extraction engine that identified and structured the relevant information, with the output stored in a relational database. The second was the front end, shaped through product-led workshops directly with end users rather than designed in isolation. That collaboration produced a bespoke web application letting people build their own visualisations, inspect raw tables, and pull up the original source text behind any given data point.

The pipeline and the application together form a simple workflow: open-source content goes in, the LLM engine extracts and structures it into the database, the web application serves it back out as visualisations, raw tables and linked source texts, and non-technical users take it from there, asking their own questions and passing insights downstream without needing an analyst in the loop for every request. Secure, indirect data-handling techniques were built into that flow to bring down the operational risk that had been inherent in the manual version.

What it delivered

Against that design, the results were validated rather than asserted: test scenarios showed the tool replicating the same geopolitical insights the client's manual process already produced, on volumes of over a million pieces of content processed within hours. The 9.5-year figure is a manual-effort-equivalent calculation, benchmarking that throughput against how long an individual would need to review, extract and structure the same volume by hand, and expressing the difference in continuous person-years. It is a modelled comparison illustrating the scale of the automation, not a cash saving; no salary-based or monetary figure was calculated.

Beyond raw speed, the proof-of-concept reduced operational risk through its secure handling techniques and empowered users directly, giving non-technical staff a dynamic front end to interrogate data and communicate insights onward without waiting on specialist bandwidth. For the client, it represented a pioneering first step: a new, data-centric capability with the potential to enhance existing functions through both efficiency gains and insights the manual process was never positioned to surface.