IG Group — first generative AI in a financial organization: $100k+ annual savings

Context

I spent over 6 years at IG Group (first as Senior DS/AI Engineer, later Team Lead). Over that time we moved from “AI means classical ML and a few predictive models” to a world where LLMs are part of production infrastructure — but only because someone had to do it safely, in compliance, with measurable ROI.

For the last few years that someone was me.

What I built

Internal knowledge base chatbot

The first internal LLM application: a RAG chatbot over compliance procedures, HR policies, system instructions. Stack: LangChain + AzureOpenAI (GPT-4) + Vertex AI + ChromaDB. The hard part wasn’t RAG — it was proving answers were compliance-safe. I added a verification layer that checked each answer against FCA requirements.

AI-powered compliance verification

Tool that automatically scanned marketing communications (emails, posts, ad creative) for FCA compliance. Model classified risk level, flagged problematic phrasing, suggested fixes. Accepted by compliance as advisory (with human-in-the-loop for all high-risk cases).

Machine translation pipeline

NMT pipeline serving three use cases: SEO content (translating blog posts into 12 languages), agency communications (with marketing agencies in many countries), client communications (service messages). Per-use-case model selection, evaluation on human-rated benchmarks, costs optimized via tier-based routing.

Multi-modal AI deployments

Orchestrated evaluation, fine-tuning and deployment of multiple modalities: text generation, text-to-image, text-to-video, translation, audio generation. Each went through compliance review.

Trading opportunity identification (classical ML, big value)

Data-driven system identifying clients with open stock positions likely to trade around earnings announcements. Statistical analysis, A/B testing, predictive analytics. Collaborated with the CRM team. ~$40,000 in additional monthly revenue from the targeted campaign.

Client lifetime value & churn scoring

Advanced analytics + pySpark + GCP ML services. Models scoring customer value across lifecycle stages, supporting client-facing, marketing, and CRM teams in prioritization. ~$250,000 annual savings through better retention resource allocation.

Aggregate outcomes

• $100k+ annual savings from the generative AI stack.
• $40k monthly added revenue from the trading opportunities model.
• $250k annual savings from client scoring.
• Hundreds of manual hours/month eliminated in compliance, marketing, customer support.
• Cultural acceleration — paper reading sessions, internal tech talks, mentorship of juniors.

What I learned

• Compliance-first AI in finance is an edge, not a brake — companies who do it well can ship faster than competitors stuck in PoC purgatory.
• The first LLM in an organization is a political test, not a technical one — success depends on CFO/CCO/Legal trust, not on accuracy.
• Classical ML still wins on dollar value — the most money I made over 6 years came from a simple client scoring model, not an LLM.