The most valuable use of AI in financial services is far from the customer experience.

Most customers would imagine that AI in financial services means chatbots. And that’s understandable, as it is still most likely, the version of AI they know. 

Yet, both our experience and wider market research shows us that back-office automation is where AI is quietly having the biggest impact.  In retail banking, it elevates financial crime detection and customer contact centre response, in business banking it provides enhanced business lending and risk management, and in capital markets, it’s starting to impact trading, risk management and compliance.

We saw the same pattern surface repeatedly while researching our own AI in the SDLC whitepaper this year, and it holds just as true when you read across the latest reports from major UK insurers, banks, wealth managers and pension providers. 

Also, tellingly there are just as many win themes and stories that relate to machine learning and data science, as there is from the ‘hype machine’ that is generative AI

Where the value is landing

If you look past the proof-of-concepts and the pilots, the pattern across the sector is pretty consistent.

 Aviva‘s medical report summarisation tool, used across its protection underwriting business, cuts assessment time by roughly half while matching human underwriting decisions 99.9% of the time. That’s a structural change to how quickly the business can say yes.

In service operations, NatWest now generates AI call and complaint summaries across nine thousand support agents, saving over 70,000 hours a year and cutting complaint resolution times by nearly 20 minutes per case. St. James’s Place uses meeting intelligence tools to turn adviser conversations directly into compliant, structured records, and People’s Partnership has layered real-time sentiment analysis and note summarisation onto its existing contact centre platform.

Coding assistants rolled out to thousands of developers inside regulated banks are shortening deployment cycles and delivering early productivity gains in the region of 10 to 20%, but many organisations are as worried about the challenges of viable context, effective guardrails and the ever moving target of cyber security. 

How this will manifest in safe, secure and compounding gains is yet to be actualised but with NatWest giving over 12,000 engineers AI coding assistants, and Barclays equipping 19,000 developers with similar tooling, you can see the direction of travel and the hard work that is being undertaken to ensure that the blueprint for success is underpinning the work in a substrate.

Further into the back office in the risk and compliance function, Schroders has built an internal AI assistant to challenge risk assessments and summarise counterparty news for credit monitoring, work that used to consume days of analyst time. 

While these tools may lack the ability to deliver exciting press releases to market, they are directly reducing costs, cycle times, and error rates.

The governance dividend

The organisations that are getting the most out of back-office AI treat governance as an accelerant. 

Dedicated AI ethics functions, responsible AI committees sitting inside group risk, and central inventories that triage use cases before deployment are becoming standard features at the more advanced firms in the sector, rather than obstacles that are bolted on afterwards. Our own research reached the same conclusion. The bottleneck in regulated organisations is rarely the technology, and success depends on governance and delivery teams moving together, with guardrails built in from the start. We’ve made this case at length previously –  Regulation is not getting in the way. It is the way.

The same governance-first strategy holds even when the AI sits closer to the customer. When Aegon approached us to prototype a generative AI concept for customer engagement without triggering the months-long formal risk assessment its size and sector would normally demand, we built it inside a hosted, managed prototype environment designed specifically to contain that risk. The prototype was working within two days, and the production-ready version followed in two weeks. Governance didn’t slow us down, it was the thing that made moving fast defensible.

That said, nothing replaces sign-off. Deploying AI into underwriting, claims or risk workflows still needs your own compliance, legal and risk functions in the room early.

This isn’t theoretical for us

We’ve built this AI, in these conditions.

Economic crime and fraud. For one major bank’s financial crime team, we built an agentic AI system that runs automated, independent investigations on business accounts flagged during onboarding or account changes, checking its own conclusions against a sample of human analyst decisions before it’s trusted at volume.

Pensions administration. When a pension provider needed to onboard an entire scheme’s membership during a bulk annuity buy-out, a key challenge was reconciling records across multiple legacy administration systems into one accurate customer view, data engineering that determines whether a provider can actually meet its Pensions Dashboard and Consumer Duty obligations on time.

A banking carve-out. During a major ownership change in retail banking, we helped migrate a legacy data estate onto modern cloud infrastructure while keeping business-critical regulatory reporting running automatically throughout.

What good back-office AI looks like

The approach is narrower than most AI strategies admit:

Pick a specific, high-volume decision, not a department. Sales  report summarisation, entity matching, complaint triage. Not “AI for underwriting.”

Instrument the exception rate, not just the headline time saved. A tool that halves assessment time but quietly increases the volume of cases needing manual override hasn’t actually saved anyone anything.

Build the governance function alongside the pilot. Retrofitting a risk framework after a use case is already live is where most of the delay comes from.

Be honest about build versus buy. Research from MIT’s Project NANDA (Networked Agents and Decentralised Architecture) shows that internally built AI projects fail roughly twice as often as those developed with an experienced delivery partner. Be deliberate about where you build internal capability, and where you choose a trusted technology partner.

A less glamorous kind of return

Back-office AI rarely gets a launch event, and nobody demos a claims summarisation tool at a keynote the way they demo a chatbot. That’s why it’s underfunded relative to the return it’s generating, and the reason why the organisations investing there now are going to be harder to catch in two years’ time.

If you’re trying to work out where in your organisation back-office AI will move the numbers, we’d be glad to talk it through. Our AI in the SDLC research goes into more depth on what realistic gains look like inside regulated, brownfield environments, and our recent piece on de-risking pension mergers with AI-powered entity matching walks through one of these problems end to end.