User-obsessed strategy and the role of AI according to Monzo Bank · LAB

Key takeaways from Runroom LAB: Understanding complexity through data

Balancing AI and human research

The real role of “Synthetic Personas” in strategy

Drawing from their experience at Monzo Bank, Caroline Wilcock (Senior Lead Researcher), Tomaso Vido (Lead User Researcher), and Izak Nel (Lead Data Scientist) revealed that the true competitive advantage still lies in deeply understanding people.

Below are the main conclusions from the event, how Monzo balances AI with human research, and the innovative role synthetic personas play in their strategy.

Understanding complexity through data

Monzo is one of the leading digital banks in the UK, known for its high user satisfaction driven by strong investment in UX teams. When facing the complexity of designing products for businesses (from small cafés to consulting agencies or construction companies), they discovered that financial needs and user problems vary drastically.

To make sense of this complexity, Monzo uses meaningful segmentation: heuristics that simplify populations into distinct groups to intuitively understand their problems and values.

This process is neither magical nor automatic—it requires a constant “ping-pong of insights,” combining quantitative behavioral data from millions of users with deep qualitative research, fresh interviews, and surveys.

Balancing AI and human research

One of Monzo’s biggest lessons is that AI cannot replace the deep analytical work of human researchers.

In fact, when the Monzo team attempted to delegate market segmentation entirely to AI, the result was poor: noisy, unstable outputs with little real alignment to customers.

As a result, Monzo has established a clear balance:

• What remains uniquely human: empathy, deep initial analysis, and storytelling. As the team explains, AI cannot determine what is truly meaningful to people—creating meaning is a strictly human capability.
• AI as an assistant, not a creator: Monzo uses AI effectively to “stress-test” research. Researchers build frameworks and ask AI to challenge them, validating findings and identifying gaps.

“Empathy is something I hope AI will always fail at [...] A synthetic persona should not replace research—it should be a tool to ideate and play with ideas. You cannot validate a final product with a synthetic persona because you wouldn’t be able to empathize with how people actually experience what you’re building.” — Tomaso Vido

The real role of “Synthetic Personas”

To scale research insights across a company of over 5,000 employees, Monzo adopted an innovative solution to avoid insights getting lost in long presentations: synthetic personas.

However, their approach is far from AI solutionism. Synthetic personas at Monzo follow strict principles:

• An interactive repository, not a replacement: They allow engineers and Product Managers to interact, ask questions, and ideate based on data—but never replace real research or validate final products.
• Powered by transcripts, not summaries: To avoid losing nuance or generating hallucinations, they are fed with raw interview transcripts, preserving tone, doubts, and emotions.
• They must say “I don’t know”: Good prompting forces synthetic personas to admit uncertainty. AI tends to agree by default, but real value comes from identifying gaps in understanding.
• Disposable tools: They are created ad hoc for specific customer profiles and have a short lifecycle, evolving as strategy and data evolve.

Caroline Wilcock on what she would never delegate to AI:

“My instinct goes straight to initial analysis. Especially in high-risk areas, it’s on us as human researchers to look at the dataset and extract what we believe is happening thematically [...] If you let AI lead there, it’s too easy to get lost—and it doesn’t help the organization truly understand the problem.”

Izak Nel also shared that they are currently experimenting with platforms to host synthetic personas:

• Gemini Gems: Currently in use. Strong for personality simulation and interactivity, though carefully monitored for hallucinations or external data leakage.
• NotebookLM: Tested but discarded for this use case due to lack of interactivity and limited personality simulation.

However, they emphasize that the tool itself is not the key factor. What truly matters is the quality of the knowledge base and insights feeding the AI.

As Caroline notes, real value comes from grounding the system in solid data and real interview transcripts—not generic summaries.

Conclusion

Monzo’s approach shows that AI is an invaluable ally for processing, challenging, and democratizing research. However, user-centered strategy must always remain anchored in real human work, empathy, and everyday problems.