To realise an AI-powered vision of the future, banks will need to embrace and enable agentic AI by tackling roadblocks such as legacy systems, inconsistent data quality, and stakeholder resistance.
And they’ll have to do so in an environment that’s changing rapidly. When ChatGPT went mainstream in 2023, human–AI interaction reached a new intensity and saw the introduction of new risks like prompt injection. Today, AI is everywhere, but systems are harder than ever to control.
The mismatch between potential and progress is reflected in the lack of effective AI solutions. A recent MIT study found that 95% of generative AI (GenAI) pilots fail to go into production.
Yet it doesn’t need to be this way.
At Zühlke, our work with leading financial institutions over the past few years has shown that the success factors have not changed with the ChatGPT moment. We see four timeless principles that continue to determine who scales — and who stalls.
Four timeless principles
Timeless Principle 1: You should work backwards from value
In a world where anything is possible, the question is no longer what can we build — but what should we build?
Generative AI and user-friendly tools like low-code platforms have removed barriers to experimentation. Pilots are easy to start, but hard to scale. Without strategic alignment, they consume resources and deliver little value.
Successful institutions guide their teams in what to build — and what not to build. This means building a portfolio of high-impact use cases directly linked to business strategy.
Example: We often see clients eager to automate workflows with retrieval-augmented generation (RAG) chatbots, only to discover that the strategic fit is missing and operationalisation budgets are hard to secure.
Timeless Principle 2: Classical AI and machine learning still matter
In financial services, the majority of enterprise AI value still comes from classical machine learning (ML) techniques — think credit scoring, fraud detection, churn prevention.
That’s not to say GenAI isn’t an exciting development. It is. But it augments other data-driven technology, it doesn’t replace it.
Classical ML remains highly effective for structured data and scalable decision-making. Large language models (LLMs) excel in unstructured content and human interaction — such as chatbots, knowledge assistants, or document analysis.
The future is compositional, not competitive. The winners will design pipelines where LLMs, ML models, and business logic reinforce each other.
Example: A global bank uses an LLM to extract and classify client data from KYC documents and relationship manager notes. The results feed into compliance and risk models that automate due diligence checks. This powerful combination increases accuracy, consistency, and speed of client onboarding and monitoring.
AI
Start your AI journey with ZenAI
Today’s most successful banks focus on augmenting human workers and driving operational efficiency. That’s why most new generative AI applications are designed for internal use — with employees as the end-users. Providing a secure, internal, UX-focused LLM chatbot is a powerful way to start your AI journey. It allows people to experiment safely, understand the benefits, and apply AI to daily work — from answering policy questions to drafting documents.
DISCOVER ZENAITimeless Principle 3: The proven data science process remains valid
The fundamentals of data science haven’t changed. You still start with a hypothesis, run experiments with measurable KPIs, then iterate and integrate learnings.
What’s new is the human factor. User reactions now influence model behaviour and trust, making evaluation harder.
The Pareto principle holds true: reaching 80% accuracy happens fast — but achieving the final 20% requires 80% of the engineering effort. For banks, this “last mile” matters. Models must integrate into regulated workflows, handle exceptions, and meet audit standards.
Throughout the process, it’s crucial to manage expectations about the complexity of high-performing AI solutions in order to keep teams motivated and engaged.
Example: It took five months of engineering to build a sales and operations chatbot for Uniqua, a major insurance company. That's when 95% accuracy was achieved — the threshold for human-level performance and go-live readiness.
Timeless Principle 4: There’s no production without fundamentals
The hype around AI often distracts from what still matters most: organisational readiness.
For banks, this means data quality — not as a purely technical challenge, but one of governance, ownership, and accountability. You can’t fine-tune your way out of unclear ownership.
Even with a strong data foundation, system integration remains complex in legacy-heavy environments. True AI enablement is as much about organisational transformation as it is about technology.
The advent of GenAI is amplifying these challenges. AI governance must be reimagined as models become more general and outputs harder to control.
At the same time, the use of unstructured data sources — such as internal documents and reports — has exposed new weaknesses. In many institutions, such data is poorly governed, with multiple versions and unclear ownership.
Example: We partnered with BLKB to assess its data maturity, align initiatives with the bank’s corporate strategy, and identify data and AI use cases with the greatest business impact. The resulting framework ensures all initiatives are strategically aligned, enabling data and AI to be effectively used across the organisation.
It’s time to act
Generative AI has changed what’s possible, but not what makes it successful. The bottleneck isn’t technology — it’s execution discipline and leadership focus. For innovation leaders in financial services, the urgency around AI is real. But abandoning time-tested principles is not the answer.
Key takeaways
- Work backwards from value — guide teams in what to build, and what not to build.
- Combine the old and the new — classical ML and LLMs must work together.
- Manage expectations — communicate that reliable AI systems are hard to build.
- Double down on fundamentals — ensure ownership, data quality, and governance are in place.
