AI is probably one of the most talked-about topics in Finance right now.
Over the past few months, we’ve been diving deeper into it ourselves, through our AI Lab, conversations with financial institutions and, most importantly, the people actually using it.
And one thing has become crystal clear: employees are often moving faster with AI than the organisations they work for.
They’re experimenting with ChatGPT, Claude and other tools, while their employers are still figuring out what’s safe, what’s allowed and where to even begin. In fact, a recent IBM study among 1,500 senior finance leaders found that only 6% say their finance function is truly ready to integrate AI at scale.
So, are the concerns around AI valid? How are financial institutions implementing it safely? Where should you start and, perhaps more importantly, where shouldn’t you?
We sat down with Henriette Claus and Annabel Oggel from ADC Consulting, who help banks and insurers including ABN AMRO, ASN Bank, NIBC and Nationale-Nederlanden navigate exactly these questions: how do you implement AI safely, where do you start and what does it look like when it actually works?
“The person working until midnight is exactly the right person to own the tool. How can AI get them home before dinner?”et them home before dinner?”
1. First things first: is AI actually safe?
It’s probably the biggest question holding financial institutions back. But according to Annabel, AI isn't inherently safe or unsafe. It depends on the environment you create around it.
AO: “It depends heavily on the user and the environment you use it in. You can build it in a way that lets you experiment safely. However, if you don’t pay attention, you can suddenly do very harmful things with it.”
The models themselves might come from external vendors, but companies can control what information they have access to, where they run and whether any data leaves their environment. Henriette argues that this requires companies to make a clear distinction between what should be controlled centrally and where employees should have freedom.
HC: “You want to keep the guardrails central. These are the boundaries you operate in, and what specific information you feed into an LLM (Learning Language Model). But you also want to keep the flexibility for teams to experiment and test.”
The same applies to regulation. According to Annabel, financial institutions often have more of the foundations for the EU AI Act in place than they realise because they're already operating under strict security and compliance requirements.
At a Large Dutch Bank, ADC started by mapping existing safeguards against AI regulation. Today, the conversation has already moved to the next challenge: agentic governance. How do you maintain oversight when AI agents themselves start performing work across the organisation?
2. Don't bet your AI strategy on one model.
Once you've created those guardrails, the next question is: what do you actually allow people to use? Many companies have rolled out one approved tool, only to discover employees are quietly using ChatGPT, Claude and others on the side. Henriette thinks that's almost inevitable.
HC: “You can't say today who's going to win, and certainly not in six months. We often advise C-level to not just pick one tool, but to build a harness in which several tools can safely be tested.”
ADC itself gives employees access to multiple models in a protected environment, alongside a monthly AI budget. Platforms such as Localign and Orq.ai allow organisations to do something similar: employees can access different models, while the company keeps control over the environment, usage and costs.
The same principle applies when building AI products.
For a leading Dutch bank, ADC built an AI-powered lending assistant that transforms complex documents into structured credit proposals, reducing a process that took days to hours. Within months, the solution was adopted by over 100 lenders, improving speed, consistency and productivity while keeping human oversight at the centre. (Read the full case study here)
The lesson: build for AI, not for today's favourite AI company.
3. Don't start with AI. Start with the annoying work.
Okay, you have the guardrails. You have the tools. Now where do you actually start?
HC: “Whether you're big or small, we'd always start somewhere very concrete, where you can deliver value fast and show results quickly.”
ADC looks for something much simpler: work that's boring, repetitive and takes a disproportionate amount of time.
At one of their largest clients in banking, lenders spent roughly a week per report reading Word documents, PDFs and Excel sheets supplied by clients. ADC built a tool that gathers and structures that information automatically.
The lender still reviews the output, makes adjustments and signs off. But a week of gathering information has become a matter of hours.
Henriette has a simple way of finding those opportunities:
In other words: don't ask where you can use AI. Ask who's still working at midnight — and what they're doing.
4. Go beyond saving hours: rethink how work gets done
Sometimes, automating the annoying work doesn't just save a few hours. It changes the work completely.
Annabel experienced this herself. Before AI agents, she migrated a bank's risk model from legacy code to Python manually. It took around two months of full-time work.
ADC has since built a framework of AI agents to perform much of that migration automatically, with controls around what those agents can access. For example at financial institutions, the framework is being used to migrate legacy risk models.
The same task that took Annabel two months? Two days.
AO: “I migrated a model by hand before. That took about two months. With the agents, it took two days.”
When you rethink how work gets done, AI stops being a productivity hack and starts becoming transformational.
What does this mean for your team?
Perhaps the biggest takeaway is that becoming AI-ready doesn't mean implementing AI everywhere.
Start small. Find one concrete process where you can deliver value quickly. Talk to the people actually doing the work. Build something. Test it. Measure the result. And don't be afraid to conclude that AI isn't the answer.
For the next generation entering finance, it also means the fundamentals still matter. AI might be able to produce the work, but someone still needs to judge whether that work is any good.
AO: “AI is just a tool. It's the whole point of philosophy: it's about the question, not the answer. AI only gives answers. If your question is bad, the answer won't help you.”
Perhaps that's ultimately what separates the 6%.
They're not necessarily using more AI. They're getting better at deciding how AI can be an enabler for (re)designing the work.
4 AI use cases in Finance, in numbers
A lending team at a Dutch bank started using AI to turn client documents into lending reports. The result? From 1 week per report to hours.
Risk teams at three financial institutions started using AI agents to migrate legacy risk models to modern code. The result? From 2 months to 2 days.
A global insurer launched an AI chatbot that runs in its own cloud environment, with customer data kept in-house. The result? 24/7 customer support, with customer data kept in-house.
A large Dutch bank started by mapping its existing safeguards against AI regulation. The result: a foundation for bank-wide AI and agentic governance.
Curious what AI could do for your organisation?
ADC helps financial institutions turn AI ambition into AI that works - safely. From strategy and governance to training and implementation, the team works alongside clients so knowledge and ownership stay in-house.
