AI in Finance: What I've Learned by Trying to Put It to Work

I have a real stake in this conversation because I am trying to put AI to work in finance myself. The more I do, the more I realize how much work remains. Here are a few observations to start a discussion.

1. Support from the company

This is the most straightforward requirement. If a company will not cover token costs, buy servers, or provide GPUs, what are employees supposed to build with? We cannot expect people to rely on a daily WorkBuddy check-in, and paying out of pocket just to do your job is not realistic either. I have actually done that for AI.

For me, the first requirement is real investment—not just slogans. I will not speculate about a company's motives; perhaps it hopes AI will eventually replace you. What matters is that it is willing to let people use AI in practice. I am especially grateful to my company's CFO, who has strongly supported my views on AI and the work I have been doing. Without that support, I would probably just be experimenting on my own.

Most finance professionals trying to apply AI will start inside their own companies. From that perspective, organizational resources are essential.

2. Support from the finance team

This is more complicated. Some colleagues believe AI is the future and are confident they can learn the necessary skills, so they support it and make an effort to learn. Others, unsure how to use AI, choose to ignore it. There can be a subtle fear underneath: if AI does all this work, what will be left for me?

That is why the rollout of even a basic automation tool can reveal a real divide. Imagine a hard-coded workflow—not even an AI agent. Some colleagues are enthusiastic, learn how to use it, and offer feedback. Others watch from the sidelines. Some are so reluctant to use it that even testing it with real data can require bringing in a third colleague. It can be frustrating.

This challenge is not unique to AI. Any new technology can meet resistance when people feel their own interests are at stake. If finance professionals are unwilling to use the tools themselves, adoption will remain a long way off. If you are reading this and are willing to try, do not spend too much time worrying about colleagues who only look for reasons not to use it. Just do it, and keep going.

3. What counts as real AI?

The first two points are abstract, so here is a practical example. To understand AI and agents in finance, I compare them with traditional financial information systems and the integration of business and finance.

Take automated journal entries. A common approach is to map out each scenario, then hard-code it as a workflow. The data for each scenario—bank transactions, workflow records, invoices, and business data—already has explicit mapping rules. This is stable and less prone to error. If a company can achieve that, it may not need AI or agents for that task at all. If you can reliably produce correct books, what more could you ask for?

Without post-training tailored to a specific industry and context—which matters a great deal—AI can hallucinate, and its answers can be inconsistent. In my view, genuine AI implementation in finance has to go beyond the old approach to automation. It means using a large language model post-trained on real financial data, together with a knowledge base of financial standards, laws, and regulations, as well as the company's own documents. That is how I imagine building AI that understands the actual context of the work.

It is a difficult system to build, but I believe a model trained this way could take on real work. It could also be deployed locally, so financial and business data would not leave the company's environment.

So, do you see the distinction? To me, real AI in finance is not just using WorkBuddy to make slides, check accounts, or prepare reports. It means developing AI that can do substantive financial work.

These are a few of my thoughts on applying AI in finance. Not many companies have made it work, and not many people understand what it takes. I keep coming back to two ideas: anything is possible if you take it seriously, and the world is one big makeshift operation.

Put what you know into practice, and trust the process.