Applied AI in finance: 4 best practices to get started
AI in finance series
Over the past month, we’ve provided a number of tips on how finance teams can drive real results with AI. But if your AI efforts are just getting off the ground, you may have more fundamental questions: how do I write effective prompts? Which LLM tool is the best fit for my business?
This edition of our Applied AI in finance series answers those common questions and more to help you hit the ground running.
1. Structure prompts for high-quality outputs
While AI chatbots are quickly stealing market share from search engines, they don’t work the same way. The output of chatbots depends heavily on the quality of your prompts, and these usually require a bit more thought than a simple Google search (but know the result will make it worth the effort).
Those who query AI like a search engine may be left wanting and ultimately move on. Nicolas Boucher, a leading AI in finance expert who runs AI Finance Club, provided a handy framework for writing prompts called CSI + FBI that will improve the quality of results. CSI stands for context, specific, and instruction, while FBI refers to format, blueprint, and identity.
To create a ready-to-send dunning letter, Boucher wrote a prompt explaining:
- He’s an accountant (context).
- A client invoice is two months overdue (specific).
- He needs to draft communication to the client (instruction).
He expanded on this by explaining the output should be:
- A formal letter (format).
- Use a serious tone and threatening language such as “legal actions” (blueprint).
- Write from the perspective of a lawyer (identity).
2. Identify the best LLM for your business
Rankings of the best LLMs are constantly changing as leaders jostle for top position by constantly improving their capabilities. This can leave decision-makers unsure of which provider they should choose. But finance teams just starting to test the AI waters do not need to spend time understanding the technical differences between models. The capabilities of the leading players—ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), and Copilot (Microsoft)—are generally not distinct enough to affect most finance users, Boucher noted.
“Don't choose the one you think is the most fun or where somebody tells you, ‘Oh, this is the best,’” Boucher said. “Because those four, they are almost always providing the same output.”
What matters more is what will work best within your work environment. For example, if your business already uses Microsoft Outlook and Microsoft Office, Copilot will have meaningful advantages because it can directly access your data and communications. If you’re planning a project, for instance, it can automatically scan recent meeting invites to propose a list of people to bring in.
3. Choose the best model for the job
In yet another sign of just how fast this technology is progressing, leading LLM providers regularly release new models. ChatGPT enterprise subscribers, for example, now have access to eight models, each with unique strengths designed to help with different use cases. It can feel overwhelming, and you might not have the time to test several models and compare the results.
With ChatGPT, Boucher recommended focusing on just two models: GPT-4o and o3.
- GPT-4o is the best option for the vast majority of tasks—95% of the time, Boucher estimated—including quick questions, translations, or gut checks on rough drafts. It should be your default.
- o3 is excellent at solving more complex problems and can provide more comprehensive solutions. The output may be 10 to 20 times better, according to Boucher, but it’s much slower. If 4o isn’t providing the recommendations or level of detail you need, try o3.
4. Realize the power of custom GPTs
Once you’ve selected an LLM and model, you’re well on the way to adding real value with AI. But if you want to go a step further, try custom GPTs. These GPTs can be quickly trained to better execute a specific task based on training material, written “instructions,” and real-time feedback.
For finance leaders, a valuable use case is a GPT trained on their voice and style. The CFO, for example, may spend an hour training the chatbot to sound like them. The finance chief can then share this custom chatbot with team members who typically help with writing tasks, such as narrative reports.
AI can also help with presentation prep—Boucher showed that the chatbot can suggest how to structure a slide and what information to include based on that leader’s preferences. The time savings add up: AI-supported drafts should require less refining from the CFO than they would in the past, and it likely took the team member less time to put together the drafts.
Looking for more ways to use AI?
Check out the rest of our Applied AI in finance series for more actionable tips on how finance and accounting teams can find real value in AI.
In the meantime, watch the full webinar with Boucher and join the 1 million+ people following him on LinkedIn.
