Start With the Problem, Not the Tool
When a finance team asks whether AI can take over part of its analysis, Miles Ericson says that is almost always the wrong starting question. In his experience, the hype around AI leads people to look for ways to adopt it before identifying a problem that needs solving. He puts it simply: "if it ain’t broke, don’t fix it."
Instead, Ericson recommends starting with a specific pain point and then determining whether traditional automation can address it. If the analysis already follows a defined formula using a structured dataset, he suggests automating that process directly. "Adding AI can introduce uncertainty into something that should produce a consistent, repeatable result," he said.
AI shows more potential when the analysis depends heavily on human judgment, where it can improve efficiency through faster research and the ability to work through more information. But that still leaves the person responsible for interpreting the findings and deciding what matters. And it only works, Ericson notes, if the person actually wants to use the technology and understands its limitations. "They need to know what to verify, when to question an answer, and when to rely on their own expertise," he said. "Their willingness and understanding should be part of the decision from the beginning."
What the AI-for-Finance Pages Leave Out
Ericson believes what often gets left out of discussions about AI in finance is how much AI safety depends on implementation. The internet tends to sensationalize failures and stories about AI "escaping" or "hacking" something without giving enough attention to how the system was configured, what permissions it had, or what safeguards were missing.
"A preventable mistake in the setup can create an outcome that gets presented as the technology acting beyond anyone’s control," he said. "But recognizing that a failure was preventable does not make AI inherently safe. You still have to anticipate what could go wrong and build appropriate protections around it."
Whenever he considers whether AI is the right solution, Ericson wants to understand the consequences if it fails. An incorrect internal research summary and an unauthorized financial transaction require very different levels of protection. He asks: What can the system access? What actions require human approval? Can mistakes be detected and reversed?
"Even when you cannot eliminate the possibility of failure, you can often limit its impact," he said. "Finding the right balance between usefulness and safeguards is the difficult part, and I think that deserves more attention than either the hype or the horror stories."
Decisions Finance Teams Make Too Quickly
As he touched on earlier, Ericson says the first mistake is often deciding to use AI at all before establishing that it is needed. If there is a clear reason to use it, the next mistake is choosing a model or provider before deciding exactly what the system should do, how it should work, and why it would improve the existing process. Those requirements should guide the choice of technology.
He also thinks media coverage puts too much emphasis on whether ChatGPT, Claude, or another tool is currently "the best." For many everyday uses, their capabilities overlap considerably. Quality advantages can shift back and forth, and certain specialized tasks may reveal larger differences, but there is no universal winner. "What matters is how well a tool performs the work you actually need it to do," he said.
The complexity of the task should also guide how much you spend. For simpler tasks, a cheaper model can be the more sensible choice, including a local model when the hardware and operating costs make sense. Paying for a more capable model is only worthwhile if those additional capabilities meaningfully improve the outcome.
Choosing a provider too early can introduce unnecessary constraints. You can end up designing the entire process around that provider’s features, making it harder to switch later even when another option would suit you better.
Where practical, Ericson favors a more open, portable setup where the provider supplies the AI capability while the core workflow and data remain under your control. That gives you room to change models or providers as your needs evolve. "The specific model may matter today, but the ability to adapt is what gives the overall system a better chance of standing the test of time," he said.
For finance teams considering AI, Ericson's advice is consistent: define the problem first, understand the risks, and let requirements guide the technology choice. More of his work can be found at milesericson.com.