I’ve been working in the finance space for over 10 years now. First doing actual finance work, then building software for finance teams on the engineering side. In the past six months, I’ve seen more transformation in how finance people do their work than in the previous 10 years of my career.
The biggest change agent has been, well, agents themselves: AI systems packaged with tools they can call and code they can write on the fly. The biggest near-term gains in finance will not come from replacing every system with an AI platform. They will come from giving people local agents that can safely turn repetitive analysis into repeatable workflows, while keeping the work close to the data, credentials, and judgment already in place.
Here are a few AI trends I’ve been noticing in recent months.
Excel as an Artifact, Not the Workspace
“AI doesn’t do it end-to-end. It does it middle-to-middle. The new bottlenecks are prompting and verifying.” — Balaji Srinivasan
For decades the number one tool used by every finance professional has been Excel. It’s a required skill on every finance job application.
Excel was a finance person’s external brain. A place to think through problems, analyze data, and present answers to tough questions. Creating a financial model was almost a piece of art. Each one shared industry-standard formatting, such as color coding and tab layouts, but retained the unique fingerprint of its creator. Much like painters using the same watercolors but developing their own distinct styles.
AI is starting to change this. Increasingly, the first pass of thinking, analysis, and drafting happens outside Excel. Excel does not disappear, but it is becoming the review surface. The familiar artifact where assumptions, outputs, and exceptions are checked by the people accountable for the decision.
That is where coding agents come into play.
Coding Agents Are Becoming the Finance Interface
When a finance person starts their day, they may now power up GitHub Copilot or another user-friendly agent like Microsoft Scout before opening Excel. These are becoming some of the sharpest tools in their arsenal.
Why? At their core, they are coding agents. Code gives them a way to work across data sources, analysis tools, and output formats. The AI writes the code, with the human steering it in the right direction from time to time.
A coding agent can take on much of the work a human used to do manually inside Excel: - Pull data - Transform data for analysis - Analyze the data - Flag unusual movements and test plausible drivers - Iterate the analysis until a final answer is found - Prepare outputs others can quickly review
Think about a monthly variance-analysis task that once required opening five reports, reconciling account names, calculating drivers, and writing commentary. An agent can turn much of that routine into a repeatable workflow. The analyst still owns the assumptions and the conclusion. The agent removes the mechanical middle.
After an agent performs a workflow correctly, a human can document it as a repeatable workflow, perhaps with a skill or script. If the agent doesn’t have the right data, it can connect to APIs or write one-off scripts to retrieve it from approved sources.
With all of this capability, manually rebuilding the same analysis in Excel starts to feel increasingly hard to justify.
Local Still Wins Over Cloud
One of the best parts of these coding agents is that they can execute code locally on your own machine. Instead of immediately running in the cloud, the work happens on your computer, much as it would if you were doing it yourself. At first this may not sound ideal. Shouldn’t we run these things in the cloud at scale?
The answer is both yes and no. Parts of the work should be done at scale in the cloud, but much exploratory work still benefits from local execution. The biggest compliance obstacle in finance is data security: making sure you can only see and pull data you have access to. It makes sense that a coding agent doing work you assigned should be able to pull and analyze only the data you can access.
For exploratory work, local execution can be the easier path because the agent can operate through the same authenticated tools and permissions the employee already uses. Cloud deployments can be secured, but they introduce a separate identity, authorization, audit, and operational design problem before they are ready for broad use.
AI agents can have their own separate identity, where they are treated like a new coworker. But that gets complicated when different people use the same agent for work requiring different levels of data access. Designing the blended access model between the systems an agent can access overall and what it can access for each employee remains difficult. For now, local execution often wins.
Finance Teams Are Building Software by Accident
Anyone can now build something using AI coding agents. Writing a first version is easier than ever, but plugging together the Lego pieces in a software system is not solved by AI. This still requires humans to set up cloud infrastructure, follow proper security practices when storing API keys, authenticate between systems, and handle everything else required to make the code work.
It is exciting to build a local web app or AI-powered tool that works on your laptop, but sharing it with others can become a slippery slope. You might build the prototype in three hours rather than waiting three weeks for an engineering team. But scaling it in the cloud and allowing others to access it securely is where a DIY finance project can stall.
Corporate systems are more cautious about cyberattacks than ever. That means there are countless restrictions around how software can be built and deployed in the cloud. Code that connects easily to systems on your laptop can appear broken when you try to run it in the cloud. That’s a feature, not a bug.
A local prototype can be valuable precisely because it’s small. Trouble starts when a useful personal tool quietly becomes a shared production system. The moment other people depend on it, someone must own access controls, secrets, deployment, monitoring, support, and change requests. That is not a failure of the prototype. It’s the point where the work has become software engineering.
If you built a tool the entire team wants to use, the time you saved by using the tool can turn into an unpaid on-call rotation. You are now the only person on your finance team who can fix it when it breaks. If your boss wants it to work differently, it is your job to add the feature. And when you leave for another job, or more commonly get re-orged to another team, the tool can die on the vine because only you know how it works.
The reason you built the damn thing in the first place was to help you and the team save time. If all that saved time becomes maintenance, you are doing work an engineering team would have been happy to own.
Heed my warning and keep things simple in this new AI era. Just because you can build a web app to showcase your company’s financials with live commentary does not mean you want to build one. Everything has tradeoffs. If you want to build something, turn it into a reusable AI skill that teammates can run locally on their own machines. The skill can contain scripts and ways to connect to corporate systems, while keeping the work close to the permissions people already use instead of quietly creating a shared production app. Save the hosted web apps and APIs for your engineering counterparts.
Final Thoughts
The AI landscape seems to change every day, so this post may become obsolete within days of publishing. We’ll see. But for now, the near-term opportunity is not to rebuild the finance stack with AI. It is to remove the repetitive middle from the work people already do, while keeping the workflow close to the data, controls, and judgment that make finance trustworthy.
Start local. Make the workflow repeatable. But the moment other people depend on it, treat it as software, with all the ownership that implies.