The CFO AI Tech Stack: What Survived (And What Didn't)

Most Tools Don't Survive the First Quarter

I've tested over 30 tools in the last two years. Most of them were gone by month three.

Not because they were bad. Some were genuinely well-built. They died because the bar moved. Once you have a reliable AI layer running underneath your workflow, every other tool gets judged against a harder question: does this make the AI more useful, or does it just add another place to log in?

Most tools can't pass that test.

Here's what's still running, what I cut, and the framework I now use before adding anything new.

The Year I Broke My Stack

Twelve months ago, I had what looked like a reasonable setup. Make.com automations moving data between systems. ChatGPT helping with drafts and analysis. A mix of project and finance tools that each did one thing well.

Then the automations started breaking. Not catastrophically. Just quietly. A field rename in one system, an API change in another, and suddenly data stopped flowing and nobody noticed for a week. I'd built a network of dependencies that required constant maintenance just to keep it running.

The ChatGPT situation was worse. I was using it for financial analysis: variance summaries, scenario framing, commentary drafts. The outputs looked good. Confident tone, clean structure. Then I caught it inventing a number in a board summary. Not a huge number, but a number I hadn't given it. It had filled a gap in the data with something plausible. In finance, plausible and wrong is worse than obviously wrong.

That's when I realized the problem wasn't the individual tools. It was that I had built around the wrong center. I needed a reliable AI layer first. Then everything else.

How Claude Became the Center

I switched to Claude in mid-2025. The context window was the initial draw: I could feed it a full set of financials, a prior-year comparison, and commentary guidance in a single session and get coherent output. ChatGPT kept losing the thread halfway through.

The real shift happened when I started using Claude Code. At that point, it stopped being a chatbot and became infrastructure.

I now have Claude handling variance analysis drafts, board commentary, scenario output summaries, and internal memo first-drafts. It writes the Python scripts that pull and format data. It reviews model outputs for logical errors before I touch them. The sessions run in parallel, so while I'm reviewing the output from one, another is building.

The time savings are real. Board prep that used to take half a day now takes two hours, and more of that two hours is actual thinking rather than assembly. Scenario modeling that required manual formula updates across a model now runs in minutes through scripted recalculation.

The reason Claude survived while everything else got cut: it reduces friction instead of adding it. It doesn't require its own workflow. It fits into whatever I'm already doing.

The Survivors

These are the tools still running, and the reason each one made it.

Claude. The center of everything now. Not optional. If you're a CFO who hasn't moved past using AI as a search engine, the gap between what you're getting and what's possible is significant.

Excel. I know. Everyone keeps predicting its death. Excel survived because AI made it better, not obsolete. I use Claude to write formulas I'd normally look up, scaffold three-statement models, and generate sensitivity analysis structures. Excel is still the best environment for financial models when you need institutional flexibility and client-compatible output. AI made it faster, not unnecessary.

Notion. Survived because AI can read it and write to it. I use Notion for client documentation, project notes, and process documentation. Claude can pull context from a Notion export and write back into templates. Tools that work with AI workflows instead of around them earn their keep.

Slack. Communication isn't going away. Slack stays because the team is there and because I can pipe AI-generated summaries into channels without anyone needing a new tool. Low friction, high value.

Google Sheets. Not instead of Excel, alongside it. Sheets handles anything that needs live collaboration or easy client sharing. It also connects cleanly to data pipelines in ways Excel sometimes doesn't.

Loom. I record walkthroughs of models and analysis for clients and team members. Five minutes of Loom replaces a 30-minute meeting and lets people watch at their own pace. Kept it.

1Password. Non-negotiable. Every tool, every API key, every shared credential lives here.

Linear. Project tracking for the builds I'm running. It's fast and doesn't require a training course to start using. One of the few project tools I've kept because it stays out of the way.

The pattern across all of these: they work with AI, not around it. They either plug into the AI layer or they stay out of its way. Either is fine. What isn't fine is a tool that demands its own separate workflow and can't connect to anything else.

The Casualties

ChatGPT went first. The context window was too short for serious financial analysis, and the hallucination rate on numerical tasks was too high for work where the numbers matter.

Tableau surprised me. Beautiful tool. Genuinely impressive. Also completely incompatible with the pace of AI-driven iteration. When I'm running ten scenario variations in a session, I need dashboards that update in real time from a data layer. Tableau's build cycle doesn't fit that workflow. I replaced it with custom dashboards built in Next.js, which sounds like more work but took less time than I expected once AI was doing the scaffolding.

Asana was too heavy. Every workflow required too much manual setup to maintain. Projects got updated for a few weeks, then fell behind, then became inaccurate, then stopped getting opened. For solo or small-team finance work, it's designed for teams much larger than mine. Monday.com had the same problem with a worse interface. Cut them both.

Zapier I held onto longer than I should have. Switched to Make.com first, then cut most automations entirely. The maintenance cost of multi-step automation chains is underestimated by almost everyone who builds them. Every time a source system changes, the chain breaks. I now prefer scripted workflows that Claude helps maintain. More durable, easier to debug, no monthly bill.

Salesforce, Confluence, Evernote: all gone. Salesforce I replaced with a lightweight CRM with an API I can actually query without a consultant. Confluence documentation dies in Confluence because nobody goes back to it. Evernote I held on to out of habit. Notion and Claude sessions cover everything it did, faster.

Shadow IT Is the Real Risk CFOs Miss

Here's something I don't see talked about enough in finance circles.

When you don't give your team a clear, approved AI stack, they don't stop using AI. They find their own. Free ChatGPT accounts, Gemini, random browser extensions. They upload spreadsheets with client data. They paste revenue figures into a free tier model with no enterprise data agreement.

That's not a hypothetical. I've seen it happen inside a team of four people.

The data governance issue is real. Sensitive financial data in unvetted models is a liability you can't fully see until something goes wrong. Regulators and clients don't care that your analyst "just used it for drafting." The data left your environment.

The fix isn't a policy. Policies don't stop behavior; friction does. Make the approved tools easy to use and easy to access. If Claude with an enterprise agreement is three clicks away, most people will use it instead of the free alternative. If it requires a help desk ticket, they won't.

Your AI governance strategy is your tech stack. They're the same decision.

The Evaluation Framework

Before I cut the stack down, I had five questions I ran every tool through. Does it solve a real, recurring problem? Will people actually use it past week two? Can it talk to the other tools in the stack? Is the switching cost worth it if something better comes along? Does someone have to babysit it to keep it working?

Those still apply. But since AI became the center, I've added three more.

The first: does it have an API or AI-native integration? Tools that can't be scripted or queried are becoming second-class citizens. If Claude can't read from it or write to it, it lives in a separate workflow I have to manage manually. That's a tax I now refuse to pay.

The second: does it reduce AI friction or add to it? A lot of tools claim to be AI-powered because they bolted a chatbot onto a dashboard nobody liked. That's noise. The real question is whether the tool makes my AI-assisted work faster. Most fail this.

The third, and the uncomfortable one: if AI can do 70% of this tool's job in 12 months, is the switching cost worth locking in now? Software categories are getting hollowed out. Structured templates, automated reports, workflow managers. Before I commit to a three-year contract, I ask whether I'll still need it. Sometimes yes. Often not.

The reason most tools fail this framework isn't that they're poorly built. It's that the problem they solve isn't painful enough to actually change behavior. A tool that saves 20 minutes a month won't stick if it requires 20 minutes of maintenance every week. The math doesn't work.

The Actual Stack Audit

If you want to run this exercise on your own setup, start with the AI layer, not the tools.

Pick one AI assistant. Learn it well enough to use it for the recurring, high-friction work in your workflow. Board prep. Variance summaries. Scenario outputs. Spend four weeks running those through AI before you touch anything else.

Then go through your current tools one by one. For each one, ask: does this help my AI workflow or fight it? Does it store information AI can access, or does it silo it? Does it speed up the work AI started, or does it add a separate process?

Cut anything that fails. That's it.

You'll end up with a shorter list than you expected. The tools that survive are better for the cuts around them.

P.S. If you asked me what I'd keep if I had to start over tomorrow with three tools: Claude, Excel, Slack. Everything else is negotiable.