When organizations start exploring AI in finance, there’s often excitement about tools, automation, and bold use cases—cash flow forecasting, auto-generated dashboards, AI-driven reporting. But I haven’t seen a single implementation that could go straight to AI without first fixing the data.

The AI audits I conduct evaluate readiness across four key areas: processes, people, data, and governance. While each plays a role, data readiness is where I most often put on my CFO hat. Because the data I assess isn’t just about format or consistency—it’s deeply financial. I ask questions like: Do you have a proper close process, so pulling data from QuickBooks today won’t yield a different result tomorrow? Or, Is revenue recognized consistently across entities and subsidiaries?

I often work closely with the company’s CFO, and almost always, the data issues we uncover are not new. They’ve usually been on the radar for years. But without urgency, they’ve stayed unsolved.

So, if your organization is considering an AI implementation project, this might be the perfect time to address that inconsistent, clumsy data—the same issues you’ve been meaning to fix for years.

I recently started working with a company that has been around for 15 years. It operates across multiple countries, has opened and closed legal entities, bought and sold business units, and never implemented an ERP.

The company was tired of an inefficient budgeting process and wanted to introduce AI to make it smoother. And no surprise, the budgeting process was a mess. Guess where 15 years of consolidated financial data lived?

In a tab inside an Excel file—with more than fifty other tabs surrounding it, many obsolete or referencing sheets that no longer exist. The funny thing is that everyone knew how to find it.

Before we could even talk about AI, we had to rebuild the foundation. We spent time cleaning the data, aligning definitions, and building consistent mapping across business units. Only then could we move forward with anything AI-related. The fix wasn’t overwhelming—it just required focused attention.

Three Layers of Data Readiness

1. Structure

AI needs structure—reliable, repeatable logic across all inputs.

A recent client was eager to automate expense reporting using an AI tool. But when we reviewed the setup, we discovered that the same vendor might be coded under different names across departments—sometimes with typos, sometimes with completely different categorizations.

The tool couldn’t reconcile them without help. We introduced naming conventions and a master vendor mapping table. It took just a few focused days and we could move forward.

Things to check:

• Are your chart of accounts and cost centers consistent across entities or divisions?
• Do your teams follow the same reporting formats?
• Are naming conventions enforced?

You don’t necessarily need to implement an ERP. Sometimes, a simple mapping exercise is enough. And a small tip for you: AI can actually assist with this cleanup work.

2. Stability

AI models need to be built and trained on stable data.

One company I worked with was pulling data directly from QuickBooks into an AI-powered forecasting tool—but the outputs kept shifting. As it turned out, the accounting team had no month-end lock process. They were retroactively changing prior period data for weeks after the close. AI, of course, treated those changes as new inputs—completely distorting the forecasts.

We had to hit pause, build a proper close process, and implement version control. Only then could we trust the outputs.

And like I said before, the CFO knew the close wasn’t working; it just never hurt enough.

Ask yourself:

• Are past periods truly locked after close?
• Are financial statements subject to post-close adjustments?
• Does your AI pipeline reflect a frozen set of data?

You don’t need every answer today. But you do need stability if you want AI to produce reliable results.

3. Access

Even perfect data isn’t useful if AI tools can’t find or trust it.

In one case, a company was using legacy accounting software that couldn’t be connected to modern AI tools via API. Every time the team needed data for reports or analysis, they had to export CSV files manually, clean them up, and re-upload them into another tool. This not only introduced delays but also created versioning issues and increased the risk of human error.

We eventually recommended creating an interim staging environment—a cloud folder with clean, structured exports from the accounting system—to bridge the gap until they could transition to a more modern, integratable system. It wasn’t perfect, but it allowed the AI implementation to move forward.

Things to consider:

• Do you have a single source of truth for financial reporting?
• Is data stored in a place where AI tools (and your team) can access it?
• Do you know who owns each dataset?

None of this requires heavy IT infrastructure. Sometimes, it’s just a matter of digital housekeeping—and a few well-placed access controls.

There’s a misconception that AI can only work once your data is perfect. That’s not true. What AI needs is clarity, consistency, and access. Most of the foundational fixes—like aligning revenue structures or building a basic month-end close control—can be handled in weeks, not months.

So if AI is on the roadmap at your company, start here: go back to the data. Chances are you’ve been aware of the weak spots in your data—across reporting structures, revenue classifications, version control—for years. Now is the time to address them, while you still have the space to make foundational improvements without pressure.