Why You Can't Vibe Code the Office of Finance: A Build-and-Buy Framework for CFOs

Agentic coding tools have made it possible for a single developer to build working prototypes in days instead of months. That capability has reached the desks of CFOs, controllers, and VP Finance leaders already under pressure to modernize aging enterprise resource planning (ERP) systems, cut technology spend, and show a return on every AI dollar.

A new pattern is emerging with IT and Finance leaders getting on joint calls together. On those calls, leaders are asking the same thing: If AI can write the code, can Finance just build the platform itself and skip the purchase?

The answer is no. Writing code that produces a number is easy. Producing a number that survives an audit, a regulator, and a board question isn’t the same task, nor is it an easy one. And AI alone can’t close that trust gap.

A Finance application performs a calculation. A Finance system proves the calculation is correct, shows who touched it, and holds up when someone outside the company asks how it was produced. Most vibe-coded tools stop at the first definition. In short, that’s the entire problem.

Here’s where that problem shows up in practice.

Consider the scale most Finance organizations are actually working with. An accountant closing the books might upload 800 subledger records into a system. Now consider what that means at an organization with hundreds or thousands of entities. There, that single upload can explode into tens of millions of possible combinations of cells after accounting for the following:

• Chart of accounts
• Alternate planning hierarchies
• Legal entity hierarchies
• Intercompany relationships

Additionally, Finance data changes constantly. That means a homegrown tool needs to know not just what the latest actuals and forecast are, but also whether they are stale and why.

Intercompany eliminations

An AI-built tool can net two entities' intercompany balances in an afternoon. What it won’t do reliably is apply the correct elimination logic in the following scenarios:

• When the entities use different functional currencies
• When one side books a transaction a period late
• When a shared-services entity bills three subsidiaries for the same service

If you get the logic wrong, your consolidated revenue is overstated. Worse, nobody notices until an external audit asks for the elimination schedule.

Entity hierarchy changes after an acquisition

With every merger and acquisition (M&A) event, the ownership percentages, reporting currencies, and consolidation method for some set of entities all change. A homegrown tool built for last year's org chart will keep applying last year's rules unless someone remembers to update it. Controllers who have lived through a close in the quarter after a deal has closed know that this scenario is where restatements come from.

Forecast assumptions with no lineage

An AI model can generate a forecast in seconds. But what happens if nobody can say which assumptions drove which line item? Or why the model changed a number from last month? Ultimately, Finance leadership cannot defend that forecast to the board, nor can Investor Relations defend the forecast to the market.

Security on sensitive data

Not everyone should see every number. For instance, regional and entity controllers need access to exactly the data they are authorized for, nothing more. This type of cell-level security is rarely on an IT team's radar until the first access review.

None of these failures come from AI being inaccurate. Instead, they come from AI operating without the accounting rules, entity structure, and history that a real Finance system is built to carry.

...it’s that no one can explain how the number was produced.

A wrong number gets caught and corrected. An unexplainable number gets discovered by an auditor, a regulator, or a short seller. And by then, the damage is reputational, not just numerical.

Auditors and regulators expect four things, and most vibe-coded systems provide, at most, one of those things:

• Evidence: The source data behind every figure, not just the figure itself.
• Lineage: A traceable path from source transaction to reported number, across every roll-up and adjustment.
• Approvals: A record of who reviewed and signed off at each control point, not just who typed a change.
• Regulatory defensibility: The ability to reconstruct, months later, exactly how a filed number was calculated under GAAP or IFRS as the number existed at that time.

A script that changes a value in a spreadsheet satisfies none of these things. However, a platform built for controllership is designed around all four from the start.

The framing above still leaves out the option most CFOs actually want: using both. A governed platform doesn’t have to compete with an IT team's own automation efforts. Instead, the platform can be the trusted foundation underneath them.

A Finance agentic layer, for example, exposes a platform's APIs, financial integrity, and audit trails as building blocks for IT teams. IT can then assemble those building blocks into existing workflows in Databricks, Snowflake, or Azure. IT can also work with the building blocks through whichever AI tool Finance already uses (e.g., Claude, ChatGPT, Gemini, Excel, or Power BI).

The point isn’t to stop IT from building. Instead, it’s to ensure whatever they build sits on top of numbers that are already correct, secured, and auditable. In the same way, most companies don’t build their own customer relationship management (CRM) platform. Companies instead build on top of a specialist platform. Why? Because that platform has already solved for the domain. Finance platforms play the same role for the Office of Finance.

Not every Finance capability carries the same risk. The mistake is treating AI in Finance as a single decision. Why? Because that single decision is actually several decisions, each belonging in a different category.

Build with AI, on top of a governed platform:

• Reporting and analytics analysts that answer ad-hoc questions against already-audited data
• Variance commentary drafting that writes a first-pass explanation of budget-to-actual swings, which a financial planning and analysis (FP&A) analyst edits before it goes in the board deck
• Scenario and what-if modeling assistants that pressure-test a planning assumption in natural language

Keep on a trusted platform, not homegrown:

• Consolidation and intercompany elimination logic
• Close processes and the controls embedded in them
• Entity hierarchy and ownership structure management
• Audit trail, evidence retention, and approval workflows
• Statutory and regulatory reporting

The pattern is simple. AI belongs wherever a human reviews the output before it matters. AI does not belong wherever the output becomes the official record without a controlled, auditable path to get there. Ultimately, CFOs should be able to answer a simple question: "Who’s accountable for this number if it's wrong?" If they can’t answer with a name and a control, the capability belongs on the platform side, not the build side.

The old model had IT build systems and Finance use them. Today, that division no longer holds. AI governance and master data integrity are now joint accountabilities. Treating them as separate workstreams is how organizations end up with the following:

• An IT-built tool that Finance cannot defend to auditors
• A Finance-built spreadsheet that IT never security-reviewed

The practical fix is structural. First, put a technical owner and a Finance process owner on the same small team. Then give them joint sign-off before any AI capability touches a number that leaves the building. Neither role reviews the other's work after the fact. Instead, they build it together. The result? The accounting logic and the technical implementation are correct at the same time, not reconciled after something breaks.

Stop asking whether AI can build a Finance capability. It almost always can. Instead, ask whether that capability produces a number your company will report to a regulator, disclose to a board, or defend to an auditor. If the capability does, put it on a governed platform and build the AI layer on top of the platform. If it doesn’t, build the capability with AI and ship it. Then let AI move as fast as it wants. Ultimately, that’s the line CFOs need to draw and hold.

CFOs who blur that line will eventually explain a restatement to their board. CFOs who hold the line won’t. Why? They’ll get the efficiency gains from AI without gambling the one thing Finance cannot rebuild quickly: the organization's trust in its own numbers.

For more insight into how to approach the future of Finance, read our AI operating model: Forward Finance.