Finance's Shadow IT Problem: AI-Built Apps Gone Wrong

When the Department That Says No Becomes the Biggest Security Risk

Sixty percent of the security incidents Pluto Security detects don't come from developers. They come from finance teams building their own tools.

Let that sit for a moment. The function we trust to control spending, verify receipts, and approve vendors is now the biggest source of unreviewed, unmanaged software risk the company has measured across its customer base. Marketing and legal are in the mix too — but finance leads the pack. The actual developers? Fourth place.

I've watched this movie before. Just with different actors.

The Department That Said No Now Ships Code

Finance has always been the control function. The team that said no to the unapproved vendor. No to the expense without a receipt. No to the system nobody vetted.

Now finance builds the systems.

I was talking with a head of tax recently about all the "apps" his team is creating for their work. Not spreadsheets. Apps. Tools with databases, interfaces, logic. "I spend more time developing than calculating taxes," he told me.

Last month Pluto Security showed TechCrunch what happens when that goes sideways: a finance employee at a Fortune 500 company used AI to build an app for tracking customer information and churn. Useful tool. Built fast. Left open to the internet. Pluto's scanning tools found it sitting there, fully exposed.

Every one of those tools turns a finance team into an unsupervised software producer. No procurement process. No security review. No patch management plan. No vendor contract to mark up and negotiate.

Nobody signs anything. Somebody just gets clever on a Tuesday.

We've Seen This Pattern Before

In the 1990s, departmental computing meant finance teams got their own servers. Suddenly you had Oracle databases running under desks, Excel macros that became the source of truth, and Access applications holding customer records. IT didn't know half of them existed until something broke or an audit surfaced them.

We called it "shadow IT" and spent fifteen years trying to wrestle it back into managed infrastructure. We mostly succeeded — by centralizing application development, enforcing change control, and requiring every system to have an owner in the CMDB.

Then we gave everyone AI that writes code.

The difference this time is speed and surface area. A 1990s Access database served ten people in one office. A 2025 AI-generated app can be live on the internet with a login page — or without one — by lunch. It can pull from your CRM, write to your data warehouse, and email your customers. All before anyone in IT knows it exists.

The railroad didn't kill the towns on the day it arrived. The towns just slowly realized that commerce now flowed somewhere else. Finance isn't trying to bypass IT. Finance is solving problems with the tools AI put in front of them, and the old control points aren't in the path anymore.

The $2.6 Billion Spreadsheet Error (And What It Teaches Us Now)

Tax has its own cautionary tale. In 1994, Fidelity's Magellan fund — the largest mutual fund in the world at the time — told shareholders to expect a $4.32-per-share year-end distribution. Big number. Public announcement. Then Fidelity had to retract it.

An accountant had dropped the minus sign on a $1.3 billion capital loss in a spreadsheet, turning it into a $1.3 billion gain. A $2.6 billion swing from a single keystroke. Nobody caught it before it went to shareholders.

Fidelity had controls. They had review processes. They had smart people checking the math. But the error was in a tool the team built themselves, with logic only they understood, reviewed by people who trusted the same mental model.

Now multiply that by every AI-generated app sitting in your environment that nobody outside one team knows exists.

AI just turned the error-prone spreadsheet into a web app with a database backend and an API. Sometimes it even includes the login page.

The Three-Column Test You Should Run Monday

Here's what I'd bring to the next finance leadership meeting — or any functional leadership meeting, honestly.

Make a list. Every employee-built app or tool that touches customer data, financial data, or anything regulated. Every AI assistant with access to your CRM. Every script someone wrote to "automate the boring stuff." Every Zapier workflow. Every internal app someone spun up to "just make this easier."

Next to each one, write three names:

1. Who owns access control? Who decides who gets in, reviews permissions, handles offboarding?

2. Who owns the data? Who's accountable if it leaks, who ensures it's handled according to your data classification policy, who knows where it's stored?

3. Who keeps it running? Who patches it, who fixes it when it breaks, who makes sure it still works after your SSO provider changes its API?

An empty column? You don't have an app. You have an orphan holding your customers' data.

I'm not asking this to shame anyone. The tax lead I talked to is doing impressive work — his team is solving real problems faster than the traditional IT request process would allow. But speed without ownership is just risk with a countdown timer.

Which column does your team leave blank most often?

What Actually Happens Next

The uncomfortable truth: you can't put this genie back in the bottle.

Finance teams — and marketing, and legal, and operations — are going to keep building tools. AI makes it too easy, and the business problems are too urgent. Telling them to stop and wait for IT to build it "the right way" just means they'll stop telling you about it.

So what's the alternative?

I think it starts with treating citizen development the same way we treated BYOD ten years ago. We didn't ban personal phones. We built MDM, we segmented networks, we created acceptable use policies that acknowledged reality instead of wishing it away.

The equivalent here:

• Guardrails, not gates. Give teams approved AI tools that log what they build and enforce basic security hygiene. You'll never see what they build in the free consumer version of ChatGPT.

• Ownership as a deployment requirement. Nothing goes live without names in those three columns. Doesn't matter if it's a ten-line Python script or a full app.

• Periodic discovery. Use the same tools attackers use — scanners like what Pluto runs — to find what's exposed before someone else does.

The other option is pretending it's not happening and waiting for the TechCrunch article with your company's name in it.

The Real Question Nobody Wants to Ask

Here's where I'll leave you, because I don't have a clean answer:

If your finance team can build a working application in an afternoon using AI, and your official IT process takes six months and costs $200K for the same functionality... which one is the actual problem?

We built those IT controls for good reasons. They caught real risks. They prevented real disasters. But if the delta between "approved path" and "Tuesday afternoon workaround" is that wide, we're not controlling risk anymore. We're just deciding whether risk happens in the light or in the dark.

I've watched three waves of shadow IT in my career. Departmental servers in the '90s. Cloud services in the 2010s. Now AI-generated apps. Every time, the initial response was "we need to shut this down." Every time, that failed. What worked was recognizing the need, meeting it halfway, and building the oversight into the flow of work instead of trying to dam the river.

The finance team that says no is now writing code. The question isn't whether that should happen. The question is whether you'll know about it before an outside scanner does.

---

What to do Monday morning: Book thirty minutes with your finance, tax, and legal leads. Ask one question: "Show me what your team built in the last six months that IT doesn't know about." Don't come in looking to shut things down. Come in looking to write names in those three columns. The tools exist. The question is whether they're orphans or assets.

And if you're the one building these tools — good. Solve the problem. Just make sure someone other than you knows it exists before you go on vacation.