Shadow AI Spend: The Off-Books Bill CFOs Can't See

Quick question for every CFO and Controller reading this. How much did your company spend on AI last month?

You probably have a number in your head. It's likely your enterprise Copilot or ChatGPT licence, plus whatever engineering burned through on API credits. Here's the uncomfortable part: that number is almost certainly too low.

The gap between what you think you spend on AI and what you actually spend has a name. It's shadow AI spend. Right now it's hiding in expense reports and inside "software – misc" GL lines. And it's growing faster than almost anything else in your budget.

What is shadow AI spend?

Shadow AI spend is any money your organisation pays for AI tools, subscriptions, tokens or API usage outside your formal procurement and approval process.

A few everyday examples:

• A marketing manager expenses a ChatGPT Plus subscription every month
• A sales team puts an AI meeting notetaker on a corporate card
• An engineer runs model API calls on a personal account and claims the bill back
• A designer pays for an image generator and codes it to "subscriptions"

Put simply, it's AI that finance pays for but never approved.

Shadow AI vs shadow AI spend

You'll hear "shadow AI" a lot, mostly from security teams. They use it to describe employees using AI tools the company hasn't sanctioned, whether those tools are free or paid. That's a data problem, and a real one.

Shadow AI spend is the paid slice of that problem, the part that actually hits your books. Free-tier use carries data risk. Paid use carries data risk and a financial commitment nobody signed off on. This post looks at shadow AI from the finance side: where the money goes and how to bring it back into the light without killing the productivity your teams get from these tools.

How big is the AI spend problem? The 2026 numbers

The data from the past year is pretty striking.

• Expense-based SaaS spend jumped 267% year over year, and ChatGPT is now the single most expensed app. That's from Zylo's 2026 SaaS Management Index, which analysed more than 40 million SaaS licences and $75 billion in spend. Zylo calls this out directly as AI entering organisations outside formal procurement.
• Spend on AI-native apps rose 108% overall, and 393% at companies with more than 10,000 employees (Zylo, 2026).
• Eight of the top 50 most expensed apps are now AI-native, and ChatGPT climbed from #2 to #1 (Zylo, top expensed apps).
• 55.9% of US businesses were paying for AI in September 2026. Among medium-sized businesses the figure was 62.5%, according to transaction data in the Ramp AI Index.
• Only 40% of companies have bought official LLM subscriptions, yet employees at over 90% of companies regularly use personal AI tools for work. MIT's Project NANDA called this the "shadow AI economy" in its State of AI in Business 2025 report.
• 78% of AI users bring their own AI tools to work, according to the Microsoft and LinkedIn Work Trend Index. The same research found 52% are reluctant to admit using AI on their most important tasks (The Decoder).

And the total pie keeps growing. Gartner now forecasts worldwide AI spending of $2.7 trillion in 2026, up 49.5% on 2025. Menlo Ventures estimates enterprise generative AI spend hit $37 billion in 2025, more than triple the $11.5 billion of the year before.

When a category grows that quickly, the controls almost never keep up.

Why AI spend slips past procurement

Your procurement process was designed for a world where software was bought by IT, on annual contracts, after a vendor review. AI tools don't arrive that way. Here's why.

1. The price sits below every approval threshold

A $20 or $30 monthly seat fits neatly on a corporate card. It never triggers a PO and lands in an expense report weeks after the decision was made. Multiply that by fifty employees and you've got a five-figure annual commitment that nobody actually approved.

2. AI vendors sell to people, then to companies

AI tools are built to be adopted by individuals first. Menlo Ventures found that 27% of AI application spend comes through product-led growth, close to four times the 7% rate for traditional software. In other words, a big chunk of AI spend starts with someone entering their card details on a website.

3. Usage-based pricing makes the bill unpredictable

Seats are easy to forecast. Tokens aren't. When teams pay for API usage, spend can swing wildly from month to month. The Ramp AI Index shows how wide the range already is: among firms paying for AI, the median company spent $15.70 per employee per month in September 2026, while the top 1% spent a median of $9,567.53. Zylo's own 2026 research frames this as a new era of SaaS cost volatility driven by AI.

4. AI is creeping into the tools you already pay for

Even your sanctioned software is changing under you. AI add-ons show up at renewal, and a tool you approved two years ago might now send your data to a model provider. Zylo's take is that nearly every app is turning into an AI app. That means shadow AI spend can hide inside an approved vendor, too.

What shadow AI spend actually costs you

The subscription fees themselves are rarely the headline. The real cost is everything that comes with them.

Duplicate tools and seats. When three departments independently pick three different AI writing assistants, you're paying for overlap, and none of them is on a volume agreement.

Lost negotiating leverage. Forty individual monthly subscriptions give you zero bargaining power. One enterprise agreement does. Shadow spend splits your buying power into pieces too small to negotiate.

Forecasting misses. Variable token bills and surprise renewals make your software budget harder to plan, and the variance usually shows up after month-end close.

Risk that turns into money. This is where the security conversation becomes a finance conversation. IBM's 2026 Cost of a Data Breach Report found that shadow AI was involved in 43% of security incidents, more than double the 20% reported a year earlier (Baker Donelson summary). Breaches involving shadow AI averaged $5.39 million, up from $4.63 million, and 68% of breached organisations had no AI governance in place to manage AI or detect shadow AI (Sentra analysis).

Terms nobody read. Consumer and individual tiers come with their own terms on how your data is used and when the subscription renews. If nobody from finance or legal saw the agreement, nobody checked whether those terms fit your policies.

Gartner expects this to get worse before it gets better. It predicts that by 2030, more than 40% of organisations will suffer security and compliance incidents linked to unauthorised AI tools.

So should you just ban AI tools?

It's tempting, right? Lock it all down and reject every AI expense claim.

Here's the problem. Your people are using these tools because they work. MIT's research found that this shadow AI often delivers better ROI than formal AI initiatives, and that employees who get used to flexible personal tools grow less tolerant of rigid enterprise ones (Fortune).

Bans also have a track record. Samsung temporarily banned generative AI tools in 2023 after engineers pasted confidential chip designs into ChatGPT (New Straits Times). The risk was real, but a blanket ban mostly teaches people to hide their usage better. Remember that 52% of AI users already don't want to admit they're using it.

A ban moves the spend somewhere you can see even less of it, like personal cards or vaguely worded expense claims. Finance ends up with less visibility than before.

The better goal? Make the approved path the easiest path.

How to govern AI spend without blocking your teams

Here's a practical playbook you can start this quarter.

Step 1: Find what you're already paying for

Pull at least 12 months of corporate card and expense data, plus your AP history. Search vendor names and descriptors for AI providers and tools, and look closely at small recurring charges under $100. They're the classic signature of shadow AI spend. Gartner's advice to CIOs points the same way: run regular audits for shadow AI activity and build GenAI risk into your SaaS reviews.

Step 2: Sort every tool into a bucket

For each AI tool you find, decide whether to consolidate it into an enterprise agreement, approve it as-is, replace it with a sanctioned alternative or retire it. The goal here is a clean, known list of AI vendors.

Step 3: Give people an easy front door

Most shadow spend happens because asking is harder than expensing. Create one intake path for AI requests that works where people already are, whether that's email, Slack, Teams or a simple form. Publish a short catalogue of pre-approved AI tools so common requests get a fast yes.

Step 4: Route approvals by risk as well as amount

A $20 tool that touches customer data deserves more scrutiny than a $2,000 tool that only reads public web pages. Your approval rules should weigh what data the tool can access and what the vendor's terms say, alongside the price.

Step 5: Bring contracts and terms into the loop

Before an AI tool becomes a recurring line item, check the basics: how the vendor handles your data, whether your inputs train their models, renewal terms and cancellation rights. Turn the answers into rules your approvals actually enforce.

Step 6: Watch usage-based spend continuously

Set budgets for API and token spend by team, and flag spikes as they happen. A monthly review is too slow when a single runaway workflow can double a bill in days.

Step 7: Review the list every quarter

AI tools come and go fast. Retire what nobody uses and fold the winners into proper agreements.

Where Blackbee AI fits

Here's the thing about shadow AI spend. By the time it shows up as a charge, the decision was made weeks ago in a browser tab. That's a spend intent problem, and it's exactly why we built Blackbee AI to start at intent rather than at the invoice.

It's also part of a much bigger pattern. Somewhere between 40% and 60% of enterprise spend starts outside formal financial workflows. AI tools are just the fastest-growing example.

Blackbee AI is an agentic Intake-to-Pay platform that sits above your ERP as the financial decision and control layer. For AI spend, that looks like this:

• Intake Agent captures spend requests at their source, whether someone asks for a new AI tool in emails, chats, Teams or anywhere else, so the request is tracked before anyone commits money.
• Route Agent routes approvals based on risk and policy, so a low-risk tool gets a quick yes and a tool that touches sensitive data gets the right reviewers.
• Clause Agent reads vendor contracts and turns the terms into active financial guardrails.
• Trust Agent handles vendor onboarding and continuous risk scoring for every new AI vendor.
• Signal Agent runs anomaly detection and spend analytics across your spend data, so unusual patterns like a sudden jump in API costs get flagged for review.
• Sync Agent posts validated decisions back to your ERP, whether that's NetSuite, Sage Intacct, Dynamics 365, Workday or SAP.

And you stay in charge the whole way through. You decide which steps Blackbee AI handles, and it handles only those, with every decision showing its reasoning. If you'd like to go deeper on the analytics side, our guide to spend intelligence walks through how anomaly detection and forecasting work, and our post on AI vendor risk scoring covers continuous vendor monitoring.

Ready to see every AI dollar before it's spent? Blackbee AI captures spend intent at the source and governs it all the way to payment. Book a demo and we'll show you how it works above your ERP.