The shadow AI bill is already bigger than your shadow SaaS bill

We run a private benchmark across 80+ mid-market tenants that use SpendSlicer to reconcile their SaaS and AI spend. Every tenant's finance team maintains a "sanctioned tools" list — the set of vendors that get paid from AP with a contract. Every tenant's IT team maintains an SSO inventory — the set of tools employees log into daily. And every tenant's engineering team has an AI gateway or a set of API keys wired up to vendor consoles.

In theory those three views should overlap almost entirely. In practice, the gap is the story.

The headline number

Across our sample, 34% of AI spend is "shadow" — paid for on personal or corporate cards and never reconciled against a vendor line item in AP. For SaaS, the shadow share is smaller but still meaningful at 18%.

In absolute terms, shadow AI has already surpassed shadow SaaS in 62% of tenants — even though AI is, on average, only 11% of total stack spend. The per-dollar leakage rate is 3–5× higher.

Why AI shadow grows faster

Three structural reasons:

• The minimum viable purchase is a $20/mo personal plan. Anyone with a credit card and a problem can become an AI customer today. The friction to start is essentially zero.
• Engineering teams spin up provider accounts directly. A new feature ships, an API key gets cut, a billing account starts accruing — often months before finance sees it.
• Usage scales non-linearly. SaaS spend is roughly per-seat and predictable. AI token spend can 10x in a deploy window. A team that was "barely using it" last month is the top-3 line item this month.

Breakdown by role

Who's expensing the most shadow AI? The answer surprised us — it's not engineers. Across our sample:

• Marketing and content: 31% of shadow AI spend. Heavy use of ChatGPT Plus, Claude Pro, Midjourney, Runway, and a long tail of writing assistants.
• Sales: 24%. Note-taker subscriptions, personalized-email tools, research assistants — almost all on personal cards, almost never reimbursed through AP.
• Engineering: 22%. Less than you'd think. Eng spend tends to be visible (API keys on corporate cards, gateway usage) because the dollar amounts force conversation.
• Product and design: 12%. Figma AI add-ons, image gen, prototyping tools.
• Everyone else: 11%. A long tail spread across ops, HR, recruiting, finance.

Breakdown by model

On the API side, where usage lives in vendor consoles rather than on cards, a different pattern shows up. Over 70% of tokens at the median tenant are being sent to the most expensive model in a given vendor's lineup — even when the work being done would route cleanly to a cheaper one.

We saw one tenant sending 90%+ of their OpenAI traffic to GPT-4o when about 40% of the calls were simple classification tasks a Haiku-class model would handle at a tenth of the cost. Another was paying for Claude Opus on a large portion of internal-tool chat where Sonnet would've been indistinguishable.

What actually works to close the gap

1. Pull card data, not just AP data. This is where shadow SaaS and shadow AI both live. Ramp, Brex, and corporate-card statements are the fastest path to surfacing the full picture.
2. Map provider consoles directly. OpenAI, Anthropic, Bedrock, and Vertex all expose usage APIs. Reading them is read-only, no disruption to engineering workflow, and you get per-key attribution.
3. Route, don't ration. The instinct when AI bills spike is to cap. The better move is to route — send the expensive model the expensive work and let a cheaper model handle the rest. You keep quality up and the bill drops.
4. Make it easy to bring shadow out of the shadows. If the path to "get this tool sanctioned and reimbursed" takes three weeks, employees will keep expensing it personally. A lightweight intake form that AP can act on in 48 hours pays for itself.

The bottom line

Shadow AI spend is not a compliance problem waiting to happen. It already happened. The question is whether you're looking at it or not. In our data, tenants who connected their AP + card + identity + provider consoles to SpendSlicer within their first week found an average of $78,000 in untracked AI spend already running on autopilot.

You don't have to do anything with that number. You just have to know what it is.