For most mid-market companies, the line in the budget labeled “AI” is the smallest piece of what they’re actually spending on AI. The bigger piece is scattered across SaaS renewals, productivity suite upgrades, professional services line items, and credit card statements nobody is reconciling. Call it "shadow AI" spend. It’s growing fast, and it’s almost never on the dashboard. This isn’t hypothetical.The gap between *stated* AI budget and *actual* AI run-rate routinely lands at 2-4x. The variance comes from five places, and none of them show up under “AI” in the general ledger.
Where the spend is actually hiding?
- Embedded AI features in existing SaaS - Microsoft 365 Copilot. Salesforce Einstein. Slack AI. Notion AI. Atlassian Intelligence. ServiceNow Now Assist. HubSpot Breeze. Most of these were optional add-ons two years ago. They are now default upsells at renewal, and a non-trivial share quietly auto-activated through vendor releases. Many add $10–$30 per user per month on top of existing contracts. Multiply by seat count, and the number is material before anyone files it under AI.
- Per-token and per-request consumption - OpenAI, Anthropic, Bedrock, Vertex AI, Azure OpenAI is billed by usage, not by seat. A small pilot can compound into a real line item once adoption sticks. The cost lives in cloud invoices alongside compute, not in the application stack, which means it almost never gets attributed to a business owner.
- Tier upgrades to unlock AI features - Plenty of vendors gate their AI capabilities behind the next pricing tier. The vendor books a renewal uplift. Finance sees a price increase. Operations sees more features. Nobody connects it back to the AI category.
- Employee-purchased “BYO AI.” - ChatGPT Plus. Claude Pro. Cursor. Perplexity Pro. Individual subscriptions, often expensed, often on personal cards that get reimbursed. This is the cleanest form of shadow AI. It is invisible to procurement, invisible to security, invisible to the AI strategy document.
- AI work bundled into professional service - Systems integrators and consulting firms now build AI scoping, prompt engineering, and model fine-tuning into broader engagements. The capability investment is real, but it is nearly impossible to extract from a master services line on an invoice.
Why this is happening now?
Two forces are converging. First, vendors have finished their AI productization cycle. The features that were “coming soon” in 2024 are now in production, monetized, and pushed through 2026 renewals. Second, employee adoption has outpaced procurement controls. The people doing the work found tools that helped, started using them, and finance is catching up after the fact.
The net effect: organizations are operating real AI infrastructure; touching customer data, drafting external communications, generating code in production without the budget visibility, governance, or risk posture that a deliberate AI program would require.
What to do about it?
This is not solved by tighter procurement alone. A practical starting point:
- Run an AI spend audit, not a SaaS audit - Pull the last twelve months across SaaS, cloud, professional services, and expense reports. Tag anything that touches a model, embedded or standalone. The number will be larger than expected. That number is the actual baseline.
- Force AI features out of the bundle - At renewal, ask vendors to itemize the AI portion of pricing. Most will resist; the negotiation itself surfaces the real cost.
- Assign a single owner for AI spend visibility - Not approval authority but visibility. Without a person whose job is to see the full picture, the picture stays fragmented across IT, finance, and individual cost centers.
- Build a sanctioned-tools list before banning the shadow ones - Employees using ChatGPT on a personal card are responding to a real productivity need. Cutting off the path without offering an approved alternative just pushes the spend further underground.
The companies handling this well are not the ones with the most restrictive AI policies. They are the ones treating AI spend as a real budget category; measured, owned, and reviewed rather than a footnote inside everything else. The AI line item nobody budgeted for is already there. The question is whether finance and technology leadership are looking at it together, or watching it grow in separate spreadsheets.
