Shadow AI, the card category is the real AI budget
Shadow AI is the AI tooling employees buy on corporate cards or personal accounts to ship work this quarter: it does not show up in the EA, and it does show up in the bill, the data risk, and the 2027 budget procurement is about to build. The EA is a procurement instrument that books what was negotiated, not what is being used, and most AI tools arrived after the last renewal, which is why the finance leader who reads only the EA sees a small AI line while the real bill sits in a different ledger.
Prepared by Redress Compliance · August 8, 2026 · GenAI advisory. Based on 30 to 40 enterprise software audits and AI category sweeps supported 2024 to 2025.
Executive summary
The share is 4 to 9 percent of software spend, and it is two to three times the formal line.
The off books AI bill ran between 4 and 9 percent of total enterprise software spend by mid 2025, a 6 percent median, typically two to three times the AI line finance had budgeted, with the spread reaching nearly ten times in the cases we audited.
The size varies by scale, 3 to 5 percent under 500 staff, 5 to 9 in the mid market where no centralized AI roadmap absorbs demand, 4 to 7 at large enterprises where the share looks modest and the absolute dollars are not.
And once a tool clears six months of card renewals it is a budget line in everything but name.
Four tool families carry 60 to 80 percent of the bill, not a long tail of unknowns.
ChatGPT on personal and team plans leads at roughly 22 to 28 percent of card spend, AI coding tools led by Cursor take 14 to 20, Claude 10 to 16, Perplexity and AI search 6 to 10, individual Copilot or Gemini seats 5 to 9, and image, audio.
And video tools 4 to 7, with the fragmented long tail of fifty plus tools at 12 to 18 percent. The concentration is the opportunity, because four category negotiations cover most of the spend, and the engineering pattern is structural: seat prices sit below approval thresholds.
So leaders buy per team and procurement never sees it.
The buyers are managers with deadlines, and the bans made everything worse.
The purchasers are line managers and senior individual contributors with a card and a deadline, not rogue actors dodging procurement, and the blanket bans we audited did not reduce shadow AI volume, they moved it from cards to personal accounts, removing the only audit trail finance had.
The data exposure is the larger and less reversible risk: personal account terms are not the enterprise terms procurement would negotiate, the data handling differs from the enterprise contracts, and activity on personal accounts was never covered by any contract at all.
The pattern that works: consolidate, contract by category, and budget from the card data.
Consolidate the few high value tools onto enterprise agreements where the terms actually cover the data; contract by category, general assistants, coding tools, search, and media, rather than tool by tool.
And budget AI as its own line sized from the card category, the real meter, growing quarterly once the first subscription lands.
The 2027 budget built from the EA will be wrong by construction, and the one built from twelve months of card data will be the first accurate AI number the organization has had.
The card category map, where the bill concentrates
| Tool family | Share of card spend | How it lands |
|---|---|---|
| ChatGPT, personal or team | About 22 to 28 percent | Personal accounts and manager bought team plans |
| Cursor and AI coding tools | About 14 to 20 percent | Per team on developer cards, under approval thresholds |
| Claude, personal or team | About 10 to 16 percent | The same pattern, individual and small team plans |
| Perplexity and AI search | About 6 to 10 percent | Individual subscriptions across functions |
| Individual Copilot or Gemini seats | About 5 to 9 percent | Bought outside the enterprise agreement |
| The long tail, 50 plus tools | About 12 to 18 percent | Fragmented, small, and churning |
The contract gap is the real exposure, more than the spend.
The vendor terms on a personal ChatGPT or Claude account are not the enterprise terms procurement would negotiate, the data handling differs materially from the enterprise contracts, and everything pasted into a personal account was processed under terms nobody at the company ever accepted.
The consolidation case is therefore a risk case before it is a cost case: the enterprise agreement buys the data terms, and the seat savings are the second benefit.
Why it grows, and why bans fail
- The buyers are not rogue: line managers and senior contributors with a card and a deadline, purchasing below approval thresholds to ship this quarter's work.
- The EA cannot see it by construction: the agreement books what was negotiated at the last renewal, and most AI tools arrived in the 24 months since.
- The growth is quarterly: once the first card subscription lands in a function, adjacent teams follow, and six months of renewals makes it a de facto budget line.
- Bans relocate, they do not reduce: the bans we audited moved spend from corporate cards to personal accounts, removing the only audit trail and the only contractual coverage finance had.
- The card feed is the meter: the corporate card category, reviewed quarterly, is the only data source that tracks the real AI estate.
The AI platform contract negotiation brief
The category consolidation method, the enterprise terms that matter, and the negotiation sequence for the four families that carry the bill.
Get the white paper →Bringing it back on books, the working pattern
The sequence that worked ran in three moves: consolidate the few high value tools, the four families carrying 60 to 80 percent, onto enterprise agreements whose data terms actually cover the activity, converting the card sprawl into three or four negotiated contracts.
Contract by category rather than tool by tool, because the general assistant, coding, and search categories each consolidate to one or two vendors once usage is visible; and budget AI as its own line sized from twelve months of card data, the first AI number the organization can defend.
The enterprise side economics of the consolidation targets run through the token usage cost report, the seat side arithmetic in the Copilot licensing guide, and the wider spend context in the price increase index.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across the sweeps, 2024 to 2025
Across roughly 30 to 40 enterprise software audits and AI category sweeps our team supported between 2024 and 2025, the gap between what the EA showed and what the cards showed was the recurring surprise:
The corporate card AI category against the formal budget line, by mid 2025.
Audited bans moved spend to personal accounts instead of reducing it, deleting the audit trail.
The report's framing is directional, bands rather than points, drawn from the engagement file, vendor public pricing, and a rolling benchmark panel.
And the direction is consistent everywhere: shadow AI is not a compliance failure to punish but a demand signal to contract, because the employees buying it are shipping work with it.
The organization that reads the card category learns what its people actually use, negotiates enterprise terms for exactly that, and enters 2027 with an AI budget built from evidence.
The organization that bans it keeps the spend, loses the visibility, and carries the data exposure with no contract behind it.
Your first five moves
- Pull twelve months of the corporate card AI category, the real meter the EA cannot see by construction.
- Map the spend to the four families, where 60 to 80 percent concentrates and four negotiations cover most of it.
- Consolidate the high value tools onto enterprise terms, the data coverage case that precedes the cost case.
- Do not ban what you have not contracted, because the audited bans moved spend to personal accounts and deleted the trail.
- Budget AI as its own 2027 line from the card data, not the EA. The cost optimization practice runs the sweep with you.
Frequently asked questions
What is shadow AI spend?
The AI tooling employees buy on corporate cards or personal accounts outside procurement: it does not appear in the enterprise agreement, which books only what was negotiated at the last renewal, and most AI tools arrived in the 24 months since.
It shows up instead in the corporate card category, the data risk, and the budget gap between the formal AI line and reality.
How big is shadow AI spend?
Between 4 and 9 percent of total enterprise software spend by mid 2025, a 6 percent median, typically two to three times the formal AI budget line and reaching nearly ten times in audited cases.
Small enterprises run 3 to 5 percent, the mid market 5 to 9 where no centralized roadmap absorbs demand, and large enterprises 4 to 7 with the largest absolute dollars.
What tools dominate shadow AI?
Four families carry 60 to 80 percent of the off books bill: ChatGPT on personal and team plans at roughly 22 to 28 percent of card spend, AI coding tools led by Cursor at 14 to 20, Claude at 10 to 16, and Perplexity with AI search at 6 to 10, plus individual Copilot or Gemini seats and media tools.
The long tail of fifty plus tools holds only 12 to 18 percent.
Should companies ban shadow AI tools?
The bans we audited did not reduce volume: they moved spend from corporate cards to personal accounts, removing the only audit trail finance had and pushing activity onto terms no contract covers.
The working pattern is the opposite, consolidate the high value tools onto enterprise agreements, contract by category, and treat the demand as a signal rather than a violation.
What is the biggest shadow AI risk?
The data exposure on personal accounts, because it is the hardest to undo: personal plan terms differ materially from the enterprise contracts procurement would negotiate, and everything processed through them was handled under terms nobody at the company accepted.
The consolidation case is a risk case before it is a cost case, since the enterprise agreement buys the data terms first and the seat savings second.
How should AI spend be budgeted for 2027?
As its own budget line sized from twelve months of corporate card data, the real meter, rather than from the EA, which understates AI spend by construction.
The card category, mapped to the four dominant families and reviewed quarterly, is the first AI number the organization can defend, and the consolidation negotiations it enables are what bring the growth curve under contract.
