Transcript
Shadow AI in Financial Services: What Examiners Want | Kindo

The AI You Don't Know Will Hurt You: Shadow AI in Financial Services
Key takeaways
- Regular employee AI use tripled in one year, from 15% to 45%, and about two-thirds of it runs on personal accounts.
- Shadow AI is any AI tool, embedded feature or agent operating without approval, risk review or procurement.
- The OCC's new examiner framework, the CSBS AI Supervisory Framework and the Treasury-backed FS AI RMF all expect a complete AI inventory, including unapproved and vendor-embedded AI.
- An inventory built from approvals will miss shadow AI. An examiner who cross-checks your software and vendor records will not.
- You can start with telemetry you already collect: proxy, identity, endpoint and SIEM data.
Every AI control a financial institution owns starts from a list. The acceptable use policy, the model risk review, the vendor due diligence and the DLP rules all assume that list is complete. If the list is wrong, the controls are wrong in exactly the same places.
I have spent 30 years in security, including three tours as a CISO, and most AI lists I see are wrong. Verizon's 2026 Data Breach Investigations Report found that regular AI use by employees tripled in a single year, from 15% to 45%. Shadow AI is now the third most common non-malicious source of data leakage. Reporting on the same dataset puts 67% of that use on personal accounts running on corporate devices.
If you run security or risk at a bank, credit union, insurer or broker-dealer, this now has a deadline. Examiners are asking where AI is used, and the first thing they want is the inventory.
What is shadow AI?
Shadow AI is any AI model, tool, embedded feature or agent running in your environment that never went through approval, risk review or procurement. Almost none of it is malicious. It is your own people getting work done with whatever is one browser tab away.
Why doesn't the approved tools list catch shadow AI?
The approved list is a record of decisions. It tells you what the institution said yes to and nothing about what people did next. Shadow AI comes in through five routes, and none of them passes your review committee:
- Personal accounts on approved tools. The tool is sanctioned and the account is not. Your enterprise tenant has logging, retention and a data processing agreement. The personal account has none of those, and your web filter cannot tell them apart.
- AI features inside products you already own. Your CRM, core platform, collaboration suite and help desk have all shipped assistants in the last two years. Nobody filed a request, because nobody bought anything.
- Browser extensions and desktop assistants. Note-takers, writing helpers and summarizers read whatever the user reads, including your internal applications.
- Developers and analysts. API keys on a corporate card, coding assistants in the IDE, models pulled down to a laptop, and agent frameworks wired into internal data.
- Agents with borrowed credentials. Someone grants an AI agent OAuth access to mail, files or ticketing. That agent now holds standing permissions that no access review has looked at.
Who is using shadow AI?
Senior leaders, more than anyone. A March 2026 survey of 2,001 UK and US employees found senior decision-makers more than twice as likely to use unapproved AI as the people who report to them (65% against 31%). Their top reason was limited access to approved tools. The people who write the policy are the ones going around it.
The cost shows up later. IBM's 2026 breach research, as audited by StationX, found a shadow AI incident at 43% of breached organizations, up from 20% a year earlier.
Why are bank examiners asking for an AI inventory now?
Every other AI question depends on the inventory. A risk tier, a named owner, a vendor contract and a kill switch all attach to a line in it. No line, no control. Four developments in 2026 turned this into exam prep:
- The sector got a control framework. The Financial Services AI Risk Management Framework, released through Treasury in February with 230 control objectives, anticipates a central AI inventory covering what you built, the AI embedded in what you bought, and shadow deployments. Shadow AI is in scope by name.
- AI became a standing exam topic. Reuters reporting, summarized by Quartz in July, described OCC and Federal Reserve examiners raising AI governance, vendor risk, data boundaries and kill switches in routine exams. Regulators called it fact-finding. In my experience, fact-finding comes right before findings.
- State examiners got a playbook. The Conference of State Bank Supervisors released its AI Supervisory Framework on September 16. Its scoping questions open with whether the bank uses AI and where, then ask whether the bank has looked for AI embedded in vendor products and whether sensitive data passes through it. Even the lowest risk tier calls for an inventory entry with a named business owner.
- The OCC told its examiners what to look for. On November 4, the OCC issued Bulletin 2026-54, its "Super Intelligence" supervisory framework for bank use of AI and the first AI-specific examiner framework from a federal banking agency. King & Spalding notes that it expects named accountability and a complete inventory of AI uses: generative, agentic and vendor-embedded.
Someone on your board will point out that the OCC says its framework creates no new enforceable standards, and that the April model risk guidance (SR 26-2) left generative and agentic AI out of scope. Both are true, and neither helps you. The framework is how examiners will scope the review and write the document request. Reed Smith reads it to mean that a tool operating beyond its approved purpose can be treated as a control failure even when the output looks fine.
Unapproved AI operates beyond an approved purpose by definition. You do not want to learn what is in your environment from the person examining it.
What does FS-ISAC guidance say about AI inventories?
FS-ISAC's AI guidance asks members to inventory their data, assets and vendors, and it assumes a fourth inventory: the AI itself. Its 8 Steps to Protect Financial Services Data with GenAI calls for a data lineage inventory tracing each dataset to the GenAI use cases it feeds. Its 2026 sector risk advisory calls for a real-time asset inventory with dependencies and third-party exposure. Neither is possible if you do not know which AI is in use. Part 2 of this series maps each FS-ISAC recommendation to the inventory fields that satisfy it.
Where should a financial institution start?
Start with the logs you already pay to store. Proxy and DNS data show traffic to AI apps and model APIs. Your identity provider shows federated AI apps and the OAuth grants agents hold. Endpoint tools show installed assistants, extensions and local models. Your SIEM ties them together. No single source is enough, but read together they cover each other's blind spots. One reliable signal: a user who appears in proxy logs for an AI tool but never in SSO for that tool is almost certainly on a personal account.
Then start with amnesty, not a block list. People use shadow AI because it works. Block first and the usage moves to personal phones, where you see nothing at all. I have watched that mistake made with every technology wave since webmail. Use the inventory to find the demand, then meet it with something sanctioned.
For the full method, including the data sources, the 13 fields examiners will ask about, and how to keep the register current, read Part 2: How to Build a Shadow AI Inventory From Data You Already Have.
Get a free 30-day Shadow AI Discovery assessment
Kindo offers financial institutions a free 30-day Shadow AI Discovery assessment. We connect to the identity, proxy, firewall and SIEM data you already collect, and within 30 days you have:
- A shadow AI inventory covering tools, embedded features, extensions and agents
- A ranked view of where sensitive data reaches unapproved AI, by business unit
- A recommended disposition for every entry, mapped to the FS-ISAC, FS AI RMF, CSBS and OCC material cited here
- A summary you can put in front of your risk committee or your examiner
No cost and no obligation to buy. You keep the inventory either way.
Start your free Shadow AI Discovery assessment
The AI you don't know will hurt you. The AI you can show an examiner will not.
FAQs
Continuously, or on a regular schedule rebuilt from telemetry. A one-time survey goes stale within a quarter as new tools appear and vendors add AI to products you approved years ago.
Not as a first step. Blocking pushes usage to personal devices you cannot see. Inventory first, offer sanctioned alternatives for the demand you find, then restrict or block what remains.
The OCC says Bulletin 2026-54 creates no new enforceable standards. It does tell examiners how to scope AI reviews, and it expects a complete inventory of generative, agentic and vendor-embedded AI with named owners.
No. Senior decision-makers are more than twice as likely as their reports to use unapproved AI (65% against 31%), most often because approved tools are hard to get.

