By:
Pete Clay
Blog
AI Industry
November 17, 2026

Transcript

How to Build a Shadow AI Inventory for Banks | Kindo

How to Build a Shadow AI Inventory for Banks

Before You Secure Your AI, Find It: How to Build a Shadow AI Inventory for Financial Institutions

Key takeaways

  • A shadow AI inventory is one register of every AI use in the institution, sanctioned or not, with an owner, a risk tier and a decision for each entry.
  • FS-ISAC guidance, the FS AI RMF, the CSBS framework and OCC Bulletin 2026-54 all assume this inventory exists.
  • Proxy, identity, endpoint and SIEM data already show most AI use. The hard part is correlating them, not collecting them.

In Part 1, The AI You Don't Know Will Hurt You, I covered why examiners now ask where AI is used and why an inventory built from approvals misses most of it. This post covers how to build one: what the guidance expects, which data shows AI usage, and which fields to record.

What do FS-ISAC and regulators expect in an AI inventory?

FS-ISAC's AI guidance asks for three inventories: your data, your assets and your vendors. It takes a fourth for granted, the AI itself, and that is the one most institutions cannot produce.

What do FS-ISAC and regulators expect in an AI inventory?
FS-ISAC publicationWhat it asks you to track
More Opportunity, Less Risk: 8 Steps to Protect Financial Services Data with GenAI (January 2025)Step 3: a data lineage inventory listing each dataset used with GenAI, its classification, who may access it, its origin and the use cases it feeds, kept current enough for traceability reporting. Step 8: transparency on where vendors store your data.
Sector Risk Advisory: Preparing the Enterprise for AI-Enabled Vulnerability Discovery (2026)Recommendation 4: a real-time asset inventory including dependencies, connections and third-party exposure, good enough for same-day decisions.
AI risk white papers (February 2024)An acceptable use policy framework for external generative AI and a vendor risk assessment tool. Both presume you know which external AI tools and vendors are in use.

Put them side by side and the gap is plain. You cannot trace datasets to AI use cases you do not know about. You cannot enforce an acceptable use policy on tools you have never seen. You cannot claim a real-time asset inventory while the AI connections in your environment are missing from it.

The regulators close the loop. The Treasury-backed FS AI RMF anticipates a central inventory covering built, vendor-embedded and shadow AI. The CSBS framework asks for an inventory entry with a named owner even at the lowest risk tier. OCC Bulletin 2026-54 expects a complete inventory of generative, agentic and vendor-embedded AI.

What data already shows AI usage across the institution?

A regulated institution already collects four classes of telemetry that, read together, show who is using which AI, from where, on what account and how much. You do not need a new sensor to start. You need to read what you already pay to store

What data already shows AI usage across the institution?
SourceWhat it showsWhat it misses on its own
Proxy, secure web gateway and DNSTraffic to AI apps and model APIs: which users and devices, how often, and how many bytes went out. Upload volume is the quickest indicator of data exposure.Whether the account is corporate or personal, AI inside approved SaaS, and anything off-network.
Identity provider and SSOWhich AI apps are federated, who signs in, and which third-party apps and agents hold OAuth grants to mail, files and calendars.Sign-ins with personal email, and tools that never touch your identity provider.
Endpoint (EDR, MDM, software inventory)Installed AI desktop apps, browser extensions, coding assistants, local model runtimes and the processes that call them.Unmanaged and personal devices, and browser-only use with no install.
SIEMWhere the first three already land, plus cloud audit logs, SaaS admin logs and DLP events, in one timeline with one user identity.Nothing new. It only knows what you send it.

Each source covers another's blind spot. Proxy logs see the personal account that SSO cannot. SSO sees the OAuth grant that the proxy reads as ordinary API traffic. The endpoint sees the extension neither one notices. A user who appears in proxy data for an AI tool but never in SSO for that tool is almost certainly on a personal account.

Two more sources close most of the remaining gap. Cloud and SaaS audit logs show model API calls from your own cloud accounts and AI features switched on inside platforms you license. Expense and procurement records show subscriptions and API keys bought on corporate cards.

Collection is not the hard part. Correlation is. Thousands of AI domains, several log formats and three different ideas of who a user is will break a spreadsheet within a week. Plan for a process that resolves every record to a person and a business unit before you start.

What fields should a shadow AI inventory include?

Keep one record per AI use, sanctioned or not, with enough detail to assign an owner, a risk tier and a decision. These fields track what the OCC and CSBS frameworks have examiners ask about, so one register can serve security, risk and the exam.

What fields should a shadow AI inventory include?
FieldWhat to recordWhy it matters
Tool, model or agentName, provider, version or plan tierConsumer and enterprise tiers of the same product carry different data terms
TypeStandalone app, embedded SaaS feature, browser extension, API or SDK, local model, agentEach type needs a different control
Approval statusApproved, under review, unapproved, prohibitedSeparates shadow from sanctioned
Users and business unitsWho, how many, which lines of businessShows where demand is and who owns the risk
Account typeCorporate tenant or personal accountLogging and contract protections end at the personal account
Business purposeThe use case in plain wordsExaminers test use against approved purpose
Data exposureData classes reaching the tool (customer NPI, MNPI, source code, credentials) and volumeDrives the risk tier and any notification duty
Permissions and autonomySystems it can reach, actions it can take without a personAgents with standing access are the highest-consequence entries
Third partyVendor, subprocessors, data residency, retention and training termsThird-party risk does not transfer with the work
Named ownerAn accountable executive and a use-case ownerThe first thing both frameworks expect
Risk tierTier under your own scheme, with the reasonSets the depth of control and review
EvidenceLog sources, first seen, last seenMakes the entry auditable and repeatable
DispositionApprove, move to a sanctioned alternative, restrict or block, with a dateAn inventory with no decisions is a list of worries

How do you keep a shadow AI inventory accurate?

Three rules keep the inventory honest:

  • Treat it as a feed, not a project. A one-time survey goes stale within a quarter. New tools appear weekly, and vendors keep adding AI to products you approved years ago. Rebuild the register from telemetry on a schedule.
  • Include the approved tools. Examiners ask for a complete inventory of AI uses. Shadow and sanctioned entries belong in one register with one set of fields.
  • Start with amnesty, not a block list. People use shadow AI because it works. Block first and usage moves to personal phones, where you see nothing. Use the inventory to find the demand, then meet it with something sanctioned.

Get a free 30-day Shadow AI Discovery assessment

Building this register by hand is slow. Kindo offers financial institutions a free 30-day Shadow AI Discovery assessment that builds it for you. We connect to the identity, proxy, firewall and SIEM data you already collect, correlate the records to people and business units, and within 30 days you have:

  • A shadow AI inventory with the fields above, 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

‍

FAQs

No items found.