CisoraThe AI CISO, from the team behind CISO2CISO — now open: try it free for 7 days.
Try CISORA free →
← Executive Intelligence

AI Security

4.8Executive relevance

Your AI Inventory Claim Will Not Survive a Board Question

One 2026 survey found 86% of security leaders claiming a complete AI inventory. Another found only 17% of organizations maintain a comprehensive, current one. Both can be true, because belief and evidence are different things. Here is a method to validate the inventory before you report coverage to the board.

By Marcos Jaimovich · Founder, CISO2CISOReviewed by the CISO2CISO Advisors Team
9 min • 2026-11-16

Executive lens

Strategic signal for CISO-level decisions.

Board relevance

Strategic signal for CISO-level decisions.

Operational impact

Strategic signal for CISO-level decisions.

Executive Summary

Somewhere in your next board pack there may be a sentence like this: "We maintain a complete inventory of our AI systems." It is a simple claim, and it is the foundation for almost every other AI governance statement you will make. Risk tiering, access reviews, incident response, regulatory mapping and third-party oversight all assume you know what exists.

The research on this claim is uncomfortable. The Purple Book Community's State of AI Risk Management 2026 report, a survey of more than 650 senior cybersecurity leaders published by ArmorCode, found that 86% of security leaders claim to maintain a complete AI inventory. In the same report, 59% of respondents admitted that shadow AI is present and ungoverned. Separately, Wiz's AI Threat Readiness Survey 2026, based on 100 senior IT and security leaders, found that only 17% of organizations maintain a comprehensive, up-to-date AI asset inventory.

These two figures come from different samples and different questions, so they do not strictly contradict each other. That is the point. One measures what leaders believe. The other measures what organizations can demonstrate. A board question such as "How do you know?" lives in the space between them, and most inventory claims cannot yet answer it.

Why This Matters Now

The inventory is no longer a hygiene item. It is the evidence base for AI assurance, and the evidence is thin.

A Drata 2026 State of GRC survey of 300 U.S. professionals, as summarized by the Non-Human Identity Management Group, found that only 13% of IT and security professionals say they have full visibility into the AI tools used in their organization. A September 2026 Veeam survey conducted by Censuswide across EMEA enterprises found that 75% have no clear oversight of the AI agents handling their sensitive data. Wiz's survey adds that 52% of organizations discovered unauthorized AI applications running inside their environments over the past year, and that 35% formally allow teams to deploy AI tools without central approval.

Notice what these numbers imply about how inventories fail. They do not fail because security teams are careless. They fail because AI arrives through channels that a classic asset register was never built to watch: a business unit enabling an embedded feature in a SaaS product, a developer wiring a model API into a pipeline, an employee connecting an agent to corporate data with a personal account.

Regulation is also keeping the question alive. The EU's Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the Annex III high-risk deadline to 2 December 2027, with Annex I systems following on 2 August 2028. Legal commentators describe this as a change to the calendar, not to the obligations. You still need to know which of your systems might be in scope, and you cannot classify what you have not found.

CISO2CISO Insight

An inventory is not a list you hold. It is a claim you can defend, and the board should hear how strong the evidence behind it is.

Most AI inventories are built from what people declare: procurement records, project registers, answers to a questionnaire. Declared data is useful, but it measures cooperation, not coverage. The assets most likely to hurt you are, almost by definition, the ones nobody declared.

That is why the 86% and 17% figures can coexist. A leader who reviews a well-maintained register sees a complete list and reports it honestly. What the leader cannot see is the share of reality that never reached the register. Until you test the register against sources that do not depend on people volunteering information, "complete" is an opinion.

The CISO's job is therefore not to produce a bigger number. It is to attach an evidence grade to the number, so the board can distinguish a measured fact from a reasonable belief.

A Method to Validate the Inventory Before You Report It

Treat the inventory like any other control you would put in front of audit: define it, test it, and state the result with its limits.

Step 1: Define the asset classes. "AI" is too broad to count. Split the estate into at least four classes: models (hosted, self-hosted, fine-tuned), agents and automated workflows, pipelines (training, retrieval, orchestration and the connectors around them), and third-party AI, including embedded features in SaaS and AI used by suppliers on your data. Add a fifth bucket for employee-adopted tools. Report coverage per class, never as one blended figure.

Step 2: Define what "in the inventory" means. An entry that lists only a name is not coverage. Set a minimum record: accountable owner, business purpose, data it can access, systems and credentials it can use, environment, lifecycle status and the date someone last verified it. An asset that lacks these fields is "known", not "governed". Keep the two counts separate.

Step 3: Gather evidence from three independent directions. Declared sources are registers, procurement and owner attestations. Observed sources are cloud and identity telemetry, code repositories, SaaS and OAuth grants, network and browser data, and API gateway logs. Tested sources are results from the checks below. A claim backed by all three carries far more weight than one backed by a register alone.

Step 4: Run the reconciliation test in both directions. Draw a random sample from the observed sources and check whether each asset is in the inventory. The proportion found is your measured recall. Then draw a sample from the inventory and confirm each asset still exists and still has the named owner. This second pass finds ghosts, which matter because an agent that was never formally onboarded is unlikely to be formally retired. If you can, plant a known test asset in a business-built environment and see how long it takes to surface.

Step 5: Measure freshness and the discovery rate. An inventory that was accurate in June is a historical document by November. Track the age of the last verification per asset and the number of previously unknown AI assets found each quarter. A discovery rate that is falling means the net is working or the net is shrinking, so interpret it alongside how many discovery sources you actually cover.

Step 6: Document the blind spots. List what you cannot currently observe: personal accounts, unmanaged devices, supplier-side AI, low-code platforms. A visible list of known blind spots strengthens credibility. An empty list does not.

How to Present Confidence Versus Evidence

Boards struggle with AI because they are asked to accept numbers without understanding how solid they are. Give them a consistent structure.

  • Coverage: the measured proportion of observed assets found in the inventory, by class, from the reconciliation test.
  • Evidence grade: declared only, observed and reconciled, or independently tested.
  • Confidence: your judgment, stated as high, medium or low, with the one reason that most limits it.
  • Freshness: when the claim was last validated and when it will be validated again.
  • Known blind spots: the classes or channels not covered, and the plan to cover them.

The wording matters. "We have a complete inventory" is an assertion. "Of the agents we can observe in our identity and cloud telemetry, a measured share are in our register; third-party embedded AI is declared only and we rate our confidence as low" is a statement a director can interrogate and an auditor can reproduce.

Executive Framework

Asset classWhere ground truth livesEvidence that supports a coverage claim
ModelsCloud AI service logs, GPU and container workloads, model registriesSample of observed model endpoints reconciled to the register, with owner and data scope
Agents and workflowsIdentity provider, service accounts, low-code platform admin consolesEvery agent mapped to a unique identity and a named human sponsor
Pipelines and connectorsCode repositories, CI/CD, orchestration and retrieval configurationPipeline and connector list derived from code and config scans, not self-reported
Third-party and embedded AIContracts, SaaS admin settings, OAuth grants, supplier questionnairesSupplier AI use confirmed in writing and checked against observed data flows
Employee-adopted toolsBrowser, network and SaaS discovery dataDiscovery sources documented, with stated gaps for personal accounts and unmanaged devices
Reporting layerValidation log and board packCoverage, evidence grade, confidence, freshness and blind spots reported together

What CISOs Should Do Next

  • Freeze the wording of the claim. Stop reporting "complete" until a reconciliation test supports it. Replace it with a coverage statement that names its evidence grade.
  • Split the inventory into classes and set a minimum record. Agree on the fields that turn a name into a governed asset, and report known versus governed separately.
  • Run a two-way sample test this quarter. Check observed assets against the register and register entries against reality. Record the results and repeat them on a schedule.
  • Assign ownership of AI discovery to security, with business partners. The Cloud Security Alliance's 2026 research on shadow AI recommends that discovery, inventory and classification be explicitly owned and resourced by the security organization. Without a named owner, the inventory decays.
  • Wire the inventory into change processes. Procurement, architecture review, identity provisioning and API onboarding should each require an inventory entry. The NIST AI Risk Management Framework's Govern function expects mechanisms to inventory AI systems, resourced according to risk. Make those mechanisms part of how work gets approved.
  • Publish your blind spots. Add them to the risk register with owners and dates. Treating them as known risks is more defensible than treating them as nonexistent.
  • Link the inventory to regulatory scope. Map observed systems to risk tiers now. The extended EU timeline gives you runway to do this properly, not a reason to postpone it.

Board-Level Questions

  • When we say our AI inventory is complete, what independent evidence supports that statement, and when was it last tested?
  • What share of the AI assets we can observe in our own systems is actually recorded in the inventory, by asset class?
  • How many AI agents or workflows did we discover this year that nobody had registered, and what did we change as a result?
  • Which parts of our AI footprint, such as supplier AI and employee-adopted tools, are declared only, and how confident are we in them?
  • If a regulator or customer asked for a list of our AI systems tomorrow, how long would it take and who would sign it off?

Final Executive Takeaway

The gap between 86% and 17% is not a statistical curiosity. It is the distance between feeling in control and being able to prove it. Neither survey proves what your organization looks like, and both come from vendor-adjacent research with limited samples, which is another reason to rely on your own testing. The practical lesson is that inventory claims should be treated like any other control assertion: scoped, sampled, graded and dated. Do that before the board asks, and the number you present may be smaller than the one you hoped for. It will also be one you can stand behind when the follow-up question arrives, which is the only kind of coverage figure worth reporting.

About the author

Marcos Jaimovich · Founder, CISO2CISO

25+ years in cybersecurity, former regional CISO. Founder of CISO2CISO.