Why does AI governance start with an inventory?

Every framework's early steps, mapping context, measuring risk and managing systems, consume a list of AI in use. Without the inventory, downstream governance acts on an incomplete universe.

How often should an AI inventory be refreshed?

Quarterly is the practical floor, with monthly cadence for high-churn organizations. Log-based regeneration makes the refresh a calendar event rather than a project.

What should an AI governance inventory record per system?

System identity and category, users, usage intensity, data posture including dated training verdicts, approval state against the sanctioned list, red flags, and change since the last period.

What is wrong with interview-based AI inventories?

They are complete only on the day of the interviews, miss embedded AI features, and lack an update mechanism. Log-based inventories regenerate on demand with dated evidence.

Who should own the AI inventory process?

Split three ways: security or IT runs the quarterly export and upload, compliance reads and files with decision minutes, and business owners decide per category.

What does maintaining a living AI inventory cost?

Roughly a morning per quarter plus report pricing from $99, or plan allowances for monthly runs, using log exports the organization already produces.

policies & oversight

risk assessment

measurement

INVENTORY

Compliance / Governance inventory

Every AI governance framework stands on the inventory

Pick any framework for managing AI risk: they all begin the same way, with knowing what AI is in use, by whom, for what. Skip that layer and everything above it is architecture on sand. This page is about pouring the foundation, and keeping it poured.

Start the inventory free

shadow-ai-audit.pdf

Cover of the sample Shadow AI audit report prepared for an example company, with its headline findings.

Sample report View all 11 pages Opens in your browser · PDF

The universal step one

Why frameworks converge on "know your AI"

Risk-management frameworks, phrased generically, all route through the same early gates. Each gate consumes the inventory.

Mapping the context

Frameworks ask organizations to map where AI touches their operations and stakeholders. A map of systems nobody listed is not a map; the inventory supplies the territory.

Measuring the risk

Measurement needs units: which system, which data, which users. The audit's per-tool rows, with risk flags and dated training verdicts, are those units pre-assembled.

Managing and governing

Policies, review boards and approval gates act on known systems. Shadow AI is precisely the set your governance cannot act on, which makes shrinking it the first governance KPI.

The framing to steal for your board deck: the inventory is not a compliance chore inside the framework, it is the substrate the framework runs on. Everything else is a function of its completeness.

Static vs living

The binder inventory vs the living inventory

Most organizations have attempted an AI inventory once. The failure mode is always the same: it was a project, not a process.

The binder version

Built by interview and spreadsheet over six weeks, complete for roughly a day.
Misses embedded features and everything adopted after the last interview.
Update mechanism: someone remembers to ask again, eventually.
Governance value decays monthly; audit value decays faster.

The living version

Regenerated quarterly from network logs against a register maintained daily, 11,841 classified domains at present.
Catches new tools, new subdomains and changed vendor verdicts as deltas between runs.
Update mechanism: the calendar. No memory required.
Each run is also evidence: dated, repeatable, fileable.

The living inventory costs a morning per quarter. The binder cost six weeks and was wrong by month two. This is the rare governance upgrade that is also a cost cut.

Anatomy

What a governance-grade inventory records

Per system, the fields governance actually consumes, and where the audit supplies them.

Field	Governance use	Source in the audit
System identity	The unit every decision attaches to	Tool domain plus category and subcategory
Who uses it	Scoping for training, oversight and accountability	Per-user or per-device breakdown
Usage intensity	Separates pilots from dependencies	Hits and user counts per period
Data posture	Feeds risk assessment and privacy analysis	Training verdict with check date, sovereignty flag
Approval state	The governance status itself	Sanctioned split against your approved list
Red flags	Escalation triggers	Risk level and abusive-purpose flags
Change since last period	The living part	Deltas between quarterly reports

The inventory's own data footprint stays small: hostnames, identities, timestamps. Uploads are discarded after each run; reports are deletable before their 90-day expiry.

The cadence

A year of quarterly inventory, in practice

What each quarter's run contributes once the loop is standing.

Baseline and triage

First full report; abusive flags blocked, top tools gated for sanctioning, owners assigned per category.

Policy takes effect

Sanctioned list published; the split becomes the adoption metric. Deltas show whether governance changed behavior.

Drift management

Changed vendor verdicts between runs feed renewal decisions; new tools get gated within the quarter they appear.

The annual story

Four PDFs tell the board a measured narrative: exposure found, decisions made, trend bending the right way.

Ownership

Who owns the inventory, and what each role does

The loop survives when three roles each hold one piece. It dies when all three assume another has it.

Security or IT: the runner

Owns the export and the upload, quarterly, same window. Twenty minutes of mechanics per run, documented in the runbook once. The walkthrough is the runbook's first draft.

Compliance or the CISO: the reader

Owns the triage, the decision minutes and the filing. Presents the delta slide wherever governance reports; the compliance officer page details the evidence chain.

The business: the decider

Department owners answer for their category's tools: keep and contract, or drop. Governance that skips this role produces lists nobody follows.

Anti-patterns

Three ways governance inventories go wrong

Inventorying only the approved

A register of sanctioned tools is a shopping list, not an inventory. The governance object is the gap between approved and observed, which only measurement reveals.

Perfection before publication

Teams delay the inventory until it can be complete, which is never. Publish the observed list with its stated scope; completeness grows by cadence, not by waiting.

Inventory without consequence

If findings never change a contract, a block rule or a policy line, the loop decays into reporting theater. Every run should close with at least one decision minuted.

The consequence engine is the sequence on detect before you block: inventory, sanction, block, verify. Governance is that loop with minutes attached.

Inventory depth

What the inventory captures on each tool

Every AI tool in the report carries structured metadata that governance frameworks consume directly, without reinterpretation.

Training-on-data verdict

Whether the vendor trains models on customer inputs by default, with a dated reference to the terms of service reviewed. Changed verdicts between quarters surface as drift findings that trigger vendor reviews and governance escalation.

Data sovereignty flags

Tools flagged for cross-border data transfer, non-EU processing, or lack of data-residency commitments. These flags route directly to GDPR and AI Act compliance queues without manual classification effort.

Abusive-purpose markers

Categories including NSFW generation, deepfake creation, uncensored chat and nudification. These tools require immediate blocking decisions regardless of the governance framework, and their presence in the inventory provides the documentation trail for that action.

Opt-out availability

Whether the vendor provides a mechanism to disable training on customer data, the scope of that opt-out, and whether it applies at the account, API or enterprise tier. This determines whether a tool can move from unsanctioned to sanctioned with a configuration change rather than a vendor switch.

Per-user attribution

When the log export includes an identity column, the inventory shows which users or devices reached each tool, enabling governance decisions that target departments rather than blanket policies. This is the data that turns a tool inventory into an actionable governance dashboard.

Usage volume indicators

Relative traffic volume per tool allows governance committees to prioritize reviews by actual exposure rather than theoretical risk. A rarely accessed tool with adverse flags is a different priority than one with heavy daily use across multiple departments.

Every field is available in the CSV export, making it possible to import the inventory into existing governance tooling, risk registers or GRC platforms without manual data entry.

Framework alignment

How the inventory maps to major governance frameworks

Framework	Requirement	What the inventory provides
NIST AI RMF	Map: identify AI systems in use	Complete tool inventory from network evidence, not self-reporting
ISO 42001	AI management system inventory	Dated, reproducible inventory with classification metadata per tool
EU AI Act	Provider and deployer registration	Identifies tools deployed within the organization, per-user if available
GDPR	Processing activity records	Training verdicts flag tools that process personal data for model improvement
SOC 2	Monitoring and change detection	Quarterly delta reports show new tools, removed tools and verdict changes

Board readiness

Turning the inventory into a board-level narrative

Governance committees and boards consume the inventory differently from the teams that produce it. Four artifacts, each extracted from the same quarterly report, build the story.

The exposure summary

Total AI tools found, broken into sanctioned versus unsanctioned, with the sanctioned percentage as the headline metric. Quarter-over-quarter trend tells the board whether the program is working.

The risk highlight

High-risk and abusive-flag tools with their user counts, summarized in three lines. The board does not need the full table; it needs to know the worst findings and that they have owners.

The decision log

A one-page table of decisions taken since the last report: tools sanctioned, tools blocked, vendors contacted, DPAs signed. Each row references the report that surfaced the finding.

The forward calendar

Next audit date, outstanding vendor responses, upcoming policy reviews. Boards care about whether the loop has a next iteration, not just that it ran once.

Total preparation time for a board slide deck: under two hours, once per quarter. The report's CSV export feeds any charting tool the board already uses.

FAQ

Governance inventory questions

Is a quarterly inventory enough for AI governance?

It is the workable floor. The register behind each run updates daily, so quarterly snapshots inherit current classifications; high-churn organizations step up to monthly.

Which framework does this support?

Any of them, phrased generically: mapping, measuring and managing all consume the same inventory substrate. The audit is framework-agnostic evidence.

Do we need new tooling to maintain it?

No. Exports from your existing DNS filter, proxy or firewall plus the audit cover the loop. The per-source mechanics live in the guides.

How does the inventory handle brand-new tools?

Roughly 300,000 domains are screened daily, so new tools enter the register continuously. The unmatched-domain count in each report keeps the residual visible.

What goes to the board from all this?

Four numbers per quarter: tools found, high-risk count, sanctioned share, delta versus last quarter. One slide, dated PDFs behind it.

What does the standing loop cost?

A morning per quarter plus report pricing: $99 single, packs at $66 and $60, or plan allowances for monthly cadence. Details on pricing.