Meet your systems: mapping the AI your own organization already runs
The short answer
You cannot govern what you cannot see
Every governance mechanism in this program (risk classification, evaluation, logging, incident response) operates on a known system. Applied to a system you never discovered, all of them do nothing. Discovery is not a preliminary to governance; it is the governance that is missing when things fail.
What you will be able to do
- Build an AI systems inventory for your own organization: a working list of the AI systems it already runs, with the fields a governor needs to act on each one.
- Distinguish what a system is marketed as from what it actually is, and record the real autonomy level, the human involvement, and the vendor-versus-built origin of each system.
- Find the AI you were not told about: embedded features inside tools you already bought, shadow AI that staff adopted on their own, and third-party AI hiding inside vendor products.
- Decide which fields belong in the inventory for governance (not for engineering) and why each one earns its place: the decision the system touches, the data it uses, who owns it, and how exposed the organization is if it fails.
- Triage the inventory by criticality so the systems that can hurt someone or the organization rise to the top, without pretending you can govern everything at once.
- Set up the inventory as a living artifact, not a one-time spreadsheet, so it stays true as systems are added, changed, and retired.
The ideas, one by one
The estate is mostly hidden
Leadership reliably knows the flagship systems and misses the embedded features, shadow tools, and vendor-hidden AI that together are usually the majority of the estate. The 2024 Work Trend Index found 75 percent of knowledge workers already using generative AI at work and 78 percent of them bringing their own tools. If your count feels quick and complete, you found the tip, not the iceberg.
Record what a system actually is, not what it is called
The Amazon Just Walk Out case is the permanent reminder: a single label can conceal a workforce, a privacy surface, a cost structure, and a dependency. The "what it actually is" field, capturing real autonomy and human involvement, is the one most inventories omit and the one that most often turns out to matter.
Cast the net wide across seven places AI hides
Built, bought, embedded, third-party, shadow, hybrid, and agentic. An inventory that checks only "built" and "bought" misses most of what can hurt the organization. Over-include and prune later; do not under-include and get surprised.
The fields that earn their place are governance fields, not engineering fields
The decision a system touches, what it actually is, who owns it, who it affects, and what happens if it fails. Architecture and hyperparameters belong to the systems that own them. An inventory drowning in engineering detail is one nobody finishes and nobody reads.
Every system needs a named human owner
Ungoverned systems are almost always unowned systems. An unowned high-criticality system is one of the most important things an inventory can surface, because every later obligation needs a person to land on.
Triage by consequence, not sophistication
Criticality is about who is affected and what happens if the system is wrong, never about how impressive the technology is. A keyword filter that screens applicants outranks a dazzling model that rearranges an internal dashboard.
Shadow AI is a signal, not a crime
Most AI use in many organizations is on unofficial tools, which tells you official tooling is missing and real work is happening on unofficial rails. Surface it with amnesty, understand the data flowing through it, then decide what to sanction, replace, or stop. Punishment produces silence, not visibility.
The inventory is a living artifact
A one-time snapshot is stale within weeks. Give it an owner, a refresh trigger, and a date-last-verified stamp on every row, or it will be wrong exactly when a later module or a real incident reaches back for it.
Complete broadly, then govern in criticality order
A large estate cannot be deeply governed all at once. Give every system a row first (a thin row on a known system still governs; an unknown system does not), then spend your deepest attention on the high-criticality systems and let the low ones hold their place. Triage is the professional allocation of finite capacity, not an admission of incompleteness.
The inventory is an index, not an encyclopedia
It deliberately excludes the depth that belongs to the risk register, the data-processing record, and the model card. Its power is that it stays short enough to finish, current enough to trust, and clear enough to act on, while every deeper document hangs off it.
This is the artifact everything else hangs off
The Module 2 data audit, the Module 5 risk classification, the Module 12 successor briefing, and the capstone dossier all begin here. A weak inventory now is a weak everything later.
Listen to it
The same lesson, talked through by two hosts as episode 2 of the podcast.
Real cases
These examples show the marketed-versus-actual discipline and the seven-category net applied to real system types an organization runs. The anchor case is examined first and in depth; the rest are illustrative categories, named where a real product family makes the point concrete.
Example 1: Amazon Just Walk Out (the anchor, examined in full). Amazon's cashierless checkout technology was launched with Amazon Go stores and marketed as computer vision that let shoppers take items and leave without checking out. According to reporting summarized by Bloomberg (2024), the operation involved roughly 1,000 people in India reviewing shopping video, with a meaningful share of transactions historically passing through human review; Amazon countered that these workers mainly trained and validated the model and checked only a minority of visits. In April 2024 Amazon announced it would remove the technology from its larger Amazon Fresh grocery stores in the United States in favor of Dash Cart smart carts, while retaining Just Walk Out in smaller formats and third-party venues. Inventory reading: the marketed label ("autonomous AI checkout") and the actual system (a human-in-the-loop hybrid with an offshore review workforce, a video-privacy surface, and human-dependent economics) diverge sharply. An inventory entry that recorded only the label would have governed none of the real risks. This is the case that justifies the "what it actually is" field for every entry you will ever write.
Example 2: The embedded feature you already bought (customer-relationship and office suites). An organization licenses a major CRM or office suite for reasons that have nothing to do with AI, then the vendor ships AI features (a Salesforce Einstein lead score, a Microsoft Copilot draft, an auto-summary in a help desk) that are enabled and quietly begin influencing decisions. Inventory reading: this is the "embedded" category, and it is the most under-counted because staff think of it as "just a feature of the tool." A lead score that reorders which customers a sales team calls first is an AI system touching a real decision, and it belongs on the list even though nobody ran an AI procurement to acquire it.
Example 3: The bought decision system (resume screening). A company licenses an applicant-screening tool that ranks or filters job applications. Marketed as "AI that surfaces your best candidates," the actual system might be a model, a keyword filter, or a hybrid, and it might run with or without a human reviewing its output. Inventory reading: origin is "bought," people affected is "job applicants," exposure if it fails is "discriminatory hiring decisions and legal liability," and criticality is high. The provenance and accountability sit with a vendor you will later need to interrogate, and the legal watch on hiring AI is active in several jurisdictions. (see Topic 5.3)
Example 4: The vendor's hidden AI (third-party inside a service). A logistics or customer-service provider sells you a service, and part of it is powered by AI the provider does not foreground: an automated triage model, a voice system, an offshore-plus-model hybrid. Inventory reading: this is the "third-party" category, and it is invisible until you ask the vendor directly. The Just Walk Out pattern (a hybrid presented as automation) recurs across the vendor landscape, which is why the marketed-versus-actual field applies to bought and third-party systems just as much as to your own. (see Topic 3.3)
Example 5: Shadow AI (bring-your-own tools). Individual employees use public generative-AI tools to draft emails, summarize documents, write code, or analyze data, on tools the organization never approved. By the 2024 Work Trend Index, this is the majority of AI use in many organizations. Inventory reading: origin is "shadow," owner is initially "unowned" (a red flag in itself), and the data field is the urgent one, because the governance risk is confidential data leaving the organization through an unsanctioned channel. Surfacing these converts invisible use into a decision: sanction, replace, or stop. (see Topic 2.1)
Example 6: The pilot that became production. A team runs a small proof-of-concept AI project with a narrow scope and clear caveats. It works, people like it, and without a formal go-live decision it quietly starts handling real cases. Inventory reading: the danger is that the governance attached to a "pilot" (light, experimental, low-stakes) never got upgraded to match the reality of a production system making real decisions. The inventory's refresh trigger exists partly to catch this transition, because the day a pilot starts touching real decisions is the day its criticality flag changes.
Example 7: The built internal model (churn prediction). A data-science team trains a model that predicts which customers are likely to cancel and feeds a list to the retention team each week. Marketed internally as "our churn model," the actual system is a built model whose output a human team acts on selectively. Inventory reading: origin is "built," so provenance and accountability sit with the internal team and the data it was trained on, not a vendor. People affected are customers who may be offered or denied retention incentives; exposure if it fails is unfair treatment of some customer groups and wasted spend. Criticality is medium: it shapes real customer treatment but does not decide rights, credit, or safety. The "built" origin matters because when the Module 2 data audit follows this system's data trail, the trail leads to the organization's own datasets and choices, with no vendor to interrogate. (see Topic 2.1)
Example 8: The translation or transcription feature (quietly consequential). A support team enables automatic translation of customer messages and automatic transcription of calls, treating both as harmless conveniences. The actual systems are AI models whose errors can distort the meaning of a complaint, misattribute a statement, or mistranscribe a critical instruction. Inventory reading: it is tempting to rate these low because they feel like plumbing, but criticality depends on what rides on the output. If a mistranslated complaint drives a wrong resolution, or a mistranscription enters a record that later matters in a dispute, the exposure is real. This example exists to warn against the reflex of rating "utility" features low without checking what decision depends on their accuracy.
Where people go wrong
- "We already know what AI we run." This is the confident-and-wrong answer, and it is almost universal. Leadership knows the flagships and misses the embedded features, the shadow use, and the vendor-hidden AI, which together are usually the majority of the estate. The cure is not confidence but discovery: invoices, interviews, tool audits, vendor questions, and a shadow-use survey. If your first count feels complete and quick, you have found the tip, not the iceberg.
- "Record what the system is called." The single most damaging inventory habit. The label ("autonomous AI checkout," "AI-driven forecasting," "smart routing") routinely conceals a different operational truth. Record the mechanism, not the marketing. The Just Walk Out case exists in your inventory as a permanent reminder that the label and the reality can be entirely different systems.
- "An AI system means our own trained model." Too narrow by far. Bought products, embedded features, third-party AI inside vendors, and shadow tools are all AI systems you run and must govern. An inventory that lists only home-built models misses most of what can hurt you.
- "More fields make a better inventory." No. An inventory drowning in engineering detail (architecture, hyperparameters, training-set sizes) is one nobody finishes and nobody reads. The governance fields are the ones a governor acts on: the decision it touches, what it actually is, who owns it, who it affects, and what happens if it fails. Hand the engineering detail to the systems that own it.
- "Shadow AI is a discipline problem to stamp out." Treating shadow AI as misconduct drives it further underground and destroys the trust you need to see it. It is a signal that official tooling is missing and real work is happening on unofficial rails. Surface it with amnesty, understand the data flowing through it, then decide what to sanction, replace, or stop. Punishment produces silence, not visibility.
- "Build the inventory once and file it." A one-time inventory is out of date within weeks, because vendors ship features, staff adopt tools, and pilots become production continuously. Design it as a living artifact with an owner, a refresh trigger, and a date-last-verified stamp on every entry, or it will be wrong exactly when a later module or a real incident reaches back for it.
- "Criticality is about how advanced the AI is." Criticality is about consequence, not sophistication. A simple keyword filter that screens job applications is higher criticality than a dazzling generative model that rearranges an internal dashboard, because the filter touches people's livelihoods and the dashboard touches nobody. Rank by the decision a system touches and the harm if it is wrong, never by how impressive the technology is.
- "We do not have AI, so we do not need an inventory." By the 2024 Work Trend Index, three in four knowledge workers already use generative AI at work, most of it on tools the employer never provided. An organization that believes it has no AI almost certainly has the least governed AI of all: entirely shadow, entirely invisible. "We have no AI" is usually a statement about visibility, not reality.
- "A utility feature like translation or transcription is too minor to inventory." The criticality of a feature is set by what depends on its output, not by how humble it feels. A mistranslated complaint that drives the wrong resolution, or a mistranscription that enters a record used later in a dispute, can carry real consequences. Inventory the feature and rate it on what rides on its accuracy, not on the reflex that "it is just plumbing."
- "The inventory is engineering's job, not governance's." Handing discovery to the engineering team alone reproduces the blind spot, because engineering knows the built and integrated systems and rarely sees the shadow tools, the embedded features enabled by a business unit, or the human-in-the-loop reality of a system in daily use. Discovery is a whole-organization activity that reads invoices, interviews business owners, and surveys staff. Governance owns the inventory; engineering is one of several sources it draws on.
Questions people ask
- What is AI systems inventory?
- A working document listing every AI system an organization runs, with the governance fields needed to act on each one (name, owner, purpose and the decision it touches, what it actually is, origin, data, people affected, exposure if it fails, criticality, and date last verified). It is the first artifact a governor builds and the one every later governance activity reaches back for. The plain principle behind it: you cannot govern what you cannot see.
- What is discovery?
- The active legwork of finding the AI systems an organization runs, using invoices, interviews, tool audits, vendor questions, and shadow-use surveys together. Discovery precedes every other governance activity, because risk classification, evaluation, logging, and incident response all operate only on systems that are known. More on Discovery
- What is marketed versus actual?
- The discipline of recording what a system actually is (its real autonomy level and human involvement) rather than what it is called. Central to inventory work because a single label can conceal a completely different governance profile, as in the Amazon Just Walk Out case.
- What is Embedded AI?
- AI features switched on inside tools an organization already runs for other reasons, such as a lead score in a customer-relationship platform or a draft-writer in an office suite. The most under-counted category in most inventories, because staff think of these as "just a feature" rather than as AI systems touching decisions. More on Embedded AI
- What is Shadow AI?
- AI tools that individual staff adopt without organizational approval, from public chatbots to browser extensions. In many organizations it is the majority of AI use. It is a signal that official tooling is missing, and its main governance danger is confidential data leaving the organization through unsanctioned channels with no owner and no logs. More on Shadow AI
Keep going
This lesson builds AI inventory and use-case intake, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →