Enterprise AI Governance Best Practices

Why Establishing the Inventory Is the Hard Part

Defining what an AI Agents Inventory should contain is straightforward. Populating it is not.

AI Agents are rarely created through a governance process. They are configured inside AI platforms and assistant builders, switched on as features of products the enterprise already licenses, assembled by business teams without engineering involvement, embedded in automation tooling, or created by other agents.

The agent estate is therefore not a list waiting to be collected. It is a population that must be found, and the agents carrying the most risk are frequently the ones no one thought to declare.

Establishing the AI Agents Inventory is an active discovery effort first and an intake process second.

Discover the AI Agents That Already Exist

Discovery should begin with a deliberate search rather than an announcement and a form.

The enterprise should look where agents are actually created and operated: AI platforms and assistant builders, automation and workflow tooling, customer service and contact channels, development environments, data and analytics platforms, vendor products with agent or assistant capabilities, and business-managed tools acquired outside central procurement.

Discovery should also include people. Business teams, engineers, and support staff routinely know about agents that no automated scan will reveal.

Every discovered agent should be recorded even when it is unapproved, experimental, or expected to be retired, because an inventory that records only approved agents describes the governance program rather than the enterprise. Discovery should then repeat on a defined cadence and after events that tend to create agents, such as platform upgrades, vendor feature releases, reorganizations, and acquisitions.

Harvest Records from the Platforms That Create Agents

Many platforms that host or create AI Agents maintain their own list of the agents configured within them. These platform-native registries are usually the highest-yield source of initial inventory content, and they often surface agents that no person would have reported.

A platform-native registry is not a substitute for the AI Agents Inventory. Each governs only the agents inside its own boundary, describes them in its own vocabulary, and disappears when the enterprise changes platforms.

The AI Agents Inventory is the vendor-neutral system of record that spans all of them and connects agents to owners, use cases, models, data, obligations, controls, and evidence.

Harvested records should be marked as harvested rather than treated as validated. Each should be reviewed by a person, assigned an owner, reconciled against duplicates, and completed to the minimum standard the enterprise has defined.

Discover Shadow Agents Through Identity and Access

Agents that were never registered still leave evidence. They authenticate, hold credentials, call interfaces, reach data, and consume resources.

Identity and access records are therefore one of the most reliable discovery surfaces available. Service accounts, machine identities, integration credentials, tokens, and interface keys can each indicate an agent that no one has declared.

An identity that acts continuously, at machine speed, or with unusually broad permissions deserves investigation. It may belong to a governed agent, an ungoverned agent, or an agent whose owner has left the enterprise.

Shadow agents are the agent-specific expression of Shadow AI and should be treated the same way: as a risk to be governed and as a signal showing where business demand exists and where approved capability is missing. They should be brought into the inventory rather than simply disabled, because removing an agent without understanding the need it served usually recreates it somewhere less visible.

Start with a Minimum Viable Record

An inventory that waits for complete records will never launch.

The enterprise should define a minimum viable record small enough to collect quickly and sufficient to govern: what the agent is, what it does, who owns it, what authority it holds, what systems and data it can reach, and its lifecycle state. Attributes such as evaluation results, monitoring coverage, evidence references, and obligations can be added as the inventory matures.

A partial record with a named owner is more valuable than a complete record that does not exist. Ownership is what makes every later improvement possible.

Establish a Registration Gate

Discovery closes the gap that already exists. A registration gate prevents it from reopening.

The registration gate prevents an AI Agent from reaching production, or from retaining production access, until it has a governed inventory record.

The gate should be enforced where agents actually become operational rather than through policy statements alone: at deployment and publication, credential issuance, permission granting, integration approval, and platform administration. A gate that exists only as a written expectation will be bypassed by the same teams and tools that created the original gap.

The gate should also be proportionate, and its exceptions explicit, time-bound, owned, recorded, and reviewed. A low-authority agent assisting a single team should not face the same intake burden as an autonomous agent acting on customer-facing systems, and an undocumented exception is indistinguishable from a failure of the gate.

Maintain the Inventory Over Time

An AI Agents Inventory decays faster than most enterprise inventories because agents change faster than most enterprise assets. An agent may be given a new instruction set, a different model, additional tools, broader permissions, or a wider audience without any formal change event taking place.

The enterprise should establish a review cadence proportionate to each agent’s authority and risk. High-authority and externally facing agents warrant frequent review; low-authority internal agents need far less.

Re-attestation asks the named owner to confirm that the record remains accurate and that the agent is still needed. An owner who cannot confirm a record is itself a governance finding.

Drift detection compares an agent’s current behavior, authority, permissions, connections, and model against what was approved and recorded. Drift is normal; undetected drift is not. Reassessment should also be triggered by events rather than by the calendar alone, including model changes, instruction changes, permission changes, incidents, and ownership changes.

Decommission Agents and Remove Their Access

Retirement is an inventory event, not only an operational one. When an agent is retired, its record should be updated, its credentials revoked, its permissions removed, its integrations disconnected, and its evidence retained for the required period.

A retired agent that keeps its identity and access is not retired. It is an unmonitored agent with live permissions and no owner paying attention to it.

Orphaned credentials, dormant machine identities, and disconnected integrations should be treated as discovery findings and traced back to the agents that created them. The enterprise should be able to show when an agent stopped operating, who authorized the retirement, and what access was removed.

Measure Inventory Completeness and Freshness

An inventory that is never measured will be trusted more than it deserves.

Useful measures include the proportion of discovered agents that have been registered, the proportion of records reviewed within cadence, the proportion of agents with a named and responsive owner, the age of the most recent attestation, the volume of drift findings, and the time between an agent being created and being registered.

A registration backlog that grows faster than it is cleared is an early warning that the gate, the intake process, or the ownership model is not working. These measures belong in AI governance reporting alongside risk, incident, and compliance measures, because each of those depends on the inventory being right.

Relationship to the AI Agents Inventory and Attributes Document

This chapter addresses how an enterprise establishes and sustains its AI Agents Inventory. It does not define the full attribute set that each agent record should carry.

The IF4IT AI Agents Inventory and Attributes document provides that baseline. It defines the AI Agent as a governed Noun Type and sets out a suggested attribute baseline organized into categories, with maturity guidance indicating which attributes are practical early and which belong to a more mature inventory.

The two documents are complementary: this document establishes why the inventory is needed and how it is built and maintained, while the AI Agents Inventory and Attributes document defines what each record contains and how it relates to other governed inventories. Enterprises should begin with the minimum viable record described here and adopt the fuller baseline as the inventory matures.

Governance Questions Establishing and Maintaining the AI Agents Inventory Should Answer

For establishing and maintaining the AI Agents Inventory, governance should answer which agents exist, which remain undiscovered, which records are current, which have gone stale, who owns each, and which risks, obligations, controls, evidence, incidents, changes, and gaps require action.