What is ISO/IEC 38507:2022?

ISO/IEC 38507:2022 is an international standard providing guidance to governing bodies on enabling and governing the use of AI. ISO's public abstract says the guidance seeks effective, efficient and acceptable AI use. It applies to current and future uses of AI, their implications for the organization, and organizations of any type or size. The public record identifies the first edition and its April 2022 publication date. A licensed copy supplies the clause-level text needed for a formal implementation mapping.

Is ISO 38507 a management system standard?

ISO's public page describes ISO 38507 as governance guidance for members of a governing body and a wider community of executives, specialists, authorities, service providers and auditors. A management system requires operating processes, assigned responsibilities, reviews and retained evidence below that governance layer. The board can use ISO 38507 to frame direction and oversight while management uses an operating model and related standards to run controls. Keep the governing decision and the management evidence connected through stable AI system and use-case identifiers.

What should directors receive for an ISO 38507 review?

Give directors a bounded set of high-consequence uses with intended outcomes and accountable executives. Include the approved risk tolerance, user and data scope, model destination, human review point, action ceiling and third-party dependency. Add material exceptions, incidents, provider changes and overdue remediation. Finish each item with a decision requested from the board, an owner and a due date. Include one traced operating sample for each high-consequence use so directors can test the join between approval and production evidence.

Can an AI gateway satisfy ISO 38507?

An AI gateway supplies one evidence stream for authenticated HTTP LLM traffic deliberately routed through it. It can enforce policy on caller context, prompt classification and model destination, then record the policy version and outcome. Governance also covers uses and consequences beyond that route. Training systems and local execution need separate controls. Vendor-managed inference, downstream business actions, human review and supplier contingency evidence stay with their assigned owners. The board combines those sources into its organizational decision record.

How often should the board review AI governance?

Set a regular cadence for portfolio direction, risk tolerance and resources, then add event-driven triggers for material changes. A quarterly board review may fit many organizations. A new high-consequence use, provider change, expanded data scope, increased agent authority, serious incident or failed control should activate the delegated escalation path sooner. Record who investigates each trigger, the response deadline and the condition that returns the matter to the board.

ISO/IEC 38507 for Boards: Turn AI Governance Into Reviewable Decisions

ISO/IEC 38507:2022 addresses a specific audience: the people who govern an organization. Its public scope asks governing bodies to enable and govern AI so its use remains effective, efficient and acceptable. That puts the board above individual model approvals. Directors need a repeatable way to set direction, review evidence and intervene when actual use departs from the approved position.

I would reject a board pack built around a green maturity score. A single red exception row, tied to an accountable owner and a production event, tells directors more than a polished 40-slide heat map.

TL;DR

ISO/IEC 38507:2022 gives governing bodies guidance for effective, efficient and acceptable AI use across organizations of any size.
Build the board packet around decisions: approved uses, accountable owners, risk tolerance, third-party exposure, exceptions and required action.
Connect policy to operating evidence with stable system IDs and named review triggers. Quarterly self-reporting misses changes between meetings.
An authenticated HTTP proxy can enforce and record routed LLM calls. Boards still need separate evidence for training, local execution, vendor-managed AI and downstream actions.

The ISO 38507 public scope starts with the governing body

The official ISO/IEC 38507:2022 publication record describes guidance for members of a governing body to enable and govern the use of AI. ISO frames the desired use as effective, efficient and acceptable. The first edition was published in April 2022 and runs to 28 pages.

Its scope reaches current and future AI use and the implications for the organization itself. ISO says the guidance applies to public and private companies, government entities and not-for-profit organizations, regardless of size or dependence on data and information technology. The intended community also includes executives and advisers, public authorities, service providers, assessors and auditors.

The public record establishes audience and scope. Clause-level implementation remains inside the licensed standard. The board packet below is my operating pattern for applying that visible scope. Teams should map its labels to their licensed copy before treating it as a formal conformity artifact.

A board packet should expose six decisions

Give the board a decision surface. Start with the AI uses whose failure could change customer treatment, regulated data handling, financial exposure or the organization's ability to operate. Give every use a stable ID that also appears in the AI model inventory and risk register.

A useful board packet exposes six decisions:

Use and intended outcome: Name the business process, affected people, allowed output and prohibited use.
Accountability: Identify the executive owner, technical owner, data owner and person authorized to accept residual risk.
Risk tolerance: State the loss, rights impact, data exposure or service interruption the organization will accept, plus the condition that forces escalation.
Operating constraint: Record the approved users, data classes, model destinations, review points and action ceiling.
Third-party position: Identify the provider, material dependencies, contract evidence, change notice, incident route and exit option.
Board action: Ask for approval, restriction, remediation funding, risk acceptance or retirement, with a date and named owner.

A packet that ends with "for awareness" lets every unresolved issue drift into the next quarter. Each item needs a requested decision.

NIST turns governance into accountable operating work

The NIST AI Risk Management Framework 1.0 provides a materially different official source for operationalizing governance. GOVERN 1 calls for policies and practices that manage AI risks, periodic review with clear responsibilities, and an inventory resourced according to risk priorities. GOVERN 2 assigns documented roles and lines of communication, training for responsible people, and executive responsibility for AI development and deployment decisions.

NIST also addresses third-party AI. GOVERN 6 calls for policies covering risks arising from third-party software, data and supply chains, plus contingency processes for failures in high-risk third-party AI systems. Apply that test to every high-consequence use by connecting policy, ownership, inventory, review cadence and supplier contingency to evidence that the arrangement operates.

The AI governance operating model can assign those decision rights below the board. ISO 38507 keeps the governing body focused on organizational direction and oversight. NIST supplies a vocabulary for the management processes and accountable roles that make the direction executable.

Evidence should connect an approved use to observed operation

A board approves a position at one point in time. Production changes daily. New model versions arrive, users activate connectors, prompts begin carrying different data, and agents gain wider action permissions. The review packet needs evidence that catches those changes before the next scheduled presentation.

For each use, preserve the approval record and current owner. Link the approved provider route and data classes to observed model traffic. Add the active policy version and recent test results. Show exceptions, incidents, vendor changes and overdue remediation in the same record. The AI governance audit framework separates control design from proof that the control operated.

I want directors to see one traced sample for every high-consequence use. Open the use record, follow a dated production event, inspect the applied policy, and finish at the human or system disposition. The sample exposes broken joins that aggregate percentages hide. A missing identity, stale approval or unrecorded destination becomes visible immediately.

Review triggers belong beside the quarterly cadence

Quarterly review supports strategy and resource decisions. Defined triggers handle material changes between meetings. The governing body should approve the trigger categories and delegation path, while management assigns the people who investigate and act.

Useful triggers include a new high-consequence use, a material model or provider change, expansion into a new data class, an increased agent action ceiling, a control failure, a serious incident, an overdue remediation item and a supplier contingency event. Each trigger needs a response owner and a deadline. It should also identify which decisions return to the board.

The committee charter remains a separate artifact. An AI governance committee charter defines membership, authority, cadence and evidence rights. The ISO 38507 board packet records the governing decisions and the evidence supporting them. Keeping those jobs separate prevents the charter from becoming a quarterly status report.

The authenticated HTTP proxy covers one bounded evidence stream

DeepInspect's product boundary is HTTP AI traffic deliberately routed between authenticated users or agents and LLM endpoints. The calling application supplies identity and use context. At that point, a proxy can inspect prompt and response classifications, evaluate the role and destination against versioned policy, and permit, redact or deny the request before allowed traffic proceeds.

That stream can support the board packet with observed caller identity, model destination, data classification, policy version, decision, action and timestamp. It can also preserve a denied regression test beside a permitted test after a policy change. The resulting record covers routes that traverse the enforcement point.

Other evidence has different owners. Training data and model weights sit with data, model registry, IAM and infrastructure controls. Local models, browser sessions and STDIO tools can bypass the proxy. Vendor-managed inference may stay inside a SaaS provider. Business systems own the action taken after a model response, and human reviewers own their dispositions. The board packet should name each boundary and use its designated evidence source.

DeepInspect

DeepInspect provides the bounded enforcement and evidence stream for authenticated HTTP traffic between users or agents and LLM endpoints. The application supplies caller identity and approved use context. DeepInspect evaluates that context with prompt classification, model destination, role and active policy before allowed traffic reaches the model.

Each routed decision can produce a signed, tamper-evident record outside the calling application's write path. The record contributes observed route evidence to an ISO 38507 board packet. Training, model storage, local execution, vendor-internal processing, downstream actions and human dispositions remain with their designated owners.

Book a technical deep dive at deepinspect.ai.