Table of Contents

Reach SOC 2 Compliance in 6 Weeks or Less.

  /

  / Shadow AI Policy Template: Ready-to-Use IT Framework

Shadow AI Policy Template: Ready-to-Use IT Framework

One in five organizations has already suffered a breach traced back to shadow AI. Meanwhile, 63% of breached organizations either have no AI governance policy at all or are still drafting one. Below is a complete, copy-ready shadow AI policy template with twelve sections, plus guidance on adapting it for your company size, your industry, and the regulatory frameworks you answer to.

The template assumes one hard truth up front: your employees are already using unapproved AI tools. A policy that pretends adoption hasn’t started yet fails on day one, so this one starts from the assumption that it has.

What Is a Shadow AI Policy?

A shadow AI policy is a formal document that defines how your organization discovers, evaluates, approves, and governs AI tools that employees adopt outside official IT channels. The term borrows from shadow IT, the older problem of unsanctioned software and hardware, but the AI version carries sharper risks: data pasted into a public model may be retained, used for training, or exposed in ways the organization can’t reverse.

The policy does three jobs: it separates approved use from unapproved use, gives employees a fast and visible way to request new tools so the sanctioned route beats the workaround, and spells out what happens when someone crosses the line, including how the organization detects it and responds.

Shadow AI Policy vs. General AI Acceptable Use Policy

Many organizations already have an AI acceptable use policy (AUP) and assume it covers shadow AI. It usually doesn’t. An AUP tells employees how to behave inside approved tools. A shadow AI policy governs the tools themselves: which ones exist in your environment, which ones are allowed, and what happens with the rest.

You need both. The AUP handles conduct; the shadow AI policy handles inventory and control. If you only have room for one document, fold the AUP’s data-handling rules into Section 6 of the template below.

The Shadow AI Policy Template (Download Link and Copy-Ready Sections)

We’ve created a compliance safe template for Shadow AI Policy, use the link below to create a copy and customize for your company:

Copy the sections below into your policy management system and replace the bracketed placeholders. The language is plain on purpose. Legalese gets skimmed.

Shadow AI Policy Template Summary

Section 1: Purpose and Scope

This policy governs the acquisition, approval, and use of artificial intelligence tools, features, and services at [Company]. It applies to all employees, contractors, interns, and third parties with access to [Company] systems or data. It covers standalone AI applications, AI features embedded in existing software, browser extensions, AI agents, APIs, and personal AI accounts used for work purposes, on both corporate and personal devices. The purpose of this policy is to enable productive AI use while protecting [Company] data, customers, and legal obligations. This policy does not prohibit AI. It prohibits ungoverned AI.

That last sentence matters. Employees read the purpose statement first, and it decides whether they see the policy as an enabler or a blocker.

Section 2: Definitions and Terminology

Shadow AI: any AI tool, feature, agent, or service used for work purposes without formal approval under this policy. Approved AI Tool: an AI tool listed in the Approved AI Tools Registry (Section 4) and used under a [Company]-managed account. Personal AI Account: an account on any AI service registered to a personal email address or paid for personally. AI Feature: AI functionality embedded within otherwise approved software (e.g., an AI assistant added to a project management tool), which requires separate evaluation. Sensitive Data: data classified as [Confidential] or [Restricted] under [Company]‘s data classification policy, including the prohibited data classes in Section 6.

Define “AI feature” explicitly. Vendors now ship AI additions into already-approved SaaS products every month, and without this definition, those features inherit approval they never earned.

Section 3: Roles and Responsibilities

The CISO (or designated security lead) owns this policy, maintains the Approved AI Tools Registry, and runs the approval workflow. Department heads ensure their teams know the policy and surface tool requests rather than suppressing them. Legal and Compliance review tools that touch regulated data or fall under the EU AI Act, GDPR, HIPAA, or client contractual restrictions. IT operates detection and monitoring controls (Section 9). Every employee is responsible for using only approved tools for work, reporting unapproved AI use they discover, and requesting new tools through the workflow in Section 7 rather than adopting them directly.

Insider Note: In organizations under roughly 200 people, the “CISO” in this section is often the same overworked IT lead who manages laptops. Name a real person, not a title that doesn’t exist yet. A policy that assigns duties to a phantom role is unenforceable, and auditors notice.

Section 4: Approved AI Tools Registry

[Company] maintains a registry of approved AI tools at [location/URL]. For each tool, the registry records: tool name and vendor, approved use cases, prohibited use cases, permitted data classes, account type (enterprise/team/individual), data retention and training settings, risk tier (Section 5), approval date, and next review date. Only tools listed in the registry may be used for work. Tools not listed are unapproved by default. The registry is reviewed [quarterly].

Keep the registry somewhere employees actually look, such as your intranet homepage or IT help center, not buried in a GRC platform they can’t access. An invisible registry recreates the problem the policy exists to fix.

Section 5: Risk Tier Classification (Low, Medium, High)

Each tool in the registry is assigned a risk tier. Low: the tool processes only public or internal non-sensitive data, runs under an enterprise agreement with training opt-out, and produces output that a human reviews before use. Approval by IT Security alone. Medium: the tool processes internal business data or connects to [Company] systems via API or integration. Approval by IT Security plus the data owner. High: the tool processes sensitive data, customer personal data, or regulated data; makes or influences consequential decisions (hiring, credit, medical, legal); or operates autonomously as an agent. Approval by IT Security, Legal, and the executive sponsor, with a documented risk assessment.

Tiering is what keeps the approval workflow fast. Without it, a request for a grammar checker sits in the same queue as a request for an autonomous agent with database access, and both take six weeks.

Section 6: Prohibited Data Classes and Use Cases

The following data may never be entered into any AI tool that is not explicitly approved for it in the registry: customer or employee personal data; protected health information; payment card data; authentication credentials, API keys, and access tokens; source code designated proprietary; unreleased financial results; legal matters and privileged communications; client-confidential material governed by NDA; and any data classified [Restricted]. The following use cases are prohibited in all tiers without executive and Legal approval: automated employment decisions, biometric identification, emotion inference in the workplace, and any use case restricted or prohibited under the EU AI Act.

IBM’s 2025 report found that 65% of shadow AI breaches involved customer personal data, well above the 53% average across all breaches. This section is where you keep yours out of that statistic.

Section 7: Tool Request and Approval Workflow

Any employee may request a new AI tool by submitting [form/link] with: the tool name, intended use case, data classes involved, and account type. IT Security triages the request within [3 business days] and assigns a provisional risk tier. Low-tier requests receive a decision within [5 business days]; medium within [10]; high within [20]. If a decision misses its deadline, the request escalates automatically to [role]. Rejections must state the reason and, where possible, name an approved alternative. Approved tools are added to the registry before use begins.

The deadlines are the whole point. A workflow without service-level targets is a suggestion box, and employees will route around it within a week.

Pro Tip: Publish your approval statistics internally

Publish your approval statistics internally each quarter: requests received, approved, rejected, and average turnaround. When employees can see that 80% of requests get approved in under a week, the sanctioned path becomes the obvious one. Secrecy around approvals feeds the assumption that asking is pointless.

Section 8: Identity, Access, and Authentication Requirements

All approved AI tools must be accessed through [Company]-managed accounts provisioned via SSO where the vendor supports it. Multi-factor authentication is mandatory. Personal accounts may not be used for work data under any circumstances, including free tiers of tools whose enterprise version is approved. API keys for AI services are issued, stored, and rotated through [Secrets Manager] and may not be embedded in code, shared in chat, or pasted into other AI tools. Access to high-tier tools follows least privilege and is reviewed [quarterly]. Departing employees have AI tool access revoked as part of standard offboarding.

See our full breakdown of Identity, Access, and Authentication Requirements for the tooling behind this section. The personal-account clause deserves extra attention in training. Netskope’s 2026 research found that most generative AI users access tools through personal accounts, which sit entirely outside enterprise logging, retention controls, and legal hold.

Section 9: Monitoring, Detection, and Auditing

IT operates technical controls to detect unapproved AI use, which may include DNS and web proxy monitoring of known AI service domains, CASB or SSPM discovery of SaaS and OAuth grants, browser-level data loss prevention for paste and upload events into AI services, and review of network traffic to AI API endpoints. Detection exists to protect [Company] data, not to surveil individuals; findings are handled under Section 10 with a remediation-first approach. IT Security audits the environment for unsanctioned AI at least [quarterly] and reports findings to [governance committee]. Monitoring practices comply with applicable employment and privacy law in each jurisdiction where [Company] operates.

Only 34% of organizations with AI governance policies actually audit for unsanctioned AI, per IBM. Writing this section is easy. Resourcing it is the real commitment.

Section 10: Incident Response for Shadow AI Events

A shadow AI event includes sensitive data entered into an unapproved tool, an unapproved tool connected to [Company] systems, or an approved tool used outside its permitted data classes. Upon discovery, the responder must: (1) identify the tool, account, and data involved; (2) preserve evidence, including prompts and outputs where retrievable; (3) contain by revoking access, disabling integrations, and requesting vendor data deletion; (4) assess whether the event constitutes a reportable breach under GDPR, HIPAA, or contractual obligations, with Legal making the determination; (5) notify affected parties per legal requirements; and (6) complete a post-incident review that feeds back into the registry, training, and this policy. Self-reported events within [24 hours] of occurrence are treated as near-misses, not violations, provided no intentional misconduct occurred.

That final sentence is your early-warning system. IBM found shadow AI breaches take about a week longer than average to identify and contain, largely because nobody wants to admit what happened. A self-reporting safe harbor shortens that timeline more than any monitoring tool.

Section 11: Enforcement and Consequences

Violations are handled proportionally. First unintentional violations result in retraining and documentation. Repeated violations, or any violation involving sensitive data, escalate through [Company]‘s standard disciplinary process up to termination. Intentional concealment of AI use after a direct inquiry, falsification of approval records, or entering prohibited data classes knowingly are treated as serious misconduct. Contractors and third parties in violation may have access revoked and contracts reviewed. Enforcement decisions are documented by [HR/Legal] and reported in aggregate to the [governance committee].

Proportionality isn’t softness. Employees who think a single honest mistake ends their career will hide every mistake, and hidden mistakes turn into the breaches you read about.

Section 12: Policy Review Cadence

This policy is reviewed [every six months] and after any of the following triggers: a shadow AI incident classified medium or higher, a material change in AI regulation affecting [Company], adoption of a new AI category (e.g., autonomous agents), or a significant change in the approved tool registry. The policy owner presents review outcomes to [governance committee], and material changes are communicated to all staff with [30 days] notice.

Six months is the ceiling, not the target. The AI tool market moves faster than any software category your policies have covered before.

How to Customize the Template for Your Organization

The template above is generic by design. Its value depends on how you adapt it, and three variables matter most: size, industry, and regulatory exposure.

Adapting for Company Size (SMB, Mid-Market, Enterprise)

SMBs (under ~200 employees) should collapse the structure. Merge Sections 3, 7, and 9 into a single owner, usually the IT lead, with the CEO as escalation. Cut the three-tier approval to two: “IT approves” and “IT plus outside counsel approves.” Skip CASB tooling you can’t afford and rely on SSO enforcement, a short approved-tools list, and a standing rule that free-tier AI accounts are banned for work data. The whole policy should fit on two pages.

Mid-market companies (200 to 2,000) should implement the template roughly as written, with real detection tooling and a quarterly audit. This is the size band where shadow AI grows fastest: enough employees to generate hundreds of unsanctioned tools, not yet enough governance headcount to notice.

Enterprises should extend it. Add per-business-unit registries that roll up to a central one, integrate the approval workflow into existing ITSM tooling, assign a dedicated AI governance function, and align Section 12’s review triggers with the enterprise risk committee calendar. At this scale, also add a section on AI agents and non-human identities, because agent sprawl is the next form shadow AI takes.

Adapting for Regulated Industries (Healthcare, Finance, Legal)

Healthcare organizations must treat any tool that could touch protected health information as high-tier by default and require a Business Associate Agreement before approval. Wolters Kluwer research in 2026 found that one in ten healthcare professionals has already used an unauthorized AI tool in a direct patient care context, so Section 6 should name clinical use cases explicitly rather than relying on general data-class language.

Financial services firms should map Section 9’s monitoring to existing supervision and record-keeping obligations, since prompts and outputs may qualify as business communications regulators expect you to retain. Model risk management requirements also mean high-tier approvals need documented validation, not just a security review.

Legal organizations face privilege risk on top of confidentiality risk. A prompt containing client matter details entered into an unapproved tool can arguably waive privilege. Section 6 should prohibit matter-identifiable information in any tool without explicit client consent, and Section 10 should add a privilege assessment step to the incident workflow.

Let Axipro help you build a business continuity plan that's practical, compliant, and audit-ready.

Strengthen Your Business Continuity Strategy​

Aligning with EU AI Act, GDPR, and NIST AI RMF

Three frameworks shape most shadow AI policies, and the template maps to each.

The EU AI Act matters even for shadow tools, because obligations attach to the use of AI, not just its procurement. Article 4’s AI literacy requirement, applicable since February 2025, means staff using AI need adequate training, which is impossible for tools you can’t see. Section 5’s high tier and Section 6’s prohibited use cases should mirror the Act’s prohibited and high-risk categories. Note that the Digital Omnibus proposals have shifted timelines for certain Annex III high-risk obligations toward late 2027, so keep Section 12’s regulatory trigger active rather than hardcoding dates.

Under the GDPR, an unapproved AI vendor processing personal data is a processor with no Article 28 contract, no lawful basis analysis, and no entry in your records of processing. Sections 6 and 10 carry the GDPR weight: prohibited data classes prevent the violation, and the incident workflow determines whether a 72-hour breach notification applies.

The NIST AI Risk Management Framework provides the vocabulary for the whole structure. The registry and detection controls implement the Map function, risk tiers implement Measure, the approval workflow and enforcement implement Manage, and the roles, review cadence, and governance reporting implement Govern. If you pursue ISO/IEC 42001 certification later, a shadow AI policy built this way gives you a running start on the management system’s inventory and risk assessment clauses.

Worth Knowing: IBM's 2025 Report

IBM's 2025 report found that 97% of organizations that suffered an AI-related security incident lacked basic AI access controls. Section 8 of this template, the least glamorous section, addresses the single most common failure condition in real-world AI breaches.

Why Bans Fail: The Permission Gap

Some leadership teams will read this template and ask the obvious question: why not just block everything? Because the data says blocking doesn’t work. PagerDuty’s 2026 survey of office professionals at large companies found that 66% had used AI tools at work despite believing it was against policy, and nearly half said they’d rather use AI without telling anyone than risk being told no. Microsoft and LinkedIn’s Work Trend Index put bring-your-own-AI at 78% of AI users back in 2024, before adoption accelerated further.

Call it the permission gap: the space between what employees are allowed to use and what their workload demands. When approved tools lag behind the job, employees close that distance themselves, on personal accounts and personal devices, outside of your control you own. A ban doesn’t shrink the gap. It pushes usage where you can’t see it, which is exactly where IBM measured that extra $670,000 in breach cost. The organizations with the least shadow AI aren’t the ones with the strictest bans. They’re the ones where the approved path is genuinely faster than the workaround.

 

Rolling It Out Without Killing Adoption

Sequence matters more than policy text. Start with discovery, not announcement: run your DNS, proxy, and SaaS discovery tools for two to four weeks before the policy goes live, so the initial registry reflects the tools people already depend on. Then grandfather aggressively. Take the ten most-used discovered tools, fast-track them through the approval workflow, and launch the policy with a populated registry rather than an empty one. A policy that opens by approving the tools people already rely on earns compliance. One that opens with prohibition doesn’t.

Announce with an amnesty window. Give employees 30 days to disclose the tools they use with no consequences, and route every disclosure into the request workflow. Pair the announcement with the training the EU AI Act already expects you to deliver, and keep it practical: show real examples of what may and may not go into a prompt, using your own data classes. Then measure adoption of the approved tools, not just violations. If usage of sanctioned tools is flat while detection findings climb, your approved offering is the problem, not your workforce.

Important: Don’t launch monitoring (Section 9) before the amnesty window closes. Employees who discover they were being watched during a supposed grace period will treat every future assurance from IT Security as hollow, and you’ll spend years rebuilding the trust that shadow AI governance depends on.

 

Conclusion

Shadow AI isn’t a discipline problem. It’s a governance vacuum, and vacuums get filled by whatever tool loads fastest. The twelve-section template above fills that vacuum with a registry, risk tiers, a fast approval path, real detection, and proportional enforcement, then adapts to your size, sector, and regulatory obligations. Copy it, name real owners, and populate the registry with the tools your people already use. Once asking beats hiding, the gap starts closing on its own.

Axipro Author

Picture of Pedro Dias

Pedro Dias

Pedro has been writing online for over 10 years. With experience in all things programming, cyber security, and compliance, he is our editor-in-chief at Axipro.

Blog Highlights

Explore More Articles

Two compromised versions of LiteLLM sat on PyPI for roughly 40 minutes on the morning of March 24, 2026. That window was enough to capture secrets from around 434,000 CI/CD pipeline runs across nearly 2,500 organizations, including AWS, Samsung, Cisco, Salesforce, Siemens, and Deloitte. In August, researchers at CloudSEK and Hudson Rock confirmed they had obtained the raw exfiltrated data: a 153GB archive containing 433,909 files of environment variables, cloud keys, Kubernetes secrets, and API tokens harvested live from running pipelines, as covered by Help Net Security’s reporting on the credential archive. If LiteLLM runs anywhere in your stack, or you touch any AI proxy infrastructure at all, you need answers to three things: whether you were exposed, what to rotate first, and whether the rotation you did back in March actually held. That last one matters more than it sounds, because “we rotated everything” has already burned at least one very large company. How the Breach Happened The attack didn’t start with LiteLLM. On March 19, 2026, a threat group called TeamPCP compromised the build pipeline of Trivy, a vulnerability scanner half the industry runs, and pushed a poisoned release. LiteLLM’s own CI pipeline ran Trivy, so the poisoned scanner had legitimate read access to the project’s runner environment. The attackers used that to steal LiteLLM’s PyPI publishing tokens and ship two malicious releases of their own: versions 1.82.7 and 1.82.8. KICS and the Telnyx Python SDK got hit in the same campaign. The payload design is the part worth studying. The malicious package dropped a .pth startup hook into site-packages, so the code ran the moment any Python interpreter started on the machine, whether or not anything imported LiteLLM. From there it harvested environment variables, read local credential files like .aws/credentials and .kube/config, tried to move laterally across Kubernetes clusters, and installed a systemd backdoor dressed up as a generic telemetry service. InfoQ’s coverage of the PyPI compromise put downloads of the compromised release above 40,000. For scale, LiteLLM normally gets downloaded around 3 million times a day. The exfiltration had a nasty fallback, too. According to CloudSEK, stolen data was encrypted and sent to a typosquatted domain, and when that failed, the malware created a public repository inside the victim’s own GitHub account and uploaded the loot as a release asset. Some companies were publishing their own secrets to the open internet and had no idea. Worth Knowing: The malicious code only existed in the PyPI artifacts. The GitHub source repository stayed clean the whole time, so a developer reviewing the code on GitHub saw nothing wrong. Source review isn’t artifact verification. If you don’t check that what the registry serves matches the upstream source, this class of attack is invisible to you. How to Check If You Were Exposed Three checks, from quickest to most involved. 1. Confirm whether the compromised versions ever ran The malicious versions went live on PyPI at 10:39 UTC on March 24, 2026 and got quarantined about 40 minutes later. The project’s advice: treat any install from that day before 16:00 UTC as suspect. Search your lockfiles, pip caches, SBOMs, and container image histories for 1.82.7 and 1.82.8. And check your internal artifact mirrors. An Artifactory or Nexus proxy that cached the bad release in March can keep serving it internally long after PyPI pulled it. Keep the .pth mechanism in mind when you scope this. The question isn’t “which applications import LiteLLM,” it’s “which machines had the package installed at all,” because every Python process on an infected machine triggered the payload. 2. Hunt for persistence Rotation is pointless if the attacker still has a foothold. Check developer machines, CI runners, and containers for unauthorized .pth files in site-packages and for suspicious systemd units, especially anything posing as a system telemetry service. And review activity from March 24 onward, not just the 40-minute window. Persistence is there so the access outlives the infection. Pro Tip: Don’t limit the persistence hunt to live machines. Base container images rebuilt in late March may have baked the payload into every image derived from them since. Scan your image registry for the affected LiteLLM versions and for unexpected .pth files, then trace which running workloads came from flagged images. 3. Check whether your secrets are in the dump Hudson Rock has published a domain lookup tool and is running ethical disclosures for affected organizations, and CloudSEK maintains a high-confidence victim list. Use them, but know their limits. Attribution in this dataset is genuinely hard. One dump with a siriusxm.com committer email actually traced, through its self-hosted GitLab endpoints, to AdsWizz, a SiriusXM subsidiary. And a large share of the dumps are generic pipeline configurations with no identifying domain, email, or server name at all. Absence from a victim list is not evidence of absence. If your pipelines ran the compromised versions, assume exposure no matter what a lookup tool tells you. What to Rotate, in What Order The guidance from both research teams is blunt: treat every secret the LiteLLM environment could reach as compromised. That covers secrets on disk, in memory, injected into CI jobs, and anything retrievable through instance metadata services. Work down by blast radius: Priority Credential type Why it comes first 1 Cloud IAM keys (AWS, GCP, Azure) Direct control of infrastructure, data stores, and billing. This is where attackers monetize fastest. 2 GitHub and GitLab PATs, package publishing tokens These let an attacker poison your releases and turn your company into the next link in the supply chain. 3 Kubernetes service account tokens and kubeconfigs Lateral movement across clusters was built into the payload, not a theoretical risk. 4 Database passwords and third-party API keys Dumped in plain text in the archive, often with no attribution, so nobody will warn you they leaked. 5 AI provider API keys Billing abuse, quota theft, and access to whatever data flows through your LLM routing layer. One word matters more than the rest of this article: revoke, don’t just rotate. That

The EU AI Act names recruitment AI as high-risk. Annex III explicitly lists AI systems used for recruitment, candidate selection, and employment decisions, which pulls CV screeners, video interview platforms, and assessment tools into the most demanding compliance regime the Act contains. The original compliance date for these systems was August 2, 2026. In June 2026, the EU’s Digital Omnibus moved the deadline to December 2, 2027, a 16-month extension that has led many HR and talent teams to shelve the topic entirely. That’s a mistake, for two reasons. First, one rule that directly affects recruitment technology is already in force: the ban on emotion recognition in the workplace has applied since February 2, 2025, and it catches features still shipping in some video interview products today. Second, the deferred obligations didn’t shrink. Conformity assessments, human oversight design, bias monitoring, and documentation all still arrive in full, and the practical work of auditing a recruitment stack, renegotiating vendor contracts, and training hiring teams routinely takes a year or more. Here’s what the EU AI Act actually requires of employers and vendors using recruitment tools, on the timeline that now applies. Why Recruitment Tools Are Classified as High-Risk Under the EU AI Act​ Definition of High-Risk AI Systems in Hiring​ The Act takes a list-based approach. Annex III, point 4, designates as high-risk any AI system intended for the recruitment or selection of natural persons, including placing targeted job advertisements, analyzing and filtering applications, and evaluating candidates. The same point covers AI used for decisions on promotion, termination, task allocation, and monitoring of workers, so the classification follows the tool through the entire employment lifecycle, not just the hiring funnel. The reasoning is straightforward: hiring decisions shape access to livelihoods, and algorithmic discrimination in hiring is well documented. The European Commission’s regulatory framework for AI treats employment as one of the areas where an AI error or bias causes serious harm to fundamental rights. That’s the test for the high-risk tier. Types of Recruitment Tools Affected In practice, the high-risk classification captures most of the modern recruitment stack: CV and resume screeners that rank or filter applicants, video interview platforms that score responses or delivery, psychometric and skills assessment tools that produce scores feeding a hiring decision, sourcing and matching algorithms that decide which candidates a recruiter sees, and programmatic job ad targeting systems that determine who sees a vacancy at all. If the system’s output materially influences who advances and who does not, assume high-risk until proven otherwise. Important: Emotion recognition is not high-risk in the workplace. It is prohibited. Article 5 bans AI systems that infer emotions of people in the workplace (outside narrow medical and safety cases), and that ban has applied since February 2025 with the Act’s top penalty tier attached. If your video interview vendor markets “engagement scoring” or “sentiment analysis” of candidates, that feature needs to be switched off for EU hiring now, not in 2027. Recruitment Tools That May Fall Outside High-Risk Classification Not everything in the HR stack qualifies. The Act carves out systems performing narrow procedural tasks that do not materially influence decision outcomes. An applicant tracking system that stores applications, schedules interviews, and sends templated emails is a database with a workflow, not a high-risk AI system. The same goes for tools that transcribe interviews without scoring them, deduplicate candidate records, or generate first drafts of job descriptions for a human to edit. The line is decision influence: the moment a tool ranks, scores, filters, or recommends candidates, it crosses into Annex III territory. Deployers who rely on an exemption must be able to document that assessment, so “we decided it doesn’t count” needs to exist on paper. Extraterritorial Scope: Which Employers Are Covered The Act applies to providers placing AI systems on the EU market and to deployers established in the EU, but it also reaches further: it covers providers and deployers located outside the EU where the output of the system is used in the EU. For recruitment, the consequence is blunt. A US or UK company with no EU entity that uses an AI screener to filter applicants for roles based in Berlin or Dublin, or that screens candidates located in the EU, is using the system’s output in the Union. Brexit doesn’t move UK employers out of scope when they hire into or from the EU. Providers vs. Deployers of Recruitment AI Tools The Act splits obligations between the provider (the vendor that develops the tool and places it on the market) and the deployer (the employer using it). Most employers are deployers, and deployer obligations are lighter but real. One common trap: an employer that substantially modifies a high-risk system, or puts its own name on it, can be reclassified as a provider and inherit the full provider stack. Heavy customization of a screening model, or fine-tuning it on your own hiring data, can be enough to trigger this. Key Obligations for Employers Using AI Recruitment Tools Human Oversight in Automated Hiring Decisions Deployers must assign oversight of the system to people with the competence, training, and authority to intervene. That last word matters. A recruiter who rubber-stamps whatever the ranking algorithm produces, because nobody has time to review 800 rejected CVs, doesn’t count as oversight. Regulators and courts will look at whether the human could genuinely override the system and whether they ever did. Designing review checkpoints where a person can meaningfully change the outcome, and logging when they do, is the core of compliant deployment. Transparency Requirements Toward Candidates Employers must inform workers and their representatives before putting a high-risk AI system into use at work, and candidates subjected to such a system must be told it is being used. In countries with works councils, such as Germany, this obligation lands on top of existing co-determination rights, so employee representatives may need to be consulted before the tool goes live rather than just told afterward. Burying an AI disclosure in a privacy policy paragraph is unlikely to survive scrutiny.

A green dashboard is not an audit opinion. Compliance automation platforms like Vanta, Drata, Secureframe, and Hyperproof have made SOC 2 readiness faster and cheaper, but every audit cycle produces the same pattern: controls that sat at “passing” for months come back from the auditor with exceptions or requests for re-testing. The four controls below account for a disproportionate share of those rejections, and they all fail for the same underlying reason. The tool confirmed that evidence exists. The auditor tested whether the control actually operated. This article walks through each of the four: what auditors reject, why, and how to fix the evidence before fieldwork starts. Why Compliance Tools Show “Passing” But Auditors Still Reject Controls​ The Gap Between Automated Checks and Auditor Judgment Compliance platforms run continuous control monitoring: API calls that check whether a configuration exists, a document is uploaded, or a task is marked done. That’s real value. It catches drift, keeps evidence in one place, and saves weeks of screenshot collection. An audit is a different exercise. A SOC 2 examination is an attestation performed by a CPA firm under AICPA standards, and the auditor’s job is to form an independent opinion on whether your controls met the Trust Services Criteria. That opinion rests on professional judgment, not on whether an API integration returned a 200 response. What “Passing” Actually Means in Your Compliance Dashboard​ When a control shows “passing,” the platform is telling you one narrow thing: at the moment of the last scan, an automated test found the artifact or setting it was programmed to look for: MFA enforced in the identity provider, a policy document uploaded, a training campaign sitting at 100%. The test says nothing about whether the underlying process ran the way your control narrative claims it did, or whether it ran that way across the whole audit period. How Auditors Evaluate Controls Beyond the Checkbox Auditors test two dimensions. Design effectiveness asks whether the control, as described, would meet the criterion if it worked as intended. Operating effectiveness, the core of a SOC 2 Type 2 report, asks whether it actually did throughout the audit period. To answer that, the auditor pulls a population (every access review, every change, every new hire in the period), selects a sample, and inspects the evidence item by item. A dashboard status feeds into that process. It doesn’t replace it. Insider Note: Auditors increasingly ask for evidence outside the compliance platform precisely because they know what the platform auto-collects. If every artifact you produce comes from the same tool export, expect the auditor to independently pull the population from the source system and compare. Discrepancies between the two are one of the fastest routes to an exception. Control #1: Access Reviews That Automation Marks Complete but Auditors Reject Why Auditors Reject Automated Access Review Evidence​ User access reviews sit under the logical access criteria (CC6.1 through CC6.3), and they are the single most common source of audit exceptions we see. The typical failure: the platform generated a user list, someone clicked “complete,” and the dashboard turned green. The auditor then asks a simple question the evidence can’t answer: what did the reviewer actually decide? The Missing Element: Documented Reviewer Judgment​ An access review is a judgment control. Someone with knowledge of the system must look at each account and confirm the access is still appropriate for the person’s role. A timestamped task closure proves the task was closed. It doesn’t prove anyone assessed anything, and an “approve all” review completed in ninety seconds gets exactly the skepticism it deserves. What Auditors Actually Want to See in Access Review Evidence Auditors look for four things: The full population of accounts at the time of review (including service accounts and admin roles), Evidence of who reviewed it and when, explicit dispositions per account or group (retain, modify, revoke), and Proof that flagged access was actually removed. That last item, the deprovisioning ticket showing revocation within a defined window, is the piece most companies can’t produce. How to Fix Your Access Review Control Before the Audit​ Assign a named control owner per in-scope system, run reviews quarterly, and require reviewers to record a disposition for every line, not a blanket approval. When access is revoked, link the removal ticket to the review record. If a quarter was missed, don’t backfill it. Document it honestly and show the remediation, because auditors treat fabricated retroactive evidence far more severely than a disclosed gap. Control #2: Change Management Approvals That Pass Automated Scans​ Why Ticket Closure Isn’t Proof of Approval​ Change management (CC8.1) automation typically verifies that production changes link to a ticket and the ticket is closed. Auditors test something stricter: that each sampled change was approved by an authorized person before deployment. An approval added after the merge, or a ticket closed by the same engineer who wrote the code, fails that test even though every automated check came back green. The Segregation of Duties Problem Automation Misses Segregation of duties is the requirement that no single person can develop, approve, and deploy the same change. NIST’s SP 800-53 control catalog treats it as a foundational access control principle, and SOC 2 auditors apply the same logic. Small engineering teams trip on this constantly. Self-approved pull requests, admins who can bypass branch protection, direct pushes to main: a scanner sees “changes with tickets” while an auditor sees SoD violations. Emergency Changes and Retroactive Approvals: Common Rejection Triggers​ Every audit period contains hotfixes. Auditors don’t reject emergency changes. They reject emergency changes with no documented post-hoc review. If your policy says urgent changes get retroactive approval within two business days, the auditor will sample your emergency changes and check exactly that. No policy, or a policy nobody followed, produces an exception. Rebuilding Change Management Evidence Auditors Will Accept​ Enforce the control technically: branch protection requiring at least one independent reviewer, no admin bypass, and deploy pipelines that only run from protected branches. Then write the emergency change procedure down and generate the review artifact every time it fires.