EU AI Act Compliance

How Cheqpoint satisfies the human oversight and transparency requirements for high-risk AI systems under Regulation (EU) 2024/1689. Includes article-by-article mapping, implementation guide, industry playbooks, and how to generate a regulator-ready compliance report.

What is the EU AI Act and when does it apply?

The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024. It is the world's first comprehensive AI law and applies to any organisation that develops, imports, or deploys AI systems in the EU — regardless of where the organisation is headquartered.

The Act uses a risk-tiered approach. Most AI systems are low-risk and face minimal obligations. However, AI systems listed in Annex III are classified as high-risk and carry substantial compliance requirements — including mandatory human oversight (Article 14), transparency (Article 13), risk management (Article 9), and record-keeping (Article 12).

Key enforcement timeline: Prohibited AI practices: 2 Feb 2025 · GPAI model obligations: 2 Aug 2025 · High-risk AI system obligations (Annex III): 2 Aug 2026 · All remaining obligations: 2 Aug 2027. Start your compliance programme now — the 2026 deadline for high-risk AI is already approaching for deployers building today.

Who counts as a “deployer” vs a “provider”?

The Act distinguishes between providers (organisations that develop and place AI systems on the market) and deployers (organisations that use AI systems in their own operations). Most teams using Cheqpoint are deployers: they are running AI agents built on third-party models (OpenAI, Anthropic, Google) and deploying them against real business operations.

Deployers of high-risk AI are required — under Article 14 — to implement human oversight measures and maintain records showing that oversight operated. Cheqpoint is the technical implementation of that requirement.

Annex III — Is your AI system high-risk?

The following categories from Annex III are most relevant to teams using AI agents for real-world business operations:

Annex III CategoryExample AI actions
5(b) — Creditworthiness / credit scoringCredit limit changes, loan approvals, affordability assessments
5(c) — Life & health insurance pricingAI-driven underwriting decisions, policy acceptance, premium setting
4(a) — Recruitment & HR decisionsCandidate shortlisting, performance scoring, promotion recommendations
2(a) — Safety-critical infrastructureAI actions on water, energy, transport, or digital infrastructure systems
3 — Education & vocational trainingAI-generated assessments, access decisions, grading
6 — Law enforcement & evidenceAI-flagged fraud, risk profiling, investigation prioritisation
7 — Migration, asylum, border controlAutomated eligibility decisions, document verification
Important: Cheqpoint does not turn a low-risk chatbot into a high-risk AI system. It is the oversight mechanism you add when your AI agents start touching decisions in one of the categories above. If your AI agent can issue a refund, change a credit limit, or modify a customer account, you likely need human oversight — and Cheqpoint provides it.

Article-by-article compliance map

The table below maps each relevant EU AI Act article to the specific Cheqpoint feature that satisfies it, with links to where you configure it in the dashboard.

ArticleRequirementCheqpoint featureWhere to configure
Art. 9Risk management system — identify, analyse, and mitigate risks throughout the AI system lifecycleRisk scoring per request (0–1 float), automatic routing by risk level, auto-approval rules for low-risk actionsRules → Risk & Policies
Art. 12Record-keeping — automatically log events during operation of the high-risk AI systemTamper-evident audit log of every request, decision, actor, payload, and timestamp. CSV export available.Audit Log
Art. 13Transparency — enable deployers and users to understand the AI system's outputs and make informed decisionsRequest detail shows action type, AI confidence score, risk score, full payload, and reason code. Reviewers see both Simple and Technical views.Approval Inbox (Technical view)
Art. 14Human oversight — allow natural persons to monitor, understand, override, and intervene in AI outputsApprove / Decline / Modify on every request. Kill switch to pause any AI Assistant instantly. No action executes without a human decision (unless explicitly whitelisted by policy).Approval Inbox, AI Assistants
Art. 14(4)Ability to decide not to use the AI system or disregard its outputReviewers can decline any request regardless of AI recommendation or confidence score. Modified decisions create a separate audit entry.Approval Inbox → Decline / Modify
Art. 17Quality management system — policies, procedures, and oversight structures for AI governanceRules engine, Review Groups (routing by team/expertise), escalation policies, SLA targets, and notification channelsRules, Review Groups, Settings
Art. 72Post-market monitoring — serious incident reporting and ongoing performance observationAnalytics dashboard (approval rates, response times, agent volumes, SLA compliance), anomaly visibility via audit logAnalytics, Audit Log, Compliance Report

Article 14 — Human Oversight in detail

“High-risk AI systems shall be designed and developed in such a way, including with appropriate human-machine interface tools, that they can be effectively overseen by natural persons during the period in which the AI system is in use.”
— Article 14(1), EU AI Act (Regulation (EU) 2024/1689)

Article 14 breaks down into four specific capabilities that deployers must implement. Here is how Cheqpoint satisfies each one:

14(4)(a)
Monitor operation
Every AI request appears in the Approval Inbox in real time. Live badge counts and SLA timers ensure no request goes unreviewed. Analytics tracks approval rates, agent volumes, and response times over time.
14(4)(b)
Detect and address deviation
Risk scoring (0–1) flags unusual or high-risk actions. Auto-routing rules send high-risk requests to specialist reviewers. Reviewers can reject or modify any action that deviates from expected behaviour.
14(4)(c)
Override or intervene
Reviewers can Approve, Decline, or Modify any AI request before it executes. The kill switch on any AI Assistant halts all future requests instantly — no code change required.
14(4)(d)
Prevent over-reliance
AI confidence scores and risk scores are shown to reviewers but decisions are never auto-submitted. Every decision requires a deliberate human action. Reviewers must select a reason code, then confirm — two clicks minimum.

Implementation guide — 5 steps to Article 14 compliance

Follow these steps to configure Cheqpoint as your Article 14 compliance layer. This takes approximately 30–60 minutes for a development team already running AI agents.

1
Register your AI Assistants
Go to AI Assistants → Add new. Give each agent a name that reflects its role (e.g. “Credit Decision Bot”, “Refund Agent”). Define its allowed action types and assign a Connection Key. The Connection Key is what authenticates the agent when it submits requests.
2
Install the SDK and route actions through Cheqpoint
Add the SDK to your AI agent code. Every action that could affect a customer, financial record, or data asset should submit an approval request before executing. The SDK call blocks execution until a human decides.
3
Configure risk routing rules
Go to Rules → Auto-decision. Set high-risk actions (riskScore ≥ 0.7) to always require human review. Set low-risk, routine actions (riskScore < 0.3) to auto-approve. Assign high-risk actions to a dedicated compliance reviewer group.
4
Set up your compliance team
Go to Settings → Team. Add your compliance officers, operations managers, or risk reviewers. Assign appropriate roles (OWNER, ADMIN, REVIEWER). Configure Slack or email notifications so reviewers are alerted immediately when a new request arrives.
5
Generate your compliance report
Go to Compliance Report in the sidebar. Select your reporting period (30d, 90d, 6m, or 1y). The report auto-generates a dated Article 14 compliance document showing: total requests reviewed, human vs auto-decided, SLA compliance, risk breakdown, per-reviewer activity. Click Print / Export PDF to share with regulators or your internal audit team.

Code example — Python SDK (fintech: credit decision)

python
from cheqpoint import Cheqpoint

cp = Cheqpoint(api_key="cq_live_...")

# Article 14: execution is blocked until a human reviewer decides
result = cp.request_approval(
    action="credit_limit_change",
    summary="Increase credit limit: £5,000 → £8,000 for customer #C-00412",
    details={
        "customer_id": "C-00412",
        "current_limit": 5000,
        "requested_limit": 8000,
        "credit_score": 742,
        "affordability_ratio": 0.31,
        "ai_recommendation": "approve",
    },
    risk_score=0.68,           # 0.0 (low) – 1.0 (high)
    timeout=300,               # seconds to wait for human decision
)

if result.approved:
    apply_credit_increase(result.modified_details or result.details)
elif result.declined:
    notify_customer_declined(result.decision_note)

Code example — Node.js SDK (fintech: refund approval)

typescript
import { Cheqpoint } from "@cheqpoint/sdk";

const cp = new Cheqpoint({ apiKey: "cq_live_..." });

// Article 14: human must review before refund executes
const result = await cp.requestApproval({
  action: "process_refund",
  summary: "Refund £1,240 to customer Sarah M. — Order #ORD-99812",
  details: {
    orderId: "ORD-99812",
    customerId: "cust_Sm42",
    refundAmount: 1240,
    currency: "GBP",
    reason: "Item not received",
    agentConfidence: 0.91,
  },
  riskScore: 0.55,
  timeout: 180,
});

if (result.approved) {
  await stripe.refunds.create({ payment_intent: order.paymentIntentId });
}

Industry playbooks

Below are the key action types for each regulated industry that should route through Cheqpoint to satisfy Article 14. Use these as a starting point when configuring your AI Assistants.

Fintech & BankingAnnex III item 5
  • Credit limit changes and loan approvals
  • Refund and chargeback decisions above £/€ threshold
  • Account access changes (password resets, MFA bypass)
  • AML or fraud flag overrides
  • Transaction limit modifications
  • Account closures or suspensions
  • Bulk data exports of customer financial data
Setup tip: Route all actions with riskScore ≥ 0.5 to your Compliance team review group. Auto-approve routine low-value refunds (< £50) with riskScore < 0.2.
InsuranceAnnex III item 5(c)
  • Claims approval and payout decisions
  • Underwriting acceptance or rejection
  • Premium adjustment decisions
  • Policy cancellations or non-renewals
  • Fraud flag decisions on claims
  • High-value claim escalations
Setup tip: Use Review Groups to route motor claims to your motor underwriters and health claims to your medical assessors.
Healthcare & MedTechAnnex III item 2(a) (safety-critical systems)
  • Clinical decision support recommendations
  • Prescription generation or modification suggestions
  • Patient record access by AI agents
  • Treatment pathway recommendations
  • Triage priority assignments
  • Diagnostic output requiring clinical sign-off
Setup tip: Configure a mandatory 2-approval workflow for any AI recommendation that could affect treatment. Enable SLA timers to track response times for clinical governance.
HR Tech & RecruitmentAnnex III item 4
  • Candidate shortlisting and screening decisions
  • Automated interview scoring
  • Salary band recommendations
  • Performance review flags
  • Promotion or termination recommendations
  • Background check processing
Setup tip: Require senior HR manager approval for any candidate rejection generated by AI. Log decision reasons for GDPR data subject access requests.
Legal Tech & RegTechAnnex III item 6
  • Contract execution decisions
  • Regulatory filing submissions
  • Compliance flag resolutions
  • Sanctions screening overrides
  • Data subject request responses
  • Litigation hold actions
Setup tip: Use the Audit Log CSV export to produce evidence packs for regulators. The Compliance Report satisfies most regulator requests for human oversight evidence.
E-commerce & RetailGeneral deployer obligations
  • Large refund decisions (set your own threshold)
  • Account fraud suspension decisions
  • Bulk promotional pricing changes
  • Supplier payment approvals
  • Inventory write-offs
  • Customer data deletion requests
Setup tip: Not all e-commerce AI is high-risk under the Act. But process-critical decisions — refunds, suspensions, payments — benefit from an audit trail regardless of regulatory obligation.

The Compliance Report — evidence for regulators

When a regulator, auditor, or procurement team asks you to demonstrate Article 14 compliance, Cheqpoint generates the evidence automatically. Go to Compliance Report in the dashboard sidebar.

The report shows, for any selected period:

  • Total AI requests submitted for review (excluding test/sandbox traffic)
  • Human decisions vs auto-approved (within policy rules)
  • Approved / Declined / Modified breakdown with percentages
  • SLA compliance rate (decisions within target response time)
  • Average, P50, and P95 response times
  • Risk profile (high / medium / low breakdown)
  • Per-AI-Assistant activity and action types
  • Per-reviewer activity and individual response times
  • Auto-generated Article 14 compliance statement naming your workspace, period, and reviewer count
  • Document reference number (CHQ-XXXXXX-YYYYMMDD) for your audit file

Click Print / Export PDF to produce a print-optimised document you can share with internal audit, external auditors, or directly with your regulator. The sidebar and controls are hidden automatically in the printed output.

Which plan do I need? The Compliance Report is available on all paid plans (Starter, Growth, Business, Enterprise). Growth and above include unlimited audit log retention and multi-reviewer workflows — both recommended for regulated use cases.

Pre-launch compliance checklist

Use this checklist before going live with AI agents in a regulated context:

AI Assistants
All AI agents registered with a name and Connection Key
Allowed action types defined
Risk thresholds configured
Rules & Routing
High-risk actions routed to specialist review group
Low-risk, policy-compliant actions auto-approved
SLA targets set (recommended: 1 hour for high-risk)
Team & Access
Compliance reviewers added with REVIEWER or ADMIN role
Notification channel connected (Slack or email)
No shared login credentials — individual accounts only
Security
2FA enabled for all reviewers (Settings → Security)
IP allowlist configured if required (Enterprise)
Data retention policy set to match your record-keeping obligations
Evidence
First compliance report generated and reviewed
Report saved to your compliance file
Audit log CSV exported for baseline period

Frequently asked questions

Does using Cheqpoint guarantee EU AI Act compliance?
Cheqpoint implements the technical requirements of Article 14 (human oversight) and Article 12 (record-keeping). It does not guarantee legal compliance — that depends on your overall governance programme, how you classify your AI systems, and your organisation's specific obligations. We recommend working with a legal advisor familiar with the EU AI Act alongside using Cheqpoint.
Does connecting my AI agent to Cheqpoint make it a high-risk AI system?
No. The risk classification is based on the type of AI system and its intended use (Annex III), not on whether it uses an approval layer. Cheqpoint is the oversight mechanism — adding it does not change your AI system's risk classification.
What about Article 13 (transparency)?
Cheqpoint's request detail view shows the action type (human label), the AI's confidence score, the risk score, the full payload, and the decision reason code. This means your human reviewers have the information they need to understand what the AI is recommending and why — which is the substance of the transparency requirement.
What plan do I need for regulated use?
Growth is the minimum recommended plan for regulated use cases — it includes multi-reviewer workflows, full audit log access, the compliance report, and SLA tracking. Business adds compliance tags, longer default retention, and SAML SSO. Enterprise adds IP allowlist, custom data retention, and dedicated support.
Can we use Cheqpoint for both low-risk and high-risk AI actions in the same workspace?
Yes. Configure auto-approval rules to pass through low-risk, routine actions automatically, and require human review only for actions above your risk threshold. The audit log records both, so you have a complete picture of all AI actions — not just the high-risk ones.
How do we respond to a regulator asking for proof of human oversight?
Go to Compliance Report, set the reporting period to match the period the regulator is asking about, and click Print / Export PDF. The report includes a compliance statement, a document reference number, and detailed metrics on every human decision made during that period. This is the starting point for most regulator conversations.
Does Cheqpoint store the AI action payloads? What are the data implications?
Yes — the full request payload is stored and shown to reviewers. This is necessary for meaningful human oversight. For sensitive data (PII, financial data), we recommend sending a structured summary rather than raw data where possible. Data is encrypted at rest (AES-256), filtered by workspaceId, and subject to your configured data retention policy. See our DPA at /legal/dpa for details.

Next steps