EU AI Act for Small Teams: A Compliance Guide for 2026
On 2 August 2026, most of the EU AI Act becomes applicable. A company using prohibited AI can face a maximum administrative fine of €35 million or 7% of worldwide annual turnover, whichever is higher under the Act’s main enforcement tier. That may sound remote for a small team, but the regulatory work is already practical: identify your role, understand how each system is used, and preserve evidence. Formal company size does not create a blanket exemption. A defensible compliance program can begin with a spreadsheet, clear ownership, vendor review, staff training, and controls proportionate to actual risk. This guide explains where to start and what to document before the deadline.
Why This Matters in 2026
The EU AI Act, Regulation (EU) 2024/1689, applies in stages rather than as a single compliance date. Understanding the specific timeline, enforcement exposure, operational obligations, and documentation requirements is now essential for organizations deploying AI systems.
First, 2 August 2026 is the central deadline. Most provisions become applicable then, including transparency duties for certain AI interactions and most high-risk systems listed in Annex III. High-risk AI embedded in regulated products, such as certain medical devices, machinery, and safety components, has a later 2 August 2027 deadline.
Second, enforcement exposure is substantial. Besides the €35 million or 7% maximum for prohibited practices, other breaches can lead to fines of up to €15 million or 3% of worldwide turnover. Incorrect, incomplete, or misleading information can produce penalties of up to €7.5 million or 1%. These are statutory maxima rather than automatic assessments, but they make documentation and governance relevant to board-level decisions.
Third, compliance increasingly affects ordinary business software. A general-purpose chatbot may have limited obligations, while the same technology used to rank job applicants can fall within high-risk employment rules. Small companies are also unlikely to own the underlying models, yet they remain responsible for how their teams deploy those models.
Finally, the evidence baseline has already moved. The AI literacy duty has applied since 2 February 2025. By 2026, a defensible position should therefore include role-specific training, attendance records, assessment materials, and a process for adding new tools—not merely a generic code-of-conduct acknowledgment.
The EU’s SME definition generally covers an organization with fewer than 50 employees and annual turnover or balance-sheet total of no more than €2 million. SME status can support proportionate documentation and simpler procedures, but it does not remove substantive AI risk-management duties.
Key Concepts, Explained Simply
1. Provider and deployer
Your legal role determines much of your obligation.
A provider develops an AI system or model and places it on the market or makes it available under its own name. A company building an applicant-ranking product, branded chatbot, or custom model is generally a provider.
A deployer uses an AI system under its authority. Examples include an HR team using a vendor’s recruitment model or a support department using an AI-generated ticket classification service.
You can be both. If a small company purchases a model, adds a proprietary interface, substantially modifies its behavior, changes its intended purpose, or markets the resulting application under its own brand, it may become the provider of that downstream system.
2. Four practical risk tiers
The Act’s categories are often simplified into four tiers. The prohibited category covers practices such as social scoring, manipulative techniques that materially distort behavior, untargeted scraping of facial images, and most prohibited workplace emotion-recognition uses. The high-risk category includes systems used for recruitment, education access or assessment, creditworthiness, certain insurance pricing, essential public services, migration, and justice-related decisions. The transparency category requires notice or labeling in situations such as direct chatbot interaction, AI-generated public-interest text, and deepfakes. Finally, systems outside these categories are often described as minimal-risk, though normal privacy, cybersecurity, consumer, and employment duties still apply.
Risk follows the actual use case. The same underlying model can be minimal-risk for proofreading and high-risk when used to screen candidates.
3. GPAI is not the same as high-risk
A general-purpose AI model, or GPAI, is a broadly capable foundation model that can integrate into many downstream systems. GPAI obligations apply mainly to the organization that develops and releases the model.
A company using a GPAI through an API is not automatically subject to every model-provider duty. However, its application of that model may still be high-risk. GPAI classification and use-case classification are separate analyses.
4. Compliance evidence
An AI compliance requirement is only manageable if you can demonstrate that it operates. Evidence can include a system inventory, a written risk classification, an owner, an approved purpose, vendor documentation, training records, human-review procedures, monitoring results, and incident records.
The objective is to avoid generating a large amount of polished but unsupported documentation.
A Practical EU AI Act Workflow
Step 1: Build a complete AI system inventory
Start with every model, API, AI feature, and agent—not only custom machine-learning projects.
Include tools such as Microsoft 365 Copilot, ChatGPT Enterprise, Claude, GitHub Copilot, customer-support automation, HR screening software, and internally built models. A team evaluating best no-code AI agent builders 2026 should classify a support agent, sales agent, or internal workflow agent in the same inventory.
For each system, record the following:
- Business owner and technical owner
- Provider, model, and deployment method
- Intended purpose
- Users and affected individuals
- Countries in which it is used
- Personal or confidential information processed
- Whether it influences employment, credit, education, or safety decisions
- Human-review arrangements
- Logging and monitoring status
- Review date
A lightweight record can begin in YAML:
system_id: applicant-ranking-v1
status: active
company_role: deployer
intended_purpose: rank_job_applications
affected_people: job_candidates
eu_exposure: yes
personal_data: cv_and_application_details
decision_impact: employment
human_review: qualified_recruiter
final_review_required: true
vendor_documentation_received: false
next_review: 2026-09-30
Credo AI or OneTrust can provide a governed registry. For a smaller team, a locked spreadsheet with controlled columns, access permissions, and version history may be sufficient.
Step 2: Classify each use case and record the reasoning
Review systems in a consistent order. First, confirm geographic scope. The Act can apply when a non-EU provider or deployer’s system output is used in the EU. Next, check prohibited practices. If a use involves one of Article 5’s prohibited practices, stop and obtain specialist advice rather than treating it as an ordinary compliance gap.
Third, evaluate Annex III and product-safety high-risk uses. Document the intended purpose and whether the system materially influences a legally or similarly significant decision. Fourth, check transparency obligations. Determine whether users must be told they are interacting with AI or whether generated or manipulated content requires a label. Finally, assess remaining privacy and security controls.
Record the conclusion, rationale, decision-maker, evidence, and review date. An unexplained spreadsheet label such as “medium risk” will not withstand scrutiny.
Examples of classification outcomes include a chatbot answering product-support questions, which may require interaction disclosure but is not automatically high-risk. An internal tool drafting meeting summaries would generally be minimal-risk, although confidentiality rules still apply. An AI system ranking candidates for recruitment would normally be high-risk. Finally, a tool generating synthetic images for public advertising may trigger provider or deployer transparency obligations.
Step 3: Assign roles and verify vendor support
For every material system, write a short role memo. State whether your company is a provider, deployer, or both, and identify which Article 26 deployer duties apply.
Request a package from the supplier covering intended use, limitations, instructions, input requirements, accuracy and robustness information, cybersecurity controls, human-oversight measures, incident contacts, and contractual allocation of responsibilities.
Review contracts to confirm whether the supplier will notify you of incidents or material changes. You should also establish whether logs and audit evidence are accessible. Additionally, check if model or system changes will be communicated, whether you can suspend processing, and who is responsible for data deletion and regulatory cooperation.
Tool selection should follow the workflow, not replace it. Credo AI is purpose-built around AI governance workflows, use-case assessment, and system visibility. OneTrust combines AI governance with broader privacy and third-party-risk capabilities. Vanta is strongest when the immediate objective is automated evidence collection, policies, training, and security-control monitoring. FairNow or ModelOp may fit narrower assessment or operational-monitoring needs. Microsoft Purview can be practical for a Microsoft-centered environment, while IBM watsonx.governance is more relevant to larger IBM-centered estates.
These products support documentation and workflows, but none makes the legal classification decision for you.
Step 4: Build proportionate controls around actual risk
For minimal-risk tools, proportionate controls may include approved-use rules, data restrictions, standard terms, training, and basic monitoring.
High-risk deployments require a substantially deeper evidence trail. Confirm that the provider supplies instructions for use, technical documentation, and conformity evidence. Establish the intended purpose and assess whether input data is relevant and sufficiently representative. Define measurable performance thresholds and establish a qualified human reviewer with authority to disregard, modify, or stop the output. You must also maintain monitoring and record-keeping. Document the process for worker or customer notification, and conduct a DPIA where personal-data processing requires one. Additionally, complete a fundamental-rights impact assessment if Article 27 applies to your organization or use case.
Article 26 generally requires deployers to retain high-risk-system logs under their control for at least six months, where possible and unless applicable law requires a different period. Configure logging before launch rather than trying to reconstruct events later.
Meaningful human oversight is critical. A recruiter who simply accepts every ranking is not meaningful oversight. The reviewer should have authority, time, training, and information sufficient to challenge the result.
Step 5: Monitor, report, and close readiness gaps
Create an evidence folder for each system rather than one global compliance archive. It should contain the inventory record, classification memo, vendor evidence, DPIA or FRIA where applicable, training records, acceptance tests, monitoring reports, incidents, and approvals for material changes.
Set an internal freeze date at least 8–12 weeks before 2 August 2026. Assign an owner and deadline to every missing artifact, then prioritize systems that influence people’s employment, education, credit, health, or access to essential services.
Monitoring should cover unusual input patterns, performance drift, manual overrides, complaints, security events, and changes to models or intended purposes. Provider incident rules can require a serious-incident report within 15 days of awareness, but the clock can be as short as two days for a widespread infringement or serious impact. Internal escalation must therefore be faster than the statutory deadline.
Common Mistakes & How to Avoid Them
Mistake 1: Treating “small” as exempt
Being below the SME threshold does not remove the Act. It may simplify documentation and support proportionate conformity processes, but prohibited systems remain prohibited and high-risk duties remain relevant.
Fix: Maintain the same decision-making process, but use a simplified template, proportionate evidence, and available SME support.
Mistake 2: Assuming vendor compliance transfers responsibility
A vendor may be the provider of the underlying model while your company remains the deployer. “The supplier says it is compliant” is not a substitute for reviewing instructions and monitoring your use.
Fix: Obtain contractual assurances and supplier documentation, map responsibilities, and retain evidence that your deployment follows the stated conditions.
Mistake 3: Classifying the model instead of the use
General-purpose models are not automatically high-risk. The risk emerges from what the system does in context.
Fix: Assess each business purpose separately, especially recruitment, education, credit, insurance, public-service, biometric, and safety-related uses.
Mistake 4: Buying software before defining ownership
A platform can produce an attractive dashboard while leaving classifications, purposes, and decisions unsupported.
Fix: Establish owners, approved uses, and minimum records first. Automate the process only after the team knows what “complete” looks like.
Cost & Effort Reality Check
A small company with no high-risk use case can often establish a baseline in 40–80 internal hours, followed by roughly three to six hours per month for updates, training, vendor changes, and monitoring. Much of the ongoing effort should be assigned to an existing operations, product, security, or legal owner rather than creating a dedicated compliance department.
External support varies by scope. In 2026, specialist AI Act consultants may charge approximately €1,500–€4,000 per day. A narrow classification or vendor-review engagement can be affordable; a full high-risk conformity program can require engineering, legal review, evaluation work, and external auditors.
Credo AI, Vanta, and OneTrust generally use sales-led pricing rather than a universal small-team list price. Useful first-year planning envelopes are approximately €10,000–€30,000 for a narrowly scoped Vanta evidence deployment, €20,000–€50,000 for a Credo AI deployment, and €25,000–€75,000 for selected OneTrust modules. These are procurement estimates, not official quotes, and should include ordinary implementation but exclude major model remediation.
For a company with no complex risks, a spreadsheet, written policies, staff training, and targeted consulting will cost much less. Conversely, remediating training data, implementing human review, adding logs, or redesigning a high-risk decision system can cost more than the governance software itself.
Getting Started This Week
1. Assign ownership and freeze unclassified high-risk uses
Name one accountable compliance owner and one business owner for each material AI system. If a tool influences hiring, credit, education, safety, or essential services, pause unapproved automated decisions until its role and controls are understood.
2. Create and classify the inventory
List every AI-enabled product and vendor. Complete the scope, purpose, data, affected-person, geographic, and decision-impact fields. Record prohibited, high-risk, transparency, or minimal-risk conclusions with a short rationale.
3. Build the minimum evidence pack
Issue AI-use and acceptable-use policies. Assign training to each user group, request vendor documentation, configure human review and logging where relevant, and establish an incident mailbox or escalation route. Book a focused legal review for any ambiguous high-risk use rather than trying to resolve complex classification internally.
FAQ
Does the EU AI Act apply to a five-person company?
Yes. There is no blanket exemption based on headcount. The Act can apply to companies established in the EU and to non-EU organizations whose AI-system output is used in the EU, depending on their provider or deployer role.
Is ChatGPT or Claude automatically high-risk?
No. A general-purpose model is not high-risk solely because of its capabilities. An application using it to rank job applications or determine access to education can create a high-risk use even if the model provider handles the underlying model obligations.
Must a small team buy Credo AI, Vanta, or OneTrust?
No. A controlled spreadsheet, written procedures, training records, supplier reviews, and technical logs may be sufficient for lower-risk deployments. Software is useful when you need automated evidence collection, integrations, recurring assessments, or a formal audit trail.
What happens on 2 August 2026?
Most remaining AI Act provisions become applicable, including many high-risk and transparency obligations. High-risk AI that is a safety component of regulated products follows the later 2 August 2027 deadline, so teams should confirm the current timetable and implementing guidance for each use case.
Small teams do not need an elaborate governance department to begin. They need a defensible inventory, accurate classifications, named owners, documented vendor responsibilities, trained users, and evidence that important risks are controlled. That is the practical starting point for EU AI Act compliance in 2026.
Disclosure: This article may contain affiliate links. We may earn a commission at no extra cost to you.
Last Updated: October 31, 2026 | Specs and prices subject to change. Please verify current pricing on Amazon.