AI QA Security Checklist for IT, Legal and Procurement

An AI QA security checklist is the list of checks IT, legal and procurement run before an AI quality assurance tool may read your support conversations. It covers the vendor’s audits and certifications, the data processing agreement and sub-processors, whether your data trains AI models, how agents are evaluated under GDPR and the EU AI Act, and the exit terms.

In short

• Use the checklist by owner: IT takes the security checks, legal takes data protection and the AI rules, procurement takes the contract and the document pack.
• AI QA gets a closer review than most support tools because it reads every conversation and evaluates employees, often with AI model providers as sub-processors.
• IT checks cover independent audits, encryption, access, logging, testing and incident response.
• Data protection checks follow the GDPR articles that apply: the processor contract, sub-processors, transfers, retention, impact assessments and automated decisions.
• AI checks cover model training, model providers, human review and where the tool sits under the EU AI Act, whose rules for high-risk uses apply from 2 December 2027.

This checklist is written for teams buying QA software that scores support conversations with AI. The law references were checked against the official texts on 4 November 2026. They are a starting point for your own legal advice, not a substitute for it.

How to use this AI QA security checklist

Give each part of the checklist to the team that owns it, send the vendor the questions in one go, and ask for documents rather than answers in a form. Most security reviews stall because questions arrive in three waves from three teams.

Mark each check as met, met with a condition, or not met. A “not met” on checks 1, 9, 10 or 17 should stop the purchase until it is resolved. If you are also comparing features, read our comparison of customer service QA software alongside this list.

Why AI QA software needs a closer security review

AI QA software needs a closer review because it reads every customer conversation, scores the employees who handled them, and often sends text to AI model providers. That puts customer data, employee data and automated evaluation in one tool.

Three things set it apart from a typical helpdesk add-on:

• Volume: a manual program reads a sample, while an AI tool may process every conversation, including card numbers, health details or complaints a customer typed in.
• Employees as data subjects: QA scores are personal data about your agents and can feed decisions about pay, promotion or dismissal. That brings in the GDPR rules on automated decisions and, under the EU AI Act, the category of AI used to evaluate workers.
• Model providers: the AI that writes scores and comments is often hosted by a third party, which becomes a sub-processor of your data.

IT security checks

The IT checks confirm that an independent auditor has tested the vendor’s controls and that the basics of encryption, access, logging and incident response are in place. Ask for the reports themselves, under NDA, not a badge on a website.

1. Independent audit report. Ask for the latest SOC 2 Type II report and check its period and scope. SOC 2 reports, defined by the AICPA, cover controls relevant to security, availability, processing integrity, confidentiality and privacy. A Type II report tests whether controls worked over a period, not only whether they were designed.
2. ISO 27001 certificate. Check the certificate’s scope covers the product you are buying, its expiry date and the certification body.
3. Penetration testing. Ask how often a third party tests the production application and how critical findings are tracked to resolution.
4. Encryption. Confirm data is encrypted in transit and at rest, and who manages the keys.
5. Access control. Ask how vendor staff access production data, whether access is need-to-know and logged, and how your own users authenticate.
6. Audit logs. Check whether you can see who in your company viewed or changed scores and settings, and on which plan.
7. Incident response. Ask for the incident response plan and how fast the vendor tells you about an incident affecting your data.
8. Continuity. Ask for the business continuity and disaster recovery plan, and whether production data is kept out of development and test environments.

Data protection and GDPR checks

The GDPR checks confirm the vendor acts as your processor under a contract that meets Article 28, that every sub-processor and transfer is known, and that you have done your own homework on employees. Read the clauses against the regulation, not against the vendor’s summary.

9. A data processing agreement that meets Article 28(3). Article 28(3) GDPR lists what the contract must say: processing only on your documented instructions, confidentiality of staff, the security measures of Article 32, rules for sub-processors, help with data subject rights and with Articles 32 to 36, deletion or return at the end, and audits. Check the agreement covers the AI features, not only the core product.
10. Sub-processors. Under Article 28(2), a processor needs your prior written authorization to use another processor, and with a general authorization it must tell you about changes so you can object. Ask for the list, with each sub-processor’s role and data location, including AI model hosts.
11. Transfers outside the EU. Where a sub-processor is outside the EEA, ask for the transfer basis under Chapter V of the GDPR: an adequacy decision (Article 45) or safeguards such as standard contractual clauses (Article 46).
12. Retention and deletion. Ask how long conversations, transcripts and scores are kept, whether you can set the period, and how data is returned or deleted after the contract, as Article 28(3)(g) requires.
13. Data minimization. Ask whether you can exclude queues, fields or channels from scoring, so the tool only processes what QA needs (Article 5(1)(c)).
14. Breach notification. Under Article 33(2), a processor must notify you “without undue delay” after becoming aware of a breach. Check the contract gives a concrete commitment, because your own 72-hour clock to notify the supervisory authority under Article 33(1) runs from when you become aware.
15. Your impact assessment. Article 35 requires a data protection impact assessment where processing is likely to result in a high risk, including systematic and extensive evaluation of people based on automated processing that leads to decisions with legal or similarly significant effects. The Article 29 Working Party’s DPIA guidelines (2017, endorsed by the EDPB) treat evaluation or scoring, systematic monitoring and employees as factors, and say processing meeting two of their criteria will in most cases need a DPIA. Ask the vendor for the information you need to write it; Article 28(3)(f) obliges a processor to help.
16. Automated decisions about agents. Article 22 gives people the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects. If QA scores feed decisions about agents, make sure a person reviews them and agents can contest a score. Check national employment rules too: Article 88 lets EU countries set more specific rules for employee data, and works councils may need to be consulted.

AI-specific checks

The AI checks find out what happens to your data inside the AI features, who can overrule the AI, and where the tool sits under the EU AI Act. Ask for written answers you can attach to the contract.

17. Model training. Ask in writing whether your conversations are used to train or fine-tune any model, by the vendor or by its model providers.
18. Model providers. Ask which AI providers process your data, where, and whether the AI features can be switched off or are opt-in.
19. Human review and override. Check a reviewer can change any AI score or comment, that the change is recorded, and that agents can dispute a score, as in our guide to designing an AI QA score appeal process. This is also your safeguard under Article 22.
20. Explainability and testing. Ask whether each score shows the evidence behind it, as described in our post on QA score traceability, and whether you can test AI scores against your own reviewers’ ratings before go-live. Our protocol to validate AI QA scoring shows how to run that test.
21. EU AI Act position. Ask how the vendor classifies the product and what it gives you as a deployer. The table below sets out the dates.

The ban and the high-risk point come from the AI Act, Regulation (EU) 2024/1689, checked against the consolidated text. The high-risk date moved from 2 August 2026 to 2 December 2027 under the amending Regulation (EU) 2026/1744, in force since 27 July 2026.

Under Article 3(39), emotion recognition means inference from biometric data, so the ban is mainly a question for voice features. Whether a QA tool falls under point 4(b) depends on how you use the scores, so have your lawyers decide it for your use.

Procurement and contract checks

The procurement checks turn the answers into contract terms and make sure you can leave cleanly. A security answer that is not in the contract is a sales statement.

22. Contract terms. Put the answers to checks 7, 10, 12, 14 and 17 into the contract or the data processing agreement.
23. Exit. Agree how long you have to export data after termination, in which formats, and when the vendor deletes it.
24. The document pack. Collect in one place: the audit report, certificates, data processing agreement, sub-processor list, pen test summary, incident response and continuity plans, and any standard questionnaire the vendor has completed, such as a SIG Lite.

How Kaizo answers this checklist

Kaizo publishes most of its answers on its Trust Center, with the reports available on request. The table maps the checks to what Kaizo’s own pages say today; where a page does not answer a check, ask us in the security review.

Kaizo scores conversations after they end. It does not monitor live calls, which keeps real-time listening out of scope for this review.

How to run the security review without stalling the deal

Start the security review in the same week as the demos, not after the business has chosen. Send the whole checklist at once, ask for the document pack, and start your impact assessment in parallel.

Step 1: Send the checklist with the RFP

Security, legal and procurement questions sent with the first vendor request get answered alongside the product questions. Sent after a decision, they become a second project.

Step 2: Ask for documents, then ask only what they leave open

Read the audit report, data processing agreement and sub-processor list first. Most questionnaire lines are answered there, and the rest become a short list for one call.

Step 3: Write the DPIA while the contract is negotiated

Your impact assessment needs the vendor’s answers on data flows, sub-processors and retention, which you now have. Involve your works council or employee representatives early if agents’ scores will be used in performance decisions. Our guide to running a QA program agents trust covers how to explain the scoring to them.

Ask for our security pack in a live demo. We walk you through Kaizo on realistic sample data and send the Trust Center documents your team needs. Let’s talk.

Frequently asked questions

What is a vendor security questionnaire?

A vendor security questionnaire is a standard list of questions a buyer sends a supplier to check its security and privacy controls before signing. Common formats include the SIG and SIG Lite from Shared Assessments and the CAIQ from the Cloud Security Alliance. Many vendors keep a completed one ready, which saves both sides weeks.

Does connecting a QA tool to Zendesk affect GDPR compliance?

Yes. When a QA tool pulls conversations from Zendesk or any helpdesk, the QA vendor processes that personal data on your behalf, so you need a data processing agreement with the QA vendor itself. The helpdesk’s own compliance does not cover a separate tool. Add the QA vendor to your records of processing.

What are the 7 principles of GDPR?

Article 5 of the GDPR sets seven principles: lawfulness, fairness and transparency; purpose limitation; data minimization; accuracy; storage limitation; integrity and confidentiality; and accountability. For AI QA, minimization and storage limitation matter most: score only the queues you need and keep conversations no longer than your QA purpose requires.

Is AI QA software high-risk under the EU AI Act?

It can be. Annex III, point 4(b) lists AI intended to monitor and evaluate the performance and behavior of people at work as high-risk, with obligations applying from 2 December 2027. A QA tool used to evaluate agents may fall in it, depending on its intended purpose and your use. Get a legal view for your case.

Do agents have to be told that AI scores their conversations?

Yes, under the GDPR’s transparency rules in Articles 13 and 14, agents must be told how their personal data is processed, including automated evaluation. Many countries add employment rules, and works councils may have a say. Telling agents early, with how they can dispute a score, also makes the rollout easier.