The person most likely to stop an AI coaching rollout has never seen the demo. They sit in IT security or procurement, and they need four answers before anything is signed: how employee data is protected, who certifies the security, where the data lives, and what the AI is for.
An AI coaching vendor security assessment checks four things: GDPR compliance for the recordings and scores the platform creates, an ISO 27001 certificate whose scope covers the vendor, where the data is hosted, and the system's intended purpose under the EU AI Act. Retorio is GDPR-compliant, ISO 27001 certified, EU AI Act aligned and hosted on GCP with EU data residency.
Example. An IT security analyst at an insurer gets a new AI coaching platform to review. Instead of sending the full company questionnaire, she asks the twelve questions below, requests three documents by name, and settles the review in one meeting with the sales enablement lead who picked the vendor.
This page is written for the reviewer, and for the champion who has to send something to the reviewer. It is not a sales page. Most of it is a checklist you could put to any AI coaching vendor, including us. Where our own answer is a matter of public record, it says so plainly. Where it is not, the checklist tells you what document to ask for instead of what to believe.
Our own position first, so nobody has to dig for it. Retorio is GDPR-compliant and ISO 27001 certified. The platform is hosted on Google Cloud Platform with EU data residency. Retorio is aligned with the EU AI Act. Everything else on this page is a question you should put to us, or to anyone else on your shortlist, in writing.
A CRM stores what a rep typed. An AI coaching platform stores how a rep sounded. A practice conversation produces a recording, often a transcript, and a score for the behaviors the session was built to coach. All three are personal data about an employee, and the score is an evaluation of that employee. That combination is why a standard SaaS questionnaire misses the questions that matter here.
Three things follow from it. First, the data protection review has to cover the employee as the data subject, not the customer on the other side of the conversation, who in a practice session is usually a virtual persona. Second, the EU AI Act review turns on what the scores are used for, which is a question about your organisation as much as about the vendor. Third, in Germany and Austria the same question lands with the works council, because a system that evaluates employee behavior is a co-determination topic before it is a technology topic.
There is also a simpler reason these reviews go badly. The person who chose the vendor is rarely the person who checks it, and the first does not reliably tell the second what they saw. In a Gartner survey of about 900 third-party relationship owners, 95% had seen a red flag with a third party in the previous twelve months, and only around half had passed it on to compliance.
Read the chart as a buyer committee. The champion is the relationship owner. They sat through the demo, asked the vendor the questions that occurred to them and formed a view. Very little of that reaches the reviewer. So the reviewer starts from zero, and a vendor who answers only the champion's questions has answered the wrong person.
The questions below are grouped by the area a reviewer signs off. The middle column names the evidence that closes each one. The last column is the answer that should make you slow down. None of it depends on which vendor you are looking at, and a vendor who cannot produce the evidence column has told you something too.
| Area | Question to ask | Evidence that closes it | Slow down if |
|---|---|---|---|
| Data protection | Is the vendor GDPR-compliant, and does it act as your processor? | A signed data processing agreement that names the processing purposes | A privacy policy is offered where a processing agreement was asked for |
| Data protection | Which personal data does one coaching session create? | A data inventory per feature: recording, transcript, score, usage data | The answer is that a practice session holds no personal data |
| Data protection | How long are recordings and scores kept, and who can delete them? | A retention schedule in the processing agreement, plus the deletion route | Retention is described as "as long as needed" |
| Information security | Is there an ISO 27001 certificate, and what does its scope cover? | The certificate itself, with issuer, scope statement and expiry date | The certificate belongs to the hosting provider, not to the vendor |
| Information security | Who at the vendor can reach customer data, and how is that logged? | The access control policy and how support access is granted | Nobody can say who has access today |
| Hosting | Where is the data stored and processed? | Hosting provider and region, written into the processing agreement | Storage is in the EU but processing location is not stated |
| Hosting | Which sub-processors touch the data, including any AI model provider? | A current sub-processor list with locations and a change-notice clause | The AI model provider is missing from the list |
| EU AI Act | What is the system's intended purpose, in the vendor's own documents? | A written intended purpose, matched against your own usage policy | Scores are expected to feed promotion or termination decisions |
| EU AI Act | Does the system infer a person's inner feelings from voice or face? | A statement of what is analysed: observable behaviors, not inner states | The marketing promises to read what a rep feels |
| EU AI Act | What does your organisation do about AI literacy for the people using it? | Your own short record of the measures taken, and for whom | Someone sells you a certificate as the answer |
| Governance | Who sees an individual rep's scores: the rep, the manager, HR? | A role and permission model, agreed with the works council where one exists | Visibility is settled after the rollout, not before |
| Operations | How are users signed in and removed, and how are incidents reported? | Authentication documentation and an incident notification clause with a time frame | Leavers keep access until someone notices |
Twelve questions is deliberately short. A reviewer who sends a generic questionnaire of several hundred items to a coaching vendor will get several hundred generic answers back, and the three rows that are specific to AI coaching (what a session stores, what the scores are for, who can see them) will be buried in the middle of it.
These are the four answers a procurement team usually copies into its own summary, so here they are in one place.
Retorio is ISO 27001 certified. Ask for the certificate and read its scope and expiry date yourself.
Hosted on Google Cloud Platform with EU data residency. Retorio is GDPR-compliant.
Enterprise clients across insurance, pharma, telecommunications, automotive and consumer goods.
The fourth fact is the EU AI Act. Retorio is aligned with it, and the next section explains what that alignment has to mean for you as the organisation using the platform, because some of the Act's questions can only be answered on your side of the contract.
We have written a longer piece on what the EU AI Act means for AI coaching in SaaS. For a vendor review, four points are enough, and each can be checked against the consolidated text of the Regulation on EUR-Lex, in the version of 27 July 2026.
Intended purpose decides the risk class. Annex III, point 4(b), lists as high-risk the AI systems intended to be used for decisions on promotion or termination, for allocating tasks based on individual behavior, or "to monitor and evaluate the performance and behaviour of persons" in work relationships. A coaching platform sits close to that wording, which is why an attentive reviewer stops there. What decides it is the documented purpose and your actual use. A system intended for voluntary practice and feedback, whose scores do not feed promotion or termination decisions, is being used for something else. Article 6(3) then adds an exception for narrow or preparatory tasks, with a hard limit: an Annex III system that profiles natural persons is always high-risk.
One use is prohibited outright. Article 5(1)(f) bans systems that infer the inner feelings of a person at the workplace, except for medical or safety reasons. That is the reason for the ninth question in the checklist. A platform that analyses observable behaviors in a conversation is doing something different from one that claims to read what a rep feels, and the vendor should be able to say which it is in one sentence.
AI literacy is a duty of effort, not a test. Article 4 has applied since 2 February 2025. It was amended by Regulation (EU) 2026/1744, and the current text requires providers and deployers to "take measures to support the development of AI literacy" of their staff. It states that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". Nothing in it asks for a pass mark or a certificate.
Article 4 carries no fine of its own in the Act. Article 99 sets three fine tiers: up to EUR 35 million or 7% of turnover for the prohibited practices in Article 5, up to EUR 15 million or 3% for a listed set of operator obligations, including those of deployers under Article 26, and up to EUR 7.5 million or 1% for misleading information given to authorities. Article 4 is not on any of those lists. Anyone quoting the top tier as the price of a missing literacy plan has attached the wrong article to it.
A certificate tells you whether a vendor protects the data. It does not tell you what the platform is for, and that is the question the EU AI Act asks.
Retorio Customer Success Team, on enterprise security reviews
Every row in the checklist ends in one of three places. Sorting them early is what keeps a review to one meeting.
The first branch is where reviews tend to lose time. A certificate is treated as the end of the question when it is the start of one: an ISO 27001 certificate says an information security management system was audited for a defined scope, by a named body, until a stated date. Read all three. The second branch is where the work is. A written retention schedule, a sub-processor list and an intended-purpose statement are the three documents that close most of an AI coaching review. The third branch should be empty by the time anyone signs.
In our own enterprise reviews, the first question from IT is almost never about encryption. It is who can see which score, and how long it is kept. When that answer is written down before the first meeting, together with the intended purpose, the review tends to run as one conversation instead of three.
If you are the sales enablement lead or the commercial owner who picked the platform, your job in this phase is not to argue for the vendor. It is to hand the reviewer a file that makes the review short. Four steps, in this order.
Write down what the platform is for
One paragraph: which reps use it, which behaviors it coaches, and that the scores are for practice and feedback. Say explicitly what the scores are not used for. This is the document the EU AI Act part of the review turns on.
Done when: the reviewer and the works council read the same paragraph.
Send the four facts and the certificate together
Data protection status, the ISO 27001 certificate, the hosting region and the vendor's AI Act position, in one message. Reviewers summarise, so give them the summary.
Done when: nothing in the first message needs a follow-up question.
Agree who sees an individual's scores
Rep, manager, enablement, nobody else, or a different split. Settle it before the rollout. It is the question that stalls reviews in co-determined companies, and it cannot be fixed afterwards without breaking trust with the reps.
Done when: the permission model is written into the rollout plan.
Map the remaining questions to documents
Take the checklist, mark each row with the document that closes it, and request the missing ones from the vendor in one request, not twelve.
Done when: every row has a document name next to it.
The same logic applies earlier in the buying process. When a team evaluates AI sales simulations against each other, the security and purpose questions are cheaper to ask of three vendors in the shortlist than of one vendor at contract stage. It also helps to know what is being scored before you decide who may see it. Our explainer on the behavioral signals an AI coach scores lays that out.
One more point for reviewers who have to justify the time. Managing AI risk is not a brake on getting value from it. In McKinsey's 2026 State of AI survey, the high performers (the 6 percent of respondents who attribute significant value and an EBIT effect of 5 percent or more to AI) are also the ones actively managing more of the risks that come with it. A short, specific review is part of how a company gets there, not a detour from it.
An AI coaching platform deserves a different review from ordinary SaaS because it records employees and scores their behavior. Close the data protection questions with a processing agreement and a retention schedule. Close the security questions with a certificate whose scope you have read. Close the hosting questions with a region written into the contract. Close the EU AI Act questions with a written intended purpose that keeps the scores out of promotion and termination decisions, and with your own record of AI literacy measures.
Then put the same twelve questions to every vendor on the shortlist. The one whose answers fit on two pages is usually the one that has been through this review before. If the shortlist is still being built, our buyer's guide to sales coaching software covers the commercial criteria. This page covers the ones IT and procurement will add.
See what your reps would actually use
Run one practice conversation and see what a session scores, before the review starts.
Test AI coach in actionYes. Retorio is GDPR-compliant and ISO 27001 certified, is aligned with the EU AI Act, and hosts the platform on Google Cloud Platform with EU data residency. For your own review, ask for the data processing agreement and the ISO 27001 certificate, and read the certificate's scope and expiry date.
Beyond the usual security items, three questions specific to AI coaching: which personal data one session creates (recording, transcript, score), what the scores are used for, and who can see an individual rep's scores. Add the retention schedule, the sub-processor list including any AI model provider, and the vendor's written intended purpose.
It depends on its intended purpose and on how you use it. Annex III lists AI systems intended to monitor and evaluate the performance and behavior of people in work relationships, or to inform promotion and termination decisions, as high-risk. A platform used for voluntary practice and feedback, whose scores stay out of those decisions, serves a different purpose. Ask the vendor for its documented intended purpose and match it to your own usage policy.
Article 4 has applied since 2 February 2025. As amended in 2026, it requires providers and deployers to take measures to support the AI literacy of the people using a system, and it states that no specific level has to be guaranteed for any individual. A short record of what you did, and for whom, is the practical answer.
That the vendor's information security management system was audited against the standard, for a defined scope, by a named certification body, and is valid until a stated date. It does not tell you what the AI is used for or how long recordings are kept. Read the scope and the expiry date, and check that the certificate belongs to the vendor and not only to its hosting provider.
IT security and procurement own the review, the data protection officer signs off on the processing agreement, and in co-determined companies the works council agrees who may see individual scores. The sales enablement lead who chose the platform should prepare the file: intended purpose, the four compliance facts, the certificate and the permission model.
Retorio is GDPR-compliant, EU AI Act-aligned, and ISO 27001-certified. Hosted on Google Cloud Platform with EU data residency. Your data stays in Europe.
Vodafone VOIS · AXA · Daimler Truck · Kraft Heinz · Merck · Madrigal
Built in Munich, Germany. 50+ enterprise clients across insurance, pharma, telecommunications, automotive, and consumer goods.
Written with help from LLMs, edited and checked by Önder Mutluer, Website Manager and Marketing Strategist, and Dr. Patrick Oehler, Co-founder and Co-CEO. How this article was researched, written and checked, and who reviews which topic: Retorio editorial policy.