Ask for the document, not the promise
A vendor security review tests whether an AI call platform can show, in writing, how it handles recordings, models, access, disclosure and failure. If you run operations, risk or procurement for a regulated or customer-facing call team, the one change that matters is this. For every question, ask for the artefact (a report, a contract clause, a log or a live setting) and do not accept "we take security seriously". Buyers keep asking where recordings live, which models touch them and who owns them, and they want proof, not a better demo. Use the 25 questions below to run your review and to know what a pass looks like.
Know where recordings live
These six questions cover where your data sits and who controls it.
| # | Question | What a good answer looks like |
|---|---|---|
| 1 | In which region is each data type hosted: audio, transcripts, metadata? | A region named per data type, not "the cloud". |
| 2 | Can we choose the region, and will the contract fix it? | A region clause in the order form, with notice required before any move. |
| 3 | Which subprocessors and cloud providers touch our data? | A named list, with a commitment to notify you of changes. |
| 4 | Who owns recordings and transcripts? | Contract text saying you do, including derived outputs. |
| 5 | What are retention and deletion terms on exit? | A written timeline and a deletion certificate. |
| 6 | How is data encrypted in transit and at rest, and who holds the keys? | Named standards, plus a clear statement on customer-managed keys. |
Residency rules differ by region, so check yours. Examples include the EU's GDPR transfer rules and Nigeria's central bank circular of June 2026, which requires payment transaction data to be kept on servers in the country by 1 January 2027. Buyers ask about hosting and ownership early. In regulated markets many make local hosting a precondition, because infrastructure hosted abroad can create compliance hurdles under strict local and European data standards.
Pin down which models see your calls
These five questions cover the AI layer. Get the model names in writing.
| # | Question | What a good answer looks like |
|---|---|---|
| 7 | Which third-party models and speech or voice providers process our audio and text? | A named list per function, updated with notice. |
| 8 | Does our data train your models or any provider's? | A contract clause saying no, not a sales assurance. |
| 9 | Does the training opt-out bind your subprocessors too? | The same clause, flowed down to every provider on the list. |
| 10 | How is personal data kept out of prompts and scripts? | A described redaction or masking step, with an example. |
| 11 | Can a prompt or script drafted from past calls leak personal detail? | A review step before anything goes live, and a way to test it. |
Redaction by pattern matching can over-redact or under-redact, especially on names and other proper nouns. Buyers raise this worry themselves. Ask the vendor to run the redaction on a sample of your own calls and read the output yourself.
If you would rather put these questions to us directly, you can book a demo with Insight7.
Prove who can see what
These five questions cover who can reach your recordings and how you would know. Buyers ask for a SOC 2 report and for proof that data fed into the AI stays safe, so ask for the evidence itself.
| # | Question | What a good answer looks like |
|---|---|---|
| 12 | Do you hold an independent attestation (SOC 2 Type II, ISO 27001)? | The report itself under NDA, with a recent date and a stated scope. |
| 13 | Can a supervisor see only their own team's calls? | A role setting shown live, not described. |
| 14 | Do you support SSO and MFA? | Both enforced for all users, shown in the admin settings. |
| 15 | Is every play and export of a recording logged, and can we export the log? | A sample log with user, time and action. |
| 16 | Who on your support staff can open our data, and how are API and CRM connections secured? | An approval and logging process, scoped keys, a field list and a rotation schedule. |
A certificate logo is not a report. Read the scope, because an attestation that excludes the call platform proves little. Type II covers a period of operation, while Type I is a snapshot on one day.
Make the AI announce itself
These five questions cover disclosure, consent and calling conduct. Ask to see each setting live, because a policy PDF is not a control.
| # | Question | What a good answer looks like |
|---|---|---|
| 17 | Does the agent say it is AI at the start of every call, and can we set the wording? | The opening line played on a test call, with the wording editable by you. |
| 18 | How is consent captured and marked, and is the do-not-contact list honoured across copies of a campaign? | A consent field on each contact, and a suppression list that applies account-wide. |
| 19 | Does the platform enforce calling hours and retry limits? | Settings you can configure and see applied, not a promise that the customer will set them. |
| 20 | How is call recording consent handled by jurisdiction? | Wording and rules set per region, with a named owner. |
| 21 | Who tracks disclosure rules as they change? | A named person or process, and a change log. |
Rules vary by region, so check yours. The EU AI Act's Article 50 has applied since 2 August 2026 and requires any AI that talks to people, including automated phone systems, to say it is AI from the first interaction. In the US, AI-voice calls already need consent under the TCPA, following the FCC's 2024 ruling. Colorado's chatbot law takes effect on 1 January 2027, and disclosure rules are spreading state by state.
Test the failure and handover
Every agent meets a question it cannot answer, so ask what it does then. Customers expect a person to be available: 87% say a company using GenAI for service must offer a way to reach a human. Banks also want a short, low-risk pilot on a small sample, with payment only after results show a measured lift.
| # | Question | What a good answer looks like |
|---|---|---|
| 22 | What does the agent do on a question outside its approved content, and can a customer ask for a person at any point? | A demonstrated handover on a test call. The agent pauses and passes the call on. It never improvises an answer. |
| 23 | Which issues go to a human, such as complaints, billing disputes or a confused customer, and who sets the rule? | A written escalation rule that you can edit, with the receiving team named. |
| 24 | What stops a campaign automatically if calls go silent or quality drops, and who is alerted? | A stated pause threshold and an alert that reaches a named person. |
| 25 | Can we test with simulated calls, then run a small pilot with success measured before we pay? | Pilot terms in writing: sample size, metrics agreed up front, payment tied to results. |
Ask for a test call and listen to the handover. A transfer that leaves the customer repeating themselves counts as a failed answer.
Score the answers before you sign
Mark every answer as shown, described or asserted. Shown means you saw the setting, report or log yourself. Described means the vendor explained it and agreed to send the document. Asserted means you were asked to trust it. Treat any asserted answer on data location, model training or disclosure as open, and do not sign until it moves up a level. Then write every shown answer into the contract or order form.
A verbal assurance stops counting once the person who gave it leaves the deal.
Related reading: How AI Voice Agents Redact Sensitive Data From Call Recordings and Transcripts, Contact Center AI Risk Assessment Checklist for Customer-Facing AI Agents and How to Write AI Voice Agent Guardrails That Keep Conversations On Script in Regulated Industries.
Questions people ask about vendor security
Is a SOC 2 report enough for an AI call vendor?
It is a strong start, not a finish. Check that the report is Type II, that its period ended within the last twelve months, and that the call platform sits inside the stated scope. Then read the listed exceptions. A report cannot tell you where your audio is hosted or whether a model provider trains on it, so questions 1 to 11 still apply.
How long should a security review take?
Run it in parallel with the pilot, not before it. Bring your IT team in at the first meeting. Deals stall when the person on the call is not the user or the budget holder, and a late security reviewer is one version of that. Data protection questions tend to arrive the moment IT sees an API integration, so have the field list and key handling ready before then.
Should we ask for data hosted in our own country?
Ask whenever a rule or your regulator requires it. Where it is only a preference, ask for the available regions and the dates they are offered, in writing. A promise of a future region has no value until it appears in the order form with a date.
What if the vendor says a feature is on the roadmap?
Treat it as not available. Either write it into the contract with a delivery date and a remedy if it slips, or leave it out of your scoring.
The strongest vendors volunteer the artefact before you ask: the report, the clause, the log sample. A vendor that hesitates over a document now will hesitate again after you sign. Which of these answers can you show me today, and which are you asking me to take on trust?

