Enterprise buyers now send an AI security questionnaire before they sign with any vendor that touches personal or confidential data. The questions repeat from one buyer to the next, and the answers are usually weak in the same places. This page lists 25 questions grouped by category, with model answers you can adapt. Answers marked for confirmation must be checked against your own contracts and controls before you send them.
A fuller, editable version of this questionnaire, with the evidence to attach to each answer, is available on request from the contact page.
1. Data
| Question | Model answer |
|---|
| What personal or confidential data will you process, and for what purpose? | Prompts, the context built from them and model outputs, processed only to provide the inference service you contracted for. |
|---|
| Do you process customer data as a processor, and do you hold a data processing agreement? | Yes. We act as processor for customer data and offer a data processing agreement covering UK GDPR terms. |
|---|
| How do you classify and label data in your service? | Customer data is treated as confidential by default and kept separate from operational data, with access limited by workspace. |
|---|
| Is customer data encrypted in transit and at rest? | Yes. Traffic is encrypted in transit, and every KV cache block is sealed with AES-256-GCM at rest, in every storage tier. |
|---|
| Do you use customer data for analytics or product improvement? | Not on prompts or cached context. Operational metrics are collected without customer content. |
|---|
2. Residency
| Question | Model answer |
|---|
| Where is customer data stored and processed? | In the UK, on dedicated GPUs in London by default. |
|---|
| Can you guarantee that data does not leave the UK, including backups and failover? | Residency is set by policy for compute and storage. [Confirm backup and failover locations before sending.] |
|---|
| Do you rely on any sub-processors outside the UK? | [Confirm the current sub-processor list and locations before sending.] |
|---|
3. Retention
| Question | Model answer |
|---|
| How long do you keep prompts, responses and logs? | Prompts and the KV cache are held for the session. Application logs follow the retention period in your contract. [Confirm log retention before sending.] |
|---|
| Can we set a shorter retention period, or request deletion on demand? | Erasure by session is available on request, and the KV cache is erased by key destruction with a certificate. |
|---|
| What happens to data when we end the contract? | Data is erased on termination, and an erasure certificate is provided. [Confirm the contractual deletion timeline.] |
|---|
| How do you handle a data subject's erasure request? | We erase by session, destroying the keys so every copy becomes unreadable, and provide a certificate. [Confirm the response timescale.] |
|---|
4. Isolation
| Question | Model answer |
|---|
| How is your service isolated between customers? | Each tenant has separate key hierarchies and KV cache. Prefix caching never crosses tenants. |
|---|
| Can one tenant’s cached context be reused by another? | No. Cached blocks are not shared across tenants. |
|---|
| Are your models and serving shared across customers? | Each deployment runs on dedicated GPUs, so compute is not shared between tenants. |
|---|
5. Training use
| Question | Model answer |
|---|
| Do you train any model, including the base models, on our data? | No. Customer prompts and outputs are not used to train any model. |
|---|
| If we fine-tune a model, who owns the weights and where are they stored? | The customer owns the fine-tuned adapter, which is stored in the customer’s UK region. [Confirm ownership terms in your contract.] |
|---|
| Can you show an evaluation for each fine-tune? | Yes. Each training run records an evaluation, kept with the weights and mapped to the EU AI Act articles it answers to. |
|---|
6. Incident response
| Question | Model answer |
|---|
| How do you detect and respond to security incidents? | Access and changes are logged to an append-only audit trail, and incidents are triaged against a documented response process. [Confirm response times and the runbook before sending.] |
|---|
| How and when will you notify us of a personal data breach? | Without undue delay, and within the timescale in our data processing agreement. [Confirm the notification window.] |
|---|
| Can we export audit evidence for an incident or audit? | Yes. Audit evidence can be exported for review, mapped to the controls you test. |
|---|
7. Access and assurance
| Question | Model answer |
|---|
| How is access to our workspace controlled? | Workspace-scoped API keys and role-based access. [Confirm the access model you offer.] |
|---|
| Which certifications do you hold? | [Confirm current certifications, such as ISO 27001 or SOC 2, before sending. Do not claim any you do not hold.] |
|---|
| Can we run a security review or audit of your service? | Yes, on reasonable notice and under the terms of your contract. [Confirm audit rights.] |
|---|
| Is multi-factor authentication enforced for staff access to customer data? | [Confirm MFA enforcement for staff before sending.] |
|---|
Questions
- Why does the questionnaire matter for a small vendor?
- Buyers use it to decide whether to proceed at all, so a clear, evidenced answer can shorten a sales cycle.
- Should I answer every question myself?
- Start with the product and legal team, and attach evidence rather than only stating claims. Answer only what you can prove.
- Where can I get the full version?
- Request it from the contact page.
