> For the complete documentation index, see [llms.txt](https://docs.askyazi.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.askyazi.com/security-and-privacy/ai-and-your-data.md).

# AI and your data

Which AI features Yazi has, which models and services run them and where, what data they receive, and how to switch AI off for your organisation.

Yazi uses AI for interviews, analysis, transcription, translation and the in-app assistant. This page sets out which models and services do that work, where they run, what personal data they see, and how to turn AI off if your governance rules require it.

{% hint style="info" %}
**In short:**

* Every AI feature runs through Amazon Bedrock in Yazi's own AWS account. The AI interviewer uses an OpenAI model; other features use models including Anthropic's Claude. Voice notes are transcribed within Yazi's AWS account in London (eu-west-2).
* Your DPA can prohibit the use of your data and your participants' data to train AI models. See [Compliance](https://docs.askyazi.com/security-and-privacy/compliance.md).
* AI can be switched off for your organisation, completely or feature by feature. Transcription and image checks keep working.

**Last checked:** 23 October 2026.
{% endhint %}

## Where Yazi uses AI

Yazi has 14 AI features, each checked against your organisation's AI setting before any model is called. They include:

* the AI interviewer in AI-moderated interviews
* **Ask Yazi**, the in-app assistant, and creating a study with AI from a brief
* AI analysis on **Graph Data**, the **Insights** tab and the generated **Report**
* highlight reels
* voice-note transcription and translation
* reading receipts and documents, and analysing photos
* checks on answers and media (response quality)
* AI audience suggestions and simulated participants for testing

## Which models run where

Every AI feature is run through Amazon Bedrock in Yazi's own AWS account, so prompts and responses stay inside Yazi's AWS environment. The AI interviewer uses an OpenAI model. Ask Yazi, AI analysis, reports and translation use large language models including Anthropic's Claude.

Other AI and media services run as follows:

| Feature                             | Model or service                     | Where it runs                           |
| ----------------------------------- | ------------------------------------ | --------------------------------------- |
| Photo and image analysis            | Claude through Amazon Bedrock        | UK/EU, since 3 September 2026           |
| Receipt and menu photo reading      | Claude in Yazi's AWS account         | UK/EU, since 16 October 2026            |
| Voice-note transcription            | Speech-to-text in Yazi's AWS account | London (eu-west-2), since 5 August 2026 |
| Image quality checks when AI is off | Amazon Rekognition                   | Yazi's AWS account                      |

For where everything else is stored, see [Data residency and hosting](https://docs.askyazi.com/security-and-privacy/data-residency-and-hosting.md).

## What data the AI receives

* **Only what the task needs.** The interviewer sees the study's questions, the participant's earlier answers and your interview configuration.
* **Names are trimmed.** Surnames are not sent to AI models, and names are shortened to a first name.
* **Participant text is treated as data.** Answers and uploaded participant lists are marked as information for the model to read, never as instructions to follow. This protects the interviewer from instructions hidden in an answer or an uploaded file.

## AI training

DPAs can include an explicit clause prohibiting the use of personal data, including survey responses, voice notes and other media, to train any AI or machine-learning model, and extending that prohibition to every sub-processor. Ask for it when you agree your DPA. See [What the DPA covers](https://docs.askyazi.com/security-and-privacy/compliance.md#what-the-dpa-covers).

Yazi's own product analytics are limited to aggregated, de-identified usage data and performance metrics. They do not include your research data, participant information or study objectives.

## Turning AI off

AI can be switched off for your organisation, either completely or feature by feature (there is a separate switch for each of the 14 AI features). This is built for organisations whose governance rules say some or all of their data must not be processed by AI.

{% stepper %}
{% step %}

### Ask for the change

An organisation admin asks Yazi to change the organisation's AI settings.
{% endstep %}

{% step %}

### The organisation owner approves

Yazi sends a 6-digit approval code to the organisation owner by email. The change only takes effect once it is approved.
{% endstep %}

{% step %}

### AI features show as disabled

Each AI feature that is switched off shows a disabled state instead of running. If a check is unclear, the feature fails closed and does not call a model.
{% endstep %}
{% endstepper %}

### What changes when AI is off

| Keeps working                                        | Stops                                                       |
| ---------------------------------------------------- | ----------------------------------------------------------- |
| Voice-note transcription                             | AI-moderated interviews (an interview study cannot go live) |
| Image quality checks, using Amazon Rekognition       | Ask Yazi and creating studies with AI                       |
| Surveys, diaries, broadcasts, Table Data, exports    | AI analysis, AI report overviews and AI summaries           |
| Reports, reels and results created before the switch | Building new highlight reels                                |
|                                                      | AI audience suggestions and simulated participants          |

This table shows the main effects, not every one of the 14 switches. When you request the change, ask Yazi for the full list and for what it means for translation in your studies.

Choosing **in-chat** or **WhatsApp Flow** for a study does not change whether AI is used. The question format and the AI setting are separate.

## AI assistants you connect (Claude and ChatGPT)

Your team can connect Claude or ChatGPT to Yazi under **Settings > Integrations > Connectors**, so they can ask questions about studies in plain English. The protections:

* The connection is read only unless an admin grants more. Separate permissions cover changing studies, seeing participant phone numbers and including demographics, and each can be revoked from the same screen.
* It can never delete anything, launch a study or see another organisation.
* Participant IDs are pseudonymous and specific to your organisation. Phone numbers are masked, and contact details are removed from answers unless you allow them.
* Participant text is labelled as data, not as instructions to the assistant.
* Every request is audited and rate limited.

Data your assistant retrieves goes to Anthropic or OpenAI under your own account's terms, not Yazi's. Check your organisation's AI policy before connecting. Set-up steps are in [Connect Claude or ChatGPT](https://docs.askyazi.com/integrations-and-api/connect-claude-or-chatgpt.md).

<figure><img src="https://358667285-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzMRQUXiGQlBwWdg6WEmG%2Fuploads%2FbZvSDGgTZvZLXl1pqDRE%2Fsettings-integrations-connectors.webp?alt=media&amp;token=aa490faf-f4e5-40c6-8c83-6f220ba2cc24" alt="Settings > Integrations > Connectors with the MCP server URL and set-up steps for claude.ai"><figcaption></figcaption></figure>

## Customising the AI interviewer

**Settings > AI Settings** holds the **Interview Prompt**, which sets the tone and behaviour of the AI interviewer for every AI interview in your organisation. It must keep the `{objective}` and `{survey_data}` parameters. Changing it does not change which model runs or where. See [AI interviews](https://docs.askyazi.com/study-types/ai-interviews.md).

<figure><img src="https://358667285-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzMRQUXiGQlBwWdg6WEmG%2Fuploads%2F0m65s1q4fmvNmZYviPpw%2Ffs-settings-ai.webp?alt=media&amp;token=2abc4923-efa1-4441-9012-a892088c0bfc" alt="Settings > AI Settings: the organisation-wide interview prompt, which must keep the {objective} and {survey_data} parameters"><figcaption></figcaption></figure>

## Limits and things to know

* **AI can be wrong.** Review AI analysis, themes, translations and transcripts before they reach a client. Translations can be edited by hand.
* **Some languages cannot be transcribed.** Several South African languages, including isiXhosa, Sesotho and Setswana, have no transcription today. Voice notes in those languages stay untranscribed. See [Language support](https://docs.askyazi.com/study-types/language-support.md) for the full list.
* **AI cannot watch video.** Checks on video answers assess the audio track only.
* **AI settings apply to the whole organisation.** They are set per feature, not per study.
* **Connected assistants are outside Yazi's residency.** Anything Claude or ChatGPT retrieves is processed by that provider.

## Related

* [Security overview](https://docs.askyazi.com/security-and-privacy/security-overview.md)
* [Data residency and hosting](https://docs.askyazi.com/security-and-privacy/data-residency-and-hosting.md)
* [Participant privacy](https://docs.askyazi.com/security-and-privacy/participant-privacy.md)
* [Response quality and validation](https://docs.askyazi.com/build-a-study/response-quality-and-validation.md)
* [Connect Claude or ChatGPT](https://docs.askyazi.com/integrations-and-api/connect-claude-or-chatgpt.md)
