# AI Provider Selection
URL: https://docs.autociso.io/product/admin/ai-providers/
Description: Which AI providers AutoCISO uses, what each one receives from your data — especially images and personal data — and what to weigh when your organisation reviews its data-processing agreements.

AutoCISO uses AI to extract user lists from access screenshots and PDFs, classify assets and suppliers, and prefill ISO control evidence. Two providers are supported:

- **Anthropic Claude** (default). Screenshots and PDFs are sent to Claude as images and read directly.
- **DeepSeek**. Images are converted to text by OCR inside AutoCISO before anything is sent out. Email addresses in that text are masked, and only the masked text goes to the provider — image bytes never leave AutoCISO.

Which provider your organisation uses is set for your AutoCISO deployment. Ask your account manager or support if you need it changed.

## What changes per flow

| Flow | What the provider receives |
|------|----------------------------|
| [Audit ingest](https://docs.autociso.io/console/access/access-audit/) (screenshot or PDF upload) | With Claude: the image. With DeepSeek: OCR text only, with emails masked. |
| Asset extraction (uploaded image or PDF) | The image, always sent to Claude — regardless of the provider you use elsewhere. |
| ISO certificate extraction (PDF) | The file, always sent to Claude. |
| Asset and supplier name and website analysis | Text only, to your provider. |
| [AI chat](https://docs.autociso.io/console/overview/ai-assistant/) | Text only, to your provider. |

Image-based asset and ISO certificate extraction always uses Claude, so Claude remains a sub-processor for your organisation even if you choose DeepSeek for everything else.

## Personal data handling with DeepSeek

Before any text is sent, AutoCISO masks email addresses and replaces each with a placeholder that means nothing outside your request. The placeholders are restored in the result, so extracted rows still match your employees. The model never sees a real email address.

Names and phone numbers are not masked — only email addresses.

If you supply a custom prompt for an asset, the rules that handle these placeholders are always appended and cannot be removed.

## Extraction accuracy when only text is sent

With DeepSeek, no model ever sees your screenshot — it reads the text that OCR recovered. That text carries predictable damage, and it is worth knowing what to expect:

- **Characters are misread.** `Sergey Ilin` is commonly recovered as `Sergey lin` (capital `I` read as lowercase `l`). Extraction is instructed to repair obvious character damage and keep the surname rather than drop it.
- **Layout is flattened.** Sidebars, headers and buttons interleave with table rows, and badges next to a name (`Owner`, `Admin`) arrive as stray characters. Extraction reads these as roles or ignores them, not as names.
- **A damaged row is never "repaired" into a different person.** An empty field is preferred over a plausible guess, and a row must appear in the source to be reported.

**The email is what carries a match.** Employee matching treats an exact email as decisive, so a row whose name arrived damaged still attributes correctly as long as its email survived. Where the name is ambiguous and no email was recovered — a row reading only `Sergey` when two employees share that first name — AutoCISO raises it for review rather than guessing. See [Correct a misrecognised employee](https://docs.autociso.io/product/access/correct-employee-match/).

If accuracy matters more to you than keeping images inside AutoCISO, ask support to send audit screenshots to Claude as images. That sends image bytes to Anthropic — a deliberate trade of the text-only guarantee for extraction accuracy.

## When to choose DeepSeek

Choose DeepSeek when:

- Your organisation requires that no screenshot or PDF image leaves the platform for audit ingest.
- You want lower cost on text-only flows such as chat and name or website classification.
- Your data-processing agreement covers DeepSeek as a sub-processor. Update your RoPA ([Privacy Register](https://docs.autociso.io/console/privacy/register/)) and
  [Data Processors](https://docs.autociso.io/console/privacy/data-processors/) accordingly.

Remember that asset and ISO certificate image extraction still uses Claude either way.

## What does NOT change

The console looks the same whichever provider you use, and the audit results you see are stored in the same structure. Personal-data tokenisation on your own inputs works as before, independent of the email masking described above.