cogny shielda cogny cloud featureswedish-hosted
send the model the analysis. keep the pii at home.
Cogny Shield routes tool results (your CRM, BigQuery, Stripe, Shopify) through a Swedish-hosted model that tags personal data and swaps it for stable placeholders before the result reaches the frontier model. A per-workspace switch, metered by tokens.
why gdpr-bound teams hold back.
- PII leaves the EU by default
Most AI agents pipe whatever the CRM, ad platform or warehouse returns straight into a US frontier model. For workspaces processing customer data, that breaks GDPR before the analysis begins.
- Redaction cripples the analysis
Naive strippers replace every email and name with [REDACTED], so the model loses any sense of identity, cohort or follow-up.
- On-prem models aren't on the table
Running a frontier model on your own hardware isn't realistic for most teams. You need the frontier model, without sending it raw PII.
- A DPA only gets you so far
A signed Article 28 agreement doesn't change the fact that the data was transferred. Sensitive industries need a technical answer, not only a contractual one.
three stops between your data and the model.
- 01
The tool returns data
HubSpot, BigQuery, Stripe, Shopify: whatever the agent calls, the raw response comes back to Cogny inside the EU.
- 02
A Swedish-hosted model tags the PII
Every name, email, phone number, address, personal ID number and IP address is wrapped in a tag.
- 03
Placeholders go to the model
Each tagged value becomes a stable placeholder. The original values never leave the EU.
- What it masks
Email addresses, phone numbers, names, personal ID numbers, street addresses, IP addresses, card numbers, dates of birth and URLs that carry a user identifier.
- What it leaves alone
Company and product names, public URLs, ad account and campaign IDs, and aggregated metrics. They aren't personal data, and the analysis needs them.
what each layer sees.
- The tool returns
Hello, peter@company.com — your invoice for Anna Lindberg (070-123 45 67) is overdue. - The tagger marks
Hello, <cogny_pii type="email">peter@company.com</cogny_pii> — your invoice for <cogny_pii type="name">Anna Lindberg</cogny_pii> (<cogny_pii type="phone">070-123 45 67</cogny_pii>) is overdue. - The model sees
Hello, <masked_email1> — your invoice for <masked_name1> (<masked_phone1>) is overdue. - You see
<masked_email1> = peter@company.com · <masked_name1> = Anna Lindberg · <masked_phone1> = 070-123 45 67
Example data.
privacy without a blunter model.
- Stays in the EU
The tagging runs in Sweden, before any data crosses to the frontier model. The mapping is stored with the tool result in your workspace.
- Stable identifiers
A repeated value gets the same placeholder, so the model can still reason about identity and cohorts: <masked_email1> in three rows is one person.
- Full transparency
Every masked result shows a Cogny Shield card with the placeholder-to-original mapping. You see what the AI saw, and what it never saw.
- Fail-closed
If the tagger times out or errors, the tool result is replaced with an error message, not the raw text.
- Injection scan
A second switch scans tool results for text shaped like instructions to the model, and quarantines it before the model reads it.
- Per-workspace opt-in
Customer-data workspaces turn it on; analytics-only ones leave it off. Token-metered, billed with your AI usage. Turn it off any time.
what you give up.
- A little cost and time
Every successful tool result becomes one extra tagging call. For a heavy multi-tool analysis that's a small cost and latency overhead, visible in your billing.
- Less personal phrasing
The model loses some grip on names and identity-flavoured signals. It still reasons about cohorts and counts, but “write a follow-up to Peter” gets blunter.
- Worth it for most regulated teams
For them it's the difference between “we can't use AI on this data” and “we can.”