AI Privacy Service
Personal data out before the LLM, back in after
The service that lets a product use an external language model without sending it anyone's personal data. Names are swapped for consistent stand-ins, figures for tokens and identifiers for placeholders, then the model's answer is put back together.
What It Is
AI Privacy Service sits between an application and an external language model. The application tells it which people it knows about, passes text through it before the text goes to a model, and passes the model's answer back through it afterwards. Names become made-up stand-ins that stay the same for the whole conversation, figures can be turned into tokens, and anything a company's redaction policy covers (emails, phone numbers, bank details, national IDs, addresses, health terms, unknown names) becomes a labelled placeholder.
On the way back it checks the answer as well as restoring it: it reports names the model invented, stand-ins it did not issue and placeholders that were lost. If a check fails, nothing is restored. The service keeps no copy of the text, and what it needs to restore an answer is sealed, tied to one company and one turn of the conversation, and expires within the hour.
Why It Exists
It began inside HR Assistant, where employees' names, salaries and circumstances run through every question, and was moved into its own service so that every product sending text to a model can share one tested implementation. The rule it enforces is simple to state and hard to keep by hand: nothing personal leaves for an outside model, and if the masking cannot run, nothing leaves at all.
What It Is Worth
Fails closed
A caller that cannot reach the service, or gets an error back, sends nothing to the model. Every caller has its own key, and an attempt to restore an answer for the wrong turn is refused.
Answers still read naturally
Stand-in names are consistent within a conversation, so the model reasons about "Nico Ridley" just as it would about the real person, and the reply comes back with the real names in place.
Catches what the model makes up
Reveal flags people the model invented and placeholders it dropped, so an application can tell a sound answer from a damaged one instead of showing both the same way.
One policy, many products
Masking rules live in one service with one test suite, rather than being re-implemented, slightly differently, in each product that calls a model.
Who It Helps
Product teams
Can use the strongest hosted model for a feature without first proving that no personal data is in the prompt; the service does that job and reports what it changed.
Data protection leads
Get a single, documented control point for what reaches an external AI provider, with metrics on every replacement and refusal, and no stored copy of the text to secure.
The people in the data
Employees, customers and anyone named in a document stay out of a third party's logs, whatever that provider's retention policy turns out to be.
At a Glance
| Status | Internal: runs inside Egon Expert Ltd and is not sold on its own; In testing |
|---|---|
| Stores | Nothing: no text, no per-turn state on the server |
| Used by | HR Assistant |
| Built by | Egon Expert Ltd |
Related Products
Need Something Like This?
We build platforms like AI Privacy Service for clients, and help in-house teams build their own. Start with a free 30-minute conversation about what you need.
Book a Free Consultation