Why we added an AI embedding model (and where it runs)
As of this release, Aquil uses an OpenAI model to search your documents, and a fair question comes with it: where does your data go now? The short answer is nowhere new. The model runs in the same Microsoft Azure region in Sweden as the one we already used, inside Sokigo's own tenant, and your document text is never sent to OpenAI.
The longer answer is worth a few minutes, because it covers the two things a compliance owner has to answer for: what the AI reads, and what it keeps. This article is meant as something you can point to internally, or send to a customer whose security team asks.
What changed
Until this release, when Aquil suggested which of your own documents to ground a generated policy on, or matched your ISMS capabilities against a procurement requirement, it looked for shared words. Useful, but literal. A document titled "Leverantörsbedömning" would not surface for a search about "supplier due diligence", even though it is exactly the right file.
Now there is a second tier that searches by meaning. text-embedding-3-small turns your document text into a numeric vector (a list of 1,536 numbers that captures roughly what the text is about), and Aquil compares those vectors to find topically related documents even when the words do not line up. You will see it in two places today: the plan step of AI document generation suggests better sources, and the Requirement Tracker's capability matcher finds better matches.
Where it runs, and where it does not
text-embedding-3-small is an OpenAI model, and that is the part worth being precise about. We do not call OpenAI's API. The copy we use is deployed on Microsoft Azure AI Foundry in the Sweden Central region, in Sokigo's own Azure tenant, on the same deployment and in the same region as gpt-oss-120b, the model that already powers generation and chat. The encryption keys for that platform sit in Azure Key Vault held by Sokigo, not by Microsoft.
Two commitments did not change and do not change here. No customer prompts, completions or embedding inputs are used to train any model, by Sokigo or by Microsoft. And everything stays inside the EU. If your reason for choosing Aquil was EU data residency, adding this model does not move your data an inch.
What gets stored
Semantic search is the only part of this change that alters what Aquil keeps, which is why it is worth spelling out. Normally Aquil retains nothing from an AI request once your session ends: not the prompt, not the response. Here there is one exception. The numeric vector for a document is stored for as long as the document exists, so we do not have to re-read and re-embed every file on every search.
What is stored is the vector, never the text. A vector is not readable; you cannot reconstruct a document from it. Delete the document and its vector is deleted in the same operation. And the search is access-scoped in the database query itself: it only ever considers documents you already have permission to open, so the AI cannot use semantics as a side door into a file your role or scope would not let you see. This is the same principle behind how we let AI agents into your ISMS. Your document bodies stay encrypted at rest, exactly as before.
Why we updated the contracts before shipping
Adding a model that processes your content, and storing a new kind of derived data, are both things a data processor has to disclose up front, not things you should discover later. So we updated the paperwork first: the AI Usage Addendum, the Data Processing Agreement, the sub-processor list, the public Trust page and the in-app privacy notice, along with our internal records (the GDPR Article 30 register, the DPIA and the transfer impact assessment).
Because the AI Usage Addendum and the DPA are required documents, you will be asked to re-accept the updated versions the next time you sign in. That prompt is not boilerplate. It is there specifically to disclose this change. The sub-processor list names the model, the region and the legal basis; the Trust page has the plain-language version.
The short version for your security team
If someone in your organisation has to answer for this to a DPO, a security lead, or a customer's procurement team, here is what they need in one place:
- The model: text-embedding-3-small, embeddings only, no text generation.
- Where it runs: Microsoft Azure AI Foundry, Sweden Central region, in Sokigo's tenant. Not OpenAI's API.
- Training: none. Your inputs are never used to train any model, by us or by Microsoft.
- What is stored: the numeric vector only, for the life of the source document, deleted with it. The text is not stored.
- Access: the search only sees documents the requesting user can already open, enforced in the database query.
- On paper: disclosed in the AI Usage Addendum, DPA, sub-processor list and Trust page, all updated this release.
That is the whole change. We added one narrow AI model that only turns text into numbers, running in the same Swedish Azure region as the model you already trusted, so search finds the right document more often. If your team wants more than this summary, the sub-processor list and the Trust page are the authoritative sources, and support is glad to walk a security reviewer through both.