Data, Access & Delivery

1. Where your data lives

In almost every engagement your data never leaves your environment. We work inside your cloud tenancy, against your systems, under your identity and access controls. What that means contractually depends on what we actually touch:

Engagement shape	Our role	What we sign
Hands-on work in systems that hold personal data	Processor on your behalf under Art. 28 GDPR	AV-Vertrag (DPA) plus documented technical and organisational measures under Art. 32
Work with anonymised, aggregated or synthetic data only	No personal data processed	NDA. We record how that determination was made, so it is auditable later
Advisory, architecture review, training - no system access	Neither controller nor processor	NDA

Worth being precise about one point, because it is often misread: personal data staying on your infrastructure does not by itself put us outside Art. 28. Processing under Art. 4(2) includes access and consultation, not just storage. Where we can see personal data, we act as your processor and paper it properly - whose disks it sits on does not change that.

2. How access works

Access is granted by you, to named people, for the work at hand:

+ Named individuals, never a shared account. You always know who has access.
+ Least privilege. The narrowest scope that lets the work happen, and no standing production access where a read replica or a scoped role will do.
+ Time-limited. Access ends when the engagement ends. We ask you to revoke it; we also tell you when we are done with it.
+ No bulk export. We do not copy your data onto ProDataAI infrastructure, and we do not keep working copies after an engagement closes.
+ Your audit trail, not ours. Because the work happens in your tenancy, your existing logging and monitoring already sees everything we do. You do not have to take our word for any of the above - you can check it.

3. Who does the work

ProDataAI is founder-led. Kamlesh Kshirsagar is on every engagement, and that is the point of hiring a small firm: the person who scopes the work is the person who does it.

For specific pieces of work we bring in specialists from a network we have worked with before. That is how the model stays senior without carrying bench cost, and we are straightforward about it:

+ Anyone joining an engagement is named to you before they start.
+ Everyone signs an NDA and a back-to-back agreement under Art. 28(4) GDPR before touching any system of yours.
+ Specialists work under the same access rules in section 2 - named, scoped, time-limited.
+ No offshore delivery unless you ask for it.

If a subprocessor list matters to your procurement or your own DPA obligations, we give you the details you need for it, and tell you before anything changes.

4. The contract position

Signed before the first invoice, not after it. Note which direction each of these runs:

+ Your AV-Vertrag, not ours. Under Art. 28 GDPR the data processing agreement belongs to the controller, which is you. We work under your template. Your legal team already trusts it, it already matches your other processors, and nobody has to negotiate a consultancy's boilerplate to get started.
+ We complete the Art. 32 annex. The technical and organisational measures schedule attached to your AV-Vertrag - access control, encryption, deletion, incident handling - filled in for how we actually work on your systems, not with generic wording.
+ Anyone who touches your systems is named to you first. If a specialist joins the engagement you know who before they start, and they sign the same terms we do. That is the substance of an Art. 28(4) subprocessor disclosure, and it is how you keep your own register accurate.
+ NDA - always, including for advisory-only work.

5. Where the models run

Most of the risk in an AI project is not the model, it is where the data goes to reach it. There are three options and we will tell you which one your case actually needs:

+ Your own cloud AI service, in your tenancy. Azure OpenAI, Amazon Bedrock or Google Vertex under your subscription. This is the default and it covers most engagements: your data stays inside the boundary you already govern.
+ Self-hosted open-weight models in your environment. For cases where data must not leave the building at all - works council constraints, sector rules, or simply a policy you do not want to argue with. We offer this; it is not our most common setup, and we will say so rather than sell it to you.
+ EU-region managed endpoints. Where a managed service is genuinely the right call, we keep processing in EU regions and document the transfer position.

No client data is used to train models. Not by us, and we configure provider settings so it is not used by them either. Where a provider's default would do otherwise, we change it and show you the setting.

6. What we do not do

+ We do not move your data to our infrastructure.
+ We do not train models on your data.
+ We do not keep copies, backups or notebooks containing your data after an engagement ends.
+ We do not subcontract offshore unless you have asked us to.
+ We do not take standing production access we have no use for.

7. This connects to the EU AI Act

Data residency, access control and subprocessor discipline are the same evidence base the EU AI Act asks for. If you are working towards an AI inventory and risk classification, the controls on this page are already part of the answer.

Our practical guide covers what high-risk classification means, the December 2026 timeline after the Digital Omnibus, and the Betriebsrat dimension under §87 and §90 BetrVG that most guidance skips: EU AI Act - a practical implementation guide.