The Data Processing Agreement: What Many Overlook

This is the first of six parts in a series. In this piece, we'll explore what a data processing agreement really is and why Icelandic businesses need to pay close attention to how these terms are changing. In part 2, I'll cover the major shifts that have shaken the market over recent months. In part 3, I'll map out every approach I know of that leads back to the same AI models. In part 4, I'll go through what this means for buyers and service providers. In part 5, I'll look at data that can leak through code during software development. And in the final part 6, I'll focus on model choice itself and what matters when you talk to a service provider.

Bringing AI into your company without reviewing your data processing agreements is a bit like inviting a mysterious guest into your office, handing them a folder with your most sensitive information, and trusting blindly that they'll put it straight into a locked storage room instead of taking it somewhere else entirely.

To make sure this guest doesn't simply walk out and wander off with your data, we need clear ground rules. That's exactly what a data processing agreement is designed to do.

A data processing agreement (often abbreviated as DPA) is a legal contract between a data controller (your company that owns the data) and a data processor (the service provider that handles the data on your behalf).

This agreement should clearly define where data is stored, who has access to it, what it's being used for, and what happens if a security breach occurs. This isn't optional. Under GDPR and Icelandic data protection laws, there are strict requirements for how a third party can handle personal information on your behalf.

The idea is simple: You should know who is working with your data and under what terms.

But reality is more complicated

It's surprising how rarely anyone actually digs into these agreements. Many companies and organizations seem to simply trust what cloud providers settled on before: that data is in Europe and that everything was fine when the processing agreement was reviewed three years ago.

But today the technology landscape is changing much faster than we can keep pace with or assume things will stay the same. Sometimes people feel uncomfortable when the issue comes up, because it's easy to think the responsibility lies elsewhere and that someone else has already thought it through.

Meanwhile, LinkedIn fills up constantly with posts about how powerful Claude and GPT models have become. Most people just want to get their hands on the newest and best tools for work. Then there's often too much trust placed in the sales rep, the big tech companies, or just "someone" having this sorted out.

The truth is, this is a promise that doesn't hold on its own.

This article is an attempt to map out the landscape and put down on paper the things I believe Icelandic companies and organizations should be discussing much more.

Three cloud providers and Iceland

The three major cloud providers, Microsoft Azure, Amazon AWS, and Google Cloud, approach AI differently:

• Microsoft Azure: Offers OpenAI models and Anthropic Claude through Microsoft Foundry (formerly Azure AI Foundry), or through Microsoft 365 Copilot, which automatically selects models based on the task at hand.
• Amazon AWS: Offers Anthropic Claude and other models through Amazon Bedrock, and as of May 2026, also through Claude Platform on AWS. OpenAI models became available in preview starting in April after their exclusive deal with Microsoft ended.
• Google Cloud: Offers Gemini and open models alongside Claude through Vertex AI or the Gemini Enterprise Agent Platform as it's called now.

On the surface, these three options seem to do pretty much the same thing: package up models and sell them as a service. But once you start asking where AI queries are actually processed, things get complicated.

None of these three providers runs their AI processing in Iceland. Their nearest regions are:

• Azure: West Europe (Netherlands), North Europe (Ireland), Sweden.
• AWS: Frankfurt, Stockholm, Ireland.
• Google Cloud: Frankfurt, Belgium, Netherlands, Finland.

This means the moment an Icelandic company calls on AI from any of these three, their queries and data cross borders for processing, even if they're originally hosted in an Icelandic data center. This isn't necessarily a problem, but it's a fact that leaders need to be aware of.

Two sides of data protection

To understand the landscape, we need to keep two separate things completely distinct, because data protection in AI has two parts:

• Model training: Can AI use your data as training material to teach the next generation of GPT or Claude?
• Data location and processing: Where in the world is your data stored, and where does the processing happen while you're waiting for an answer from the AI?

Why does location matter if data isn't used for training?

It's easy to think: "If our processing agreement guarantees that Anthropic or Microsoft won't use our data to train their models, why should I care if our queries go to the US for a few seconds?"

The answer comes down to legal protection. Once personal information (like resumes, customer lists, medical records, or even employee emails) leaves the European data area, the protection of GDPR laws partly falls away. In the United States, different laws apply (like FISA Section 702 and the Cloud Act), which give US intelligence agencies and authorities much stronger powers to demand access to data hosted or processed there, without any say from European citizens.

For an Icelandic company (especially those in finance, law, healthcare, or government), this data transfer alone can be considered illegal under data protection laws, regardless of whether anyone is "training" on your data.

As we'll see in the coming pieces, there's real tension between wanting to use the latest tools and following the laws that protect personal data.

Icelandic companies and organizations have already been fined for transferring sensitive data to the United States. While domestic penalties are generally low today, that could change if intent is proven, and maximum fines can go as high as 20 million euros or 4% of global annual revenue.

To give a taste of what's coming, here's an interactive table showing all thirteen approaches I cover in this series. You can filter it by AI model, service provider, EU Data Boundary, and jurisdiction, and we'll dive into each approach in parts 2 and 3:

In the next piece, part 2, we'll look at the major changes that have shifted the landscape over recent months and why companies choose the approach they do.