← Articles


How On-Premises AI Keeps Enterprise Data Inside the Firewall

When a company uses cloud AI to analyse its documents, emails or meeting notes, that data leaves the building. It travels to a vendor's servers, gets processed there, and returns as an answer. For many European directors and compliance leads, that journey is not an option.

The EU AI Act, GDPR (General Data Protection Regulation), NIS2 (the EU cybersecurity rules) and DORA (the EU digital operational resilience framework) all create pressure to keep sensitive business data on premises. Some companies cannot send customer information, financial records or proprietary knowledge outside their own infrastructure. Others choose not to, because the risk or the regulatory burden is too high.

On-premises AI solves this by running AI models directly on company hardware, inside the firewall. The data never leaves. The model processes it locally, returns the result, and stays put. For companies that cannot or will not send their data to the cloud, this is the way to use AI without sending that data out.

What on-premises AI actually requires

Running AI locally is not a simple toggle. It needs suitable hardware, usually a GPU (graphics processing unit) server, because AI models are computationally expensive. The speed of the model depends on that hardware. A company with a basic server will see slower responses than one with a high-end GPU. Running it also costs money: the hardware itself, the maintenance, the updates to the model, and the electricity to keep it running.

This is not faster or cheaper than cloud AI by default. It is a different trade-off. Cloud AI is fast and inexpensive per query, but the data goes somewhere else. On-premises AI keeps the data home, but the company pays for and manages the hardware.

The choice is not "cloud AI or on-premises AI". It is "do the benefits of keeping data local outweigh the cost and complexity of running it locally". For a company handling regulated data, the answer is often yes.

How on-premises AI reads company data without learning from it

A common misunderstanding: if AI runs on company data, does it train on that data and become a copy of the company's knowledge?

No. A local AI model reads and searches company documents. It does not train on them. Training is the process of building the model in the first place. That happens once, by the model's creator, using a large dataset. Once the model is built, it is frozen. When it runs on a company's premises and reads a document, it is using the document to answer a question in that moment. It is not learning or changing.

This distinction matters for compliance. The model does not absorb the company's secrets. It uses them to answer a question, then forgets them. The company retains full control of the data.

Audit trails for every AI action

Regulatory compliance also requires proof that AI was used correctly. Every question asked, every piece of data the AI touched, and every action it took need to be logged.

On-premises deployment makes this simpler. Because the data and the model are both inside the company's own infrastructure, every interaction can be recorded in one place. A company can build a complete, immutable audit trail: who asked the AI a question, which data it used to answer, what it did, and what the result was. That trail stays inside the firewall, under the company's control, and is available for regulators to inspect.

Cloud AI can also be audited, but the audit trail is split between the vendor's logs and the company's own records. On-premises, it is all one.

When on-premises AI makes sense

On-premises AI is the right choice when:

  • The company handles data it cannot send to external vendors (customer information, financial records, proprietary research).
  • Compliance rules or internal policy require data to stay on premises.
  • The company has the infrastructure team to run and maintain a GPU server.
  • The regulatory or business value of keeping data local outweighs the cost of the hardware.

It is not the right choice for every company or every use case. A small team with no infrastructure capability, or a company with no sensitive data, may find cloud AI simpler and cheaper. On-premises AI is for companies that have made a deliberate choice to keep their data home.

ERIS with local AI models

ERIS is built to run on the company's own premises with local AI models, on the company's own GPU server. Companies without their own GPU server can have ERIS hosted by Upward on its own servers in the Netherlands, with the same local AI model. The AI agents that work across ERIS modules (meetings, tasks, projects, documents, CRM, chat and email in the beta, with more modules from January 2027) all process company data locally. Every action is logged in a single audit trail.

The beta starts in November 2026. Design partners and early testers can request access now.

Request access to the ERIS beta

FAQ

What is a GPU and why does on-premises AI need one?

A GPU (graphics processing unit) is a specialised processor designed to do many calculations in parallel. AI models perform billions of calculations per second. A GPU is fast enough to do that without the user waiting minutes for an answer. A standard CPU (the processor in most computers) would be too slow.

Does on-premises AI mean the company trains the model on its own data?

No. The model is built once by its creator, using a large public or licensed dataset. Once built, it is frozen. When it runs on company premises and reads company documents, it is using those documents to answer questions in real time. It is not learning or changing. The company's data remains separate from the model.

Can on-premises AI meet GDPR and NIS2 requirements?

On-premises deployment supports compliance with GDPR and NIS2 because the data stays under the company's control and can be logged completely. However, whether a specific deployment meets these rules depends on how the company uses it and what data it processes. Compliance is assessed per deployment, not guaranteed by the software alone.

Is on-premises AI faster than cloud AI?

Not necessarily. Speed depends on the hardware. A company with a high-end GPU server may see fast responses. A company with basic hardware will see slower responses than a cloud vendor with large-scale infrastructure. On-premises AI is chosen for data control, not speed.

What happens if the on-premises server fails?

The company's data is safe, but AI features are unavailable until the server is back online. This is why companies running on-premises AI usually have a backup or redundancy plan. Some also keep a cloud option available for non-sensitive queries when the local system is down.


Tell us which work takes too much time

That is usually the most useful first conversation. Whether the answer is a new system, an AI agent, an integration between what you already have, or nothing at all, you get an honest assessment.