Private LLM or privacy proxy: what an Irish firm needs

Do you need a private LLM, a language model running on your own servers, to use AI on client data? For most Irish firms of ten to a hundred people the honest answer is no. A self-hosted model makes sense in a few specific situations, and it costs more in people than in hardware. For the rest, keeping a commercial API and pseudonymising each request before it leaves gets you most of the privacy benefit for a fraction of the effort.

Nonimo Gateway runs on your own servers. See plans and get in touch.

The question usually comes up after a partner or a practice manager reads that client data “goes to America” when staff use AI. Two different worries hide in that sentence: who can read the data, and where it is stored. A private LLM answers both, at a price. A privacy proxy answers the reading question directly and leaves storage to your contract with the provider.

Three set-ups sold as a private LLM

The phrase covers three set-ups that are often confused.

set-up	where the model runs	who sees the prompt
self-hosted open-weight model	your server or your own cloud account	your firm
dedicated model from a cloud provider	the provider’s cloud, in a region you pick	the cloud provider, under your contract
business account with a public provider	the provider’s infrastructure	the provider, under its terms

The top row alone keeps the prompt entirely inside your walls. The middle row is often marketed as private, and it may be a good arrangement, but there is still a processor reading your data under an Article 28 contract. The bottom row is what most Irish firms already have, covered in our guide to which AI Irish firms can use under the GDPR.

When hosting your own model makes sense

There are real reasons to run a model yourself. They tend to look like this:

• A contract forbids it leaving. Some client contracts, particularly in the public sector and financial services, restrict where data may be processed. If yours rules out any outside processor, the decision is made for you.
• The work cannot be done with placeholders. If the model needs the actual medical history to reason about a patient, as in some clinical coding or triage work, pseudonymising the identifiers does not hide the sensitive part. Our guide to patient data and AI in Irish practices goes into this.
• Volume is very high and steady. A model running all day on a large, predictable workload can cost less per request than an API. Occasional use rarely reaches that point.
• You want to adapt the model to your own documents in ways a provider does not allow.
• You need to work without an internet connection, on a site or in a secure room.

If none of these describes your firm, you are likely to pay for a private LLM without using what it offers.

What it costs, without the sales figures

We will not quote a hardware price, because it moves every quarter and depends on the model. The cost shape is stable, though.

1. Hardware or cloud capacity. A capable model needs GPU servers, bought or rented. Smaller models run on less, and answer less well.
2. People. Someone has to install the model, keep the server patched, watch for failures, and decide when to move to a newer model. In a firm of 30, that is either a hire or a contractor on call.
3. Evaluation. Open models vary a lot in quality between tasks. Someone has to test whether the one you chose drafts a letter to the Residential Tenancies Board as well as the service your staff compare it with.
4. Security. The server now holds every prompt. Under Article 32 of the GDPR you are responsible for its security, its access controls and its logs, and a breach of it is yours to notify.

That last point is the one people miss. A self-hosted model does not make the data less personal. It makes your firm the single party answerable for it, which is a fair trade if you have the staff.

When keeping the API and pseudonymising is enough

Most of what Irish offices ask of a language model does not depend on knowing who the client is. Summarising a file, drafting a reply, pulling dates out of a lease, rewriting a policy in plain English: swap the names, numbers and addresses for placeholders and the model does the same job.

That is what a privacy proxy does. It replaces personal data in each request before it reaches the API, the model answers with placeholders, and the proxy restores the real values before the answer reaches your user. The DPC says pseudonymised data is still personal data for whoever holds the key, and our guide on whether pseudonymised data is personal data explains the Court of Justice’s position on the recipient. In practice, the provider and its logs see far less, and you keep the strong general model.

Where data is stored is a separate lever. OpenAI’s data controls documentation describes a European data residency region for the API, which requires an approved retention arrangement such as Zero Data Retention. Residency and pseudonymisation work well together; neither replaces the other.

A worked example from a Galway clinic

The clinic and the details are invented. A physiotherapy practice in Galway wants letters to GPs drafted from its treatment notes. The partners priced a private LLM and found they would need a server and someone to look after it.

Instead, the notes go through a proxy. “Referral for Tomás Breathnach, DOB 03/02/1961, 14 Mill Street, Oranmore, H91 R2C4, lumbar disc pain since March” reaches the model as “Referral for [PERSON_1], DOB [DATE_1], [ADDRESS_1], [POSTCODE_1], lumbar disc pain since March”. The clinical description still travels, which is health data under Article 9, so the clinic keeps its lawful basis and its contract with the provider in order. What no longer travels is the name, the date of birth and the Eircode that would tie the description to one man in Oranmore.

Nonimo Gateway, briefly

Nonimo Gateway is a privacy proxy for teams that keep the API. It runs on your own servers or in your own cloud, and the content does not pass through Nonimo. It is OpenAI-compatible, so you change base_url and keep your code. It replaces personal data on the way out and restores it in the response. You use your own OpenAI key, and it is available as a LiteLLM plugin.

Plans and prices are on the Nonimo Gateway page.

It helps with pseudonymisation, which Articles 25 and 32 of the GDPR name as a measure; it does not settle your lawful basis or your contracts. If you are still deciding what sits in front of the API, our guide to choosing an AI gateway in Ireland covers the layer before the proxy, and how an AI data leak happens covers the routes data takes out of your own apps. For a wider look at the tools in this space, see our overview of PII redaction software and the comparison with Microsoft Presidio.

A private LLM is a serious undertaking that some firms genuinely need. If yours is not one of them, keep the model you already pay for and stop sending it what it does not need. See Nonimo Gateway’s plans and get in touch.

Sources

Checked 20 October 2026.

• Regulation (EU) 2016/679, the GDPR. Article 25(1) and Article 32(1)(a) naming pseudonymisation; Article 28(3) requiring a contract with a processor; Article 9 on health data.
• Data Protection Commission, Anonymisation and pseudonymisation. Pseudonymised data is still considered personal data.
• Data Protection Commission, AI, Large Language Models and Data Protection, 18 July 2024. Understanding how a supplier’s AI product handles personal data on your behalf.
• OpenAI, Data controls in the OpenAI platform. API data not used for training by default; the European data residency region and the retention arrangements it requires.