AI gateway: what it is and when a UK firm needs a private one

An AI gateway is the single doorway your applications use to reach large language models. Instead of each internal tool holding its own OpenAI key and calling the API directly, every request goes through one service that picks the model, applies limits, keeps the keys and writes a log. The term LLM gateway means the same thing.

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

Whether you need a private gateway turns on that last job. A gateway that stores prompts is storing whatever your staff and systems put in them, and in a solicitors’ practice, an accountancy firm or a supplier to the NHS that means client names, dates of birth and case details. Under the UK GDPR the log is personal data in its own right, wherever the gateway runs.

This guide sets out what a gateway does, what its logs mean for a UK controller, how to choose between a managed and a private gateway, and where a privacy proxy fits in the chain.

What an AI gateway does, job by job

Most gateways on the market share the same core. The right-hand column is the part a data protection officer will ask about.

Job	What it means in practice	The UK GDPR reading
Routing	Sends each call to the model or provider you choose, with fallbacks	Each provider is a recipient, so it belongs in your records and privacy notice
Key management	Holds provider keys centrally and issues internal keys per team or app	Good security under Article 32: fewer keys scattered across laptops
Budgets and limits	Caps spend and request rates per team	Neutral for personal data
Logging	Keeps requests and responses for debugging, cost and audit	Processing of everything in the prompt, with a retention clock
Guardrails	Checks content on the way in or out, and blocks or edits it	Can support minimisation, depending on what it actually does

Routing and keys are why IT teams adopt a gateway. Logging is why the gateway ends up on the agenda of whoever owns your record of processing.

Why an AI gateway’s logs are a UK GDPR question

Picture the log line for an ordinary request from a family law team: the matter number, the client’s full name, the other party’s address, the children’s dates of birth, and a paragraph about a non-molestation order. The model saw it once. The gateway may keep it for months.

Three provisions of the UK GDPR land on that line.

1. Storage limitation. Article 5(1)(e) says personal data may be kept in identifiable form for no longer than is necessary. A debug log with no deletion date fails that test on its face.
2. Processors. If a vendor hosts the gateway, it processes the log on your behalf. The ICO’s guidance describes a processor as one that “processes personal data on behalf of the controller”, and Article 28 requires a contract that covers instructions, confidentiality, security and deletion at the end of the service.
3. Transfers. If that vendor stores logs outside the UK, the ICO treats sending them as a restricted transfer, which needs adequacy regulations or a safeguard such as the International Data Transfer Agreement. The UK-US data bridge covers US recipients that participate in the UK Extension to the EU-US Data Privacy Framework, not every US company.

None of this makes a managed gateway unlawful. It means the gateway is a processing operation you document, with a retention period and a contract, the same as your practice management system. If what passes through it also includes health data, the stricter rules in our guide on special category data in AI apply as well.

Managed or private AI gateway: how to choose

A managed gateway is a cloud service you point your apps at. A private gateway is software you run on your own servers or in your own cloud account. Both are respectable choices, and the right one depends on what is in the traffic.

Two examples from our comparison pages show the range. Cloudflare AI Gateway runs on Cloudflare’s network, its core features are offered at no charge, and its DLP can flag or block prompts and responses. Its documentation also notes that when DLP scans responses it buffers the full streamed response before scanning. The detail is in our page on Cloudflare AI Gateway and PII.

At the other end, the LiteLLM proxy is open source and self-hosted, presents an OpenAI-style interface in front of many providers, and takes guardrails as plug-ins. With the Presidio guardrail it can mask entities and, for replace operations, put the original values back in the response. Our LiteLLM PII guardrails page covers how that is wired.

Five questions settle most cases:

• Will prompts carry client, patient or employee data? If never, a managed gateway with sensible log settings is usually enough.
• Who can read the logs? Your own IT team, the vendor’s support staff, or both.
• Where are the logs stored, and for how long? A UK or EEA location and a short retention period make the transfer and storage questions easy.
• Does it stream? Users expect answers to appear word by word. Anything that waits for the whole answer changes how your tools feel.
• What happens to personal data in the prompt? Logged as is, blocked, masked for good, or replaced and restored.

The last question is the one a gateway on its own rarely answers well, which is where a privacy proxy comes in.

Where a privacy proxy fits in the chain

A privacy proxy has a narrower job than a gateway. It reads each outgoing request, replaces personal data with consistent placeholders, forwards the clean request to the provider, and swaps the real values back into the answer before your application sees it. The model reasons over “CLIENT_1” and “ADDRESS_1”; your user reads the real names.

There are two common layouts:

1. Proxy behind the gateway. Your apps call the gateway, the gateway calls the proxy, the proxy calls OpenAI. The gateway keeps routing and budgets; the provider receives placeholders.
2. Proxy as a guardrail inside the gateway. The gateway hands each request to the proxy before sending it on. This suits teams already running LiteLLM.

To make it concrete, take an invented example. Harrow and Pike, a 22-person conveyancing practice in Leeds, has built a small internal tool that drafts replies to buyers’ enquiries. A typical prompt reads: “Draft a reply to Mr Daniel Okafor, 41 Cardigan Road, LS6, about the completion date of 14 November and the retention of £2,500 for the boundary fence.” Without a proxy, that sentence sits in the gateway log and reaches the provider intact. With one, the provider receives “Draft a reply to PERSON_1, ADDRESS_1, about the completion date of DATE_1 and the retention of AMOUNT_1”, and the draft comes back with Mr Okafor’s details restored. The firm’s own log can hold the placeholder version, which keeps the record useful for debugging without it becoming a second copy of the client file.

Placeholders do not make the data anonymous. The ICO is clear that pseudonymised data remains personal data for whoever holds the key, which is your firm. What changes is who else holds the identifiable version. Our guide to pseudonymisation under UK GDPR explains the distinction, and Article 25 names pseudonymisation as one of the measures for data protection by design.

Nonimo Gateway

Nonimo Gateway is a private proxy of exactly this kind. It runs on your own servers or in your own cloud, and the content does not pass through Nonimo. It is OpenAI-compatible, so your code keeps working after you change base_url, and you keep using your own OpenAI key. It replaces personal data in what goes out and restores the real values in the response, and it ships as a plugin for LiteLLM if you already run that gateway.

Plans and prices are on the Nonimo Gateway page.

It helps with pseudonymisation and minimisation; it does not decide your lawful basis or your retention policy for you.

Choosing your AI gateway setup

If your traffic contains no personal data, pick the gateway that suits your routing and budgeting needs and move on. If client files travel through it, decide where the logs live and what the provider receives before any tool goes live, and treat both as processing you can explain to the ICO. For the wider picture of AI tools at work under UK law, see our guide on the Data Protection Act 2018 and AI tools. If you are weighing whether to run your own model instead, read private LLM or privacy proxy, and for the ways data slips out of internal apps, how a ChatGPT data leak happens from business apps.

See Nonimo Gateway’s plans and get in touch.

Sources

Checked 20 October 2026.

• UK GDPR, Article 5, legislation.gov.uk. Data minimisation, Article 5(1)(c), and storage limitation, Article 5(1)(e).
• UK GDPR, Article 25, legislation.gov.uk. Data protection by design, naming pseudonymisation as a measure and data minimisation as a principle.
• UK GDPR, Article 28, legislation.gov.uk. Sufficient guarantees from processors and the contract terms in Article 28(3).
• UK GDPR, Article 32, legislation.gov.uk. Pseudonymisation and encryption listed as security measures.
• ICO, What are controllers and processors?. The definitions of controller and processor.
• ICO, A brief guide to international transfers, updated 15 January 2026. Restricted transfers, adequacy regulations and the IDTA.
• GOV.UK, UK-US data bridge: supporting documents. In effect from 12 October 2023, for US recipients participating in the UK Extension to the EU-US Data Privacy Framework.
• ICO, Pseudonymisation. Pseudonymised data remains personal data for whoever holds the additional information.
• Cloudflare, AI Gateway pricing and AI Gateway DLP. Core features at no charge; DLP flag and block actions; full buffering of streamed responses when responses are scanned.
• LiteLLM, documentation and PII masking with Presidio. Self-hosted gateway with an OpenAI-format interface; mask or block per entity; restoration of masked tokens for replace operations.

---

This page is for information and is not legal advice. Product details of other vendors were checked on their official pages on 20 October 2026 and can change.