Your team already pastes client work into ChatGPT. Locul gives them the same easy chat and a set of bots for the work they repeat, running on top open models on servers used by your firm alone, in your country or your own building. Prompts, files and answers never go to OpenAI, Google or anyone else.
Private AI for business is an AI assistant that runs on infrastructure reserved for your firm, so prompts, files and answers are processed there and never sent to a public AI company. In practice that means open-weight models on dedicated servers in your country, or in your own building.
The test is simple: if a client asks where their data went when your team used AI, can you name the building? With Locul, you can.
Nobody sets out to leak a client file. Someone has a messy spreadsheet, a meeting in an hour and a ChatGPT tab open. Here is the Friday afternoon version, and what the same job looks like when the AI runs inside the firm.
Sorry to do this on a friday. Anita has just told me Megan put the whole Q3 payroll export for Castell Foods into ChatGPT to 'tidy the columns' before the client meeting. Names, NI numbers, salaries, the lot. She meant well, the spreadsheet was a mess (the date column is half 03/09/26 and half Sept 3rd because two people maintain it).
Can we a) find out if that is now a reportable breach, b) stop it happening again without banning AI outright, because frankly half the office uses it and Tom in audit says he'd quit lol. Castell's FD is already twitchy after the HMRC letter.
G
From: Anita Desai · 24 October 2026 14:12
Gareth, quick one, not on email ideally. Ring me?
Gareth, short version first: treat it as a possible breach until we know more, but it is unlikely to need reporting to the ICO if the chat is deleted today and the account is a business one with training switched off. I've asked Megan which account she used. If it was her personal free account, that changes things, and we have 72 hours from when we became aware (this afternoon, 14:12) to decide.
For next week: Castell's file and every other client file can go through Locul instead. It runs on our own servers in the UK, so the same 'tidy these columns' job never leaves the building. I've also fixed the date column in the copy on the shared drive. Twelve rows were in the 'Sept 3rd' format.
An example. The people, firm and client are invented, the mess is not.
Most firms weigh the same three options. The difference is not the chat window, which looks the same in all three. It is where the model runs, who keeps it running, and what you can show an auditor afterwards.
| Business plans of public AI (ChatGPT Enterprise, Copilot) | Build it yourself (open-source UI on your own server) | Locul (managed private AI) | |
|---|---|---|---|
| Where prompts are processed | The provider's cloud, run by a US company | Your server | Servers for your firm alone, in your country or your building |
| Who keeps it running | The provider | Your IT team: drivers, updates, backups, security patches | Locul or a certified partner in your country |
| Overseas transfer to assess | Yes, covered by contract | No | No |
| Works on an isolated network | No | Yes, if you build it | Yes, on your own servers |
| Bots for repeat work | Build your own agents | Plugins, if someone maintains them | Ready-made bots with schedules |
| Asks before anything leaves | Depends on the setup | Not built in | Every email or post waits for a yes |
| Audit trail on your side | Admin exports | Whatever you configure | Every question, answer and file access logged |
| Model choice | The provider's models | Any open model you can run | Top open models, upgraded when a better one ships |
Thinking about the build-it-yourself column? Our self-hosted LLM guide lists what it really takes, and the comparisons with Open WebUI and AnythingLLM are fair to both sides.
Public AI bills by usage or by seat on shared hardware. Self-hosting looks free until you add the GPUs, the power and the person who patches them. Locul is a flat price per seat that already includes the dedicated servers, the setup and the model upgrades, so a busy month costs the same as a quiet one. The cost depends mainly on where it runs, and every option is on the pricing page.
With no overseas transfer and no third-party AI processor, the hardest parts of a data protection impact assessment get much shorter. The EU AI Act mostly asks firms using general-purpose AI to train staff and keep a record of how it is used. The audit log covers the record, and the approval step keeps a person in charge of anything that reaches a client.
"Your data is safe with us" is not an answer. These are. Ask them of us and of anyone else on your shortlist.
Not "somewhere in a region". A named data centre in your country or your own server room.
Including the vendor's support staff, and including logs kept for "abuse monitoring".
The answer should be no, with no opt-out form needed.
A junior should not get the partners' pay review because they asked nicely.
On your side, ready when a client or regulator asks who saw what.
Some contracts require it. Most firms never need it, but the answer tells you how the system is built.
A named partner in your country or the vendor's own engineers, not a community forum.
Locul comes with bots for the work every firm repeats. Each has one job, a schedule and a memory of how your firm likes things done, and each one asks before anything goes to a client.
Sorts the shared inbox, drafts the routine replies and tells you the three that need a person today.
Drafts the right reminder for each late payer. It knows Harlow Ltd pays after a phone call, not an email.
Answers staff questions from your own handbook and links the section, so nobody guesses about phones or home working.
Checks every new client file for the paperwork your rules require, like a signed engagement letter.
Drafts first answers from your past winning bids, in the word limits the buyer set.
A one-page brief on the client: open matters, last emails, the thing they complained about in March.
Your staff see the same browser app either way. The difference is who owns the hardware.
Dedicated GPU servers in a data centre in your country, for example in Germany or Finland, set up for your firm alone. Nothing to buy, nothing to run.
Locul installs on GPU servers in your comms room or data centre. It can run with no internet connection at all.
Sizing the hardware, power and cooling for the on-site option is covered in the on-premise LLM guide. Both options are compared on the deployment page.
We test open models on real office work: messy emails, spreadsheets with broken formulas, forwarded chains with the ask buried in paragraph four. See how close they get to the big cloud models.
Matter files, precedents and privileged emails stay inside your boundary.
Payroll exports, working papers and tax files, handled in your country.
Letters, admin and policy questions, with no patient data reaching an outside AI.
Specs, bids and controlled drawings on hardware your security team signs off.
Private AI for business means an AI assistant that runs on infrastructure your firm controls or rents for itself alone, so prompts, documents and answers are processed and stored there and are never sent to a public AI provider. It usually runs on open-weight models instead of a vendor's cloud model.
They promise not to train on your data, and for many firms that is enough. They still process every prompt on a US company's infrastructure, so you rely on a contract instead of on where the data physically goes. If a client, insurer or regulator asks where their data was processed, private AI lets you give a one-line answer.
For drafting, summarising, checking documents against a policy and answering questions from your own files, the best open models are now close to the big cloud models. We test this on messy, real office tasks in the Office AI Benchmark, and you can see the results before you decide.
On dedicated servers in a data centre in your country, set up for your firm alone, or on GPU servers in your own building. Both options are explained on the deployment page, and the on-site option is covered in depth in our on-premise LLM guide.
No. A certified partner in your country or our own team sets it up, connects your drives and wiki, and keeps the models updated. Your staff open it in the browser like any other web app.
Per seat per month, and never per prompt. The price depends on where the AI runs. Every option is on the pricing page.
One email. Tell us the work your team repeats and your data rules, and we reply with where Locul would run and how it handles that work.