View all articles
Local AIOpen WebUIGDPRSMEsPrivacy

Private ChatGPT for your business: three options and the cost

JG
Jacobo González Jaspe
|

Reviewed:

Abstract illustration: concentric rings of a vault mechanism with amber light in the seams. Para tu empresa
Illustration generated with AI on our own machine.
This article is also available in Spanish:ChatGPT privado para tu empresa: tres opciones y qué cuestan

By the end of this guide you will know what “private” really means in three ways of giving your office an AI chat, how to work out the monthly cost of each with your own figures, and what the law asks of you. The cheapest test is the 16 GB laptop you already have.

Who this is for, and who should not do it

It fits an office of 3 to 30 people who paste real work into a chatbot: client emails, contracts, accounts.

Do not host your own chat if any of these is true:

  • Nobody will look after it. Updates and backups are real work. Without a named person, a cloud plan with a signed contract is the honest choice.
  • Low volume. For a handful of questions a week, any machine costs more than the seats.
  • You need the strongest model. We have not tested a local model against the largest cloud ones. Try your hardest task on both before deciding.
  • Work cannot stop for a day. One machine is one point of failure.

What you need

  • A list of what your team uses AI for today, and which of it contains personal or confidential data.
  • A laptop with 16 GB of memory for the trial. In our runs, 7B and 8B models took 4.75 to 9.21 GB once loaded (6 Nov, 2 Nov 2026).
  • One person with an afternoon for two guides: installing Ollama and setting up Open WebUI.

What “private” means at each option

Everything in this section comes from each provider’s own pages, read on 15 November 2026. Prices and terms change, so open the links on the day you decide.

ChatGPT BusinessAPI with an EU regionChat on your own machine
Where your text is processedOpenAI’s servers; OpenAI does not list this plan among those with European storageIn Europe, once configuredIn your office
Used to train modelsNot by defaultNot unless you opt inNo AI provider involved
Who can read the chatsYour workspace admins; OpenAI staff and contractors in limited casesThe provider, in abuse logs kept up to 30 daysYour administrator
Contract you signData processing addendumData processing addendumNone with an AI provider
You payPer seat, per monthPer use; the region can add 10%Machine, electricity, hours

Option 1: ChatGPT with a business plan

Personal accounts are not this option. OpenAI’s pricing page says of Free, Go, Plus and Pro that content is used to train its models, with an opt-out. An employee’s personal account is also their agreement, not your company’s.

ChatGPT Business is the self-serve plan for teams, called Team until August 2025. From OpenAI’s own pages:

  • Price. A Standard seat is USD 25 per user per month, or USD 20 billed annually, with a minimum of two seats. A Premium seat with higher limits is USD 125, or USD 100 billed annually. The plan overview adds that pricing “may vary depending on your country and currency type”. We could not read a price in euros without starting a purchase, so take yours from the checkout.
  • Training. OpenAI states it does not train on Business, Enterprise or API data by default (business data page).
  • Who reads the chats. Workspace admins “can view, access, export, and delete end user conversations”. OpenAI limits its own access to authorised employees and to contractors reviewing abuse (enterprise privacy, updated 8 January 2026).
  • Retention. Admins can set how long chats are kept. Deleted or unsaved conversations leave OpenAI’s systems within 30 days, unless the law or protecting the service requires longer.
  • Location. OpenAI offers storage in Europe to eligible Enterprise, Edu, Healthcare and API customers. Business is not on that list.
  • Contract. OpenAI signs a data processing addendum for Business, Enterprise and the API. For customers in the EEA the counterparty is OpenAI Ireland Ltd. Section 4.1 covers transfers outside the EEA with standard contractual clauses or an adequacy decision.

ChatGPT Enterprise adds European storage for eligible customers and an audit log for admins. Its price is not published: the pricing page sends you to sales, on annual plans.

So with Business, “private” means a contract, no training and admin control. It does not mean your text stays in Europe.

Option 2: an API with a European region

An API is not a chat page. It is the model on tap, billed by use, and someone has to put a chat in front of it. Open WebUI, the page used in option 3, can connect to these services.

  • OpenAI. API text is not used for training unless you opt in, and abuse-monitoring logs are kept for up to 30 days (data controls). You choose Europe when you create the project and cannot change it later. Requests go to eu.api.openai.com, with a 10% uplift on models released from 5 March 2026. Account, billing and usage data may still be handled elsewhere (data residency).
  • Mistral. The EU address api.eu.mistral.ai costs 1.1 times the list price and lacks agents, batch jobs and the files API (regional inference).

“Private” here means a contract, no training and a region that somebody turned on and checked. Our local AI guide covers Microsoft’s regions and the status of the EU-US Data Privacy Framework. The EU data-residency checklist walks through the settings.

Option 3: a chat on your own machine

The text never leaves the office, because the model runs there. Nobody outside can read it, train on it or change the terms. Everything the provider did becomes your job. The next step shows what that is.

Step 1: see what a self-hosted chat consists of

Four pieces, all free software:

  1. A machine that stays on. For the trial, a laptop.
  2. A model server. Ollama downloads a model and answers requests.
  3. A chat page. Open WebUI gives each person an account in the browser. Its licence is free to use and asks you to keep its branding, with an exception for up to 50 users in 30 days.
  4. A person. Accounts, updates, backups.

To see it work on one computer, from Ollama’s site (ollama.com/download) or on Linux:

bash
curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.1:8b "Write a two-line reply confirming Tuesday's meeting."
ollama ps   # the memory the loaded model takes

The model is a 4.9 GB download.

On 15 November 2026 we set up Open WebUI 0.13.3 on a test instance with invented accounts and recorded what happened (the record). What an owner should know from that run:

  • A new colleague sees no models. The administrator saw all 14 on the server; each ordinary user saw none until a group was given access. An empty list on day one is the default.
  • Sign-up closed itself after the first account. Nobody who finds the address can register.
  • The administrator can read every chat by default. Two settings hide them in the product. Whoever controls the machine or its backups can still read the database.
  • Chats and files are not encrypted by the application. Use an encrypted disk.
  • The backup that worked was a copy of the whole data folder with the service stopped. The “download database” button gave us a file with no users and no chats, minutes after we created them.
  • Nothing is deleted on its own. We found no retention setting, so the period is yours to write down and apply.

This is the same level of admin access as ChatGPT Business. The difference is who the administrator is.

What it can and cannot do: what we measured

All runs are ours, on NVIDIA DGX Spark machines (GB10, 128 GB of memory). Each file holds the command and the raw output.

Office taskResultData
Drafting speed, one personllama3.1:8b: 43.7 tokens a second, first word in 0.39 s2 Nov 2026
A client email and a summary in Spanish, four 7B and 8B models, scored 1 to 5Best email: 4, 5 and 4 on three criteria. One model copied the minutes back unsummarised8 Oct 2026
40 invoices as text, 360 fieldsA 26B model: 360 right. Five 7B and 8B files: 321 to 3445 Nov 2026
Transcribing 36 s of SpanishWhisper medium: 2.9% of words wrong, in 23.1 s8 Nov 2026
Ten people at once, Ollama as installedMedian wait for the first word: 71.7 s; worst: 146.4 s2 Nov 2026
The same ten with vLLMMedian wait 0.17 s; full answer in about 15 s2 Nov 2026
Eight people, Ollama with eight slotsWorst wait fell from 31.68 s to 0.47 s; memory rose from 4.75 to 6.53 GB6 Nov 2026

How to read it:

  • It is a ceiling. We have not timed a 16 GB laptop or a mini PC. Yours will be slower.
  • Small models draft; they do not sign. Every Spanish draft needed an edit. The scores are one reader’s: our AI assistant with a written rubric, not a panel.
  • Detail suffers. Ten of the 40 invoices arrived inside an email. The 7B and 8B files got every field right on one to five of those ten.
  • A team needs parallel answers. As installed, Ollama served one request at a time and the rest queued. With slots configured it served eight. See Ollama or vLLM for a small office and, for the memory each slot costs, how much memory a model needs.
  • The audio and invoices are synthetic.

Step 2: work out your own cost

We give you the sum and where to read each input, because any figure we printed would be stale within months.

text
Seats, per month      = people × price per seat at your checkout
API, per month        = tokens used × price per token (× 1.1 where the EU region has the uplift)
Own chat, per month   = machine price / months of use
                      + watts × hours per day × 30 / 1000 × price per kWh
                      + hours of upkeep × what that hour costs you
Break-even in seats   = own chat per month / price per seat
InputWhere to read it
Price per seatThe provider’s checkout, in your currency, on the day
Tokens used and their priceThe API dashboard and the provider’s pricing page
Machine priceThe shop’s page on the day; zero while you test on a laptop you own
Months of useYour decision; ask your accountant how long you write equipment off
WattsA plug-in power meter, read while the machine works and while it waits
Price per kWhYour electricity bill
Hours of upkeepA note kept during the two-week trial

One worked line with a sourced price. Apple’s Spanish store listed the Mac mini from EUR 1,069 on 15 November 2026 (apple.com/es). The lowest configuration on that page has the M6 chip, 16 GB of memory and 256 GB of storage. Over 36 months that is EUR 29.69 a month for the machine alone. We have not measured that computer, so it is an example of the sum, not a recommendation.

Two warnings. The wattage printed on a power supply is a maximum, not what the machine draws, and we have not metered our own. And the hours are the line owners forget: a chat nobody maintains is cheaper only on paper.

For a full sheet by volume, use cloud vs local: your break-even. Its cloud prices date from September 2026; check them first.

Step 3: try it for two weeks before buying anything

  1. Install Ollama and Open WebUI on the laptop, following the two guides above.
  2. Create an account for three or four colleagues and give their group one model.
  3. Ask them to use it for real work that contains nothing confidential yet.
  4. Keep four notes: questions per day, the most people asking at once, the tasks where the answer was not good enough, and your own hours.
  5. Put the notes into the sum in Step 2.

If the answers are too slow, do the memory sum before blaming the model. If they are not good enough, try a larger model on the same tasks or keep those tasks in the cloud. A split is normal, and hybrid routing explains how to do it.

Data protection: what changes with each option

This is information, not legal advice. Article numbers link to the GDPR on EUR-Lex. All of it is in force.

GDPR dutyBusiness plan or APIChat on your own machine
Processing contract, Art. 28Required: sign the provider’s addendumNone with an AI provider; perhaps one with an outside technician
Transfers abroad, Arts. 44 to 46Check the region and the transfer basis in the addendumNone while machine and backups stay in house
Security, Art. 32Shared with the providerAll yours: access, encryption, updates, backups
Telling staff, Art. 13YoursYours
Retention, Art. 5(1)(e)Set it in the admin consoleWrite it down and delete by hand
Record of processing, Art. 30YoursYours

Three points owners get wrong:

  • Local does not make you compliant. It removes the provider’s contract and the transfer. The legal basis, the record, security and people’s rights stay with you. See GDPR Article 25 and local AI.
  • Reading staff chats has rules. In Spain, Article 87 of Organic Law 3/2018 lets an employer access content on work tools only to check that duties are met and to protect the devices. It also requires usage criteria that staff are told about. That applies to a ChatGPT Business admin and to your own administrator alike.
  • A technician with access is a processor. If an outside company logs in to the machine that holds the chats, the Article 28 contract comes back. Our local AI guide explains the test.

The notice the AI Act asks for

Article 50 of the AI Act has applied since 2 August 2026. Paragraph 1 asks the provider of a system that talks to people to make sure they know it is an AI, unless that is obvious. The Commission’s FAQ, updated 24 July 2026, says the exception is read restrictively and the notice comes at the start of the first interaction.

With a vendor’s chat, the vendor is the provider. If you set up your own and offer it under your name, that may be you. One sentence on the first screen settles it:

text
You are chatting with an AI assistant, not a person.
It can make mistakes, so check anything important.

The other labels are in who labels what under Article 50.

Mistakes people make

  • “We pay for ChatGPT, so it is private.” Only a business plan carries the no-training default and a contract. A personal Plus account does not.
  • “The business plan keeps data in Europe.” Not on Business, from what OpenAI publishes.
  • “On our own machine nobody can read it.” The administrator can.
  • “After the hardware it is free.” Electricity and hours are costs.
  • “We backed up with the download button.” Ours came out empty. Copy the folder with the service stopped, and restore it once to be sure.

Verdict

  • Buy (set up your own). Daily use, confidential text in the prompts, and a named person to look after it. Size the machine after the trial.
  • Wait. You do not know your volume yet. Run Step 3 and come back with the notes.
  • Skip. Ten questions a week, or nobody to maintain it. Use a business plan, sign the addendum and add the AI notice.

Questions owners ask

Can my staff paste client data into ChatGPT Business? With the addendum signed, you have the contract the GDPR requires. You still need a legal basis and a decision on the transfer, since OpenAI does not list Business among the plans with European storage. Ask whoever handles your data protection.

Can I read what my employees ask? Technically yes, in both ChatGPT Business and Open WebUI. Legally, only within the criteria you have told them about.

Does a local model learn from our documents? No. The model is a file that does not change as you use it. It knows your company only through what you paste or connect.

Is it as good as ChatGPT? Not at everything. In our tests a 26B model read 40 invoices without a mistake, and small models wrote Spanish that needed editing. We have not compared against the largest cloud models.

How many people can one machine serve? On our hardware one 7B model answered ten people at once when the server ran requests in parallel. On a smaller machine, measure it during the trial.

Can we use both? Yes. Open WebUI can connect to a cloud API next to the local model, though we did not test that. Decide in writing which work goes where.

What happens if the machine breaks? Work stops until you restore the backup on another computer. We restored one on the test instance and accounts, chats and settings came back. Keep a cloud plan as plan B if a day without chat is a problem.

Limits and what we did not test

  • No measurement on a laptop or a Mac mini, and none of electricity.
  • No comparison of answer quality against ChatGPT.
  • We read OpenAI’s public pages; we did not buy a Business workspace or sign its addendum.
  • The Open WebUI run used four invented accounts through its API, not thirty people in a browser.

Next steps

Work with us

We size the model and the machine by measuring, not by guessing, and we say so when the cloud is enough. A free 15-minute call is usually enough to know which of the three options fits: book a call or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 102 guides and 25 templates. All free, no signup.

Bonus: the EU AI Act checklist, ready to complete
Once a week No spam Unsubscribe anytime

See what you get

The EU AI Act now applies: a checklist you can complete

Tell us what your organisation needs

In 15 minutes we look at what the AI Act asks of you, whether AI on your own machines makes sense for you, and where to start. No commitment, and nothing you do not need.

First call free, 15 min Your data stays on your network AI Act and GDPR from day one

102 free guides · 17 compliance templates