Using AI with your own company data: ask your inbox and files without giving them away

You can use AI with company data privately, without handing that data to public AI models, when three things hold: the AI runs in a closed system for your company only, the AI provider is contractually barred from training on your inputs, and you decide which data it may access. That is how the Verselo AI assistant works: a dedicated, isolated server in the EU, connected to your inbox, that knows your business and puts answers in front of you as drafts.

Why does your most important company knowledge sit in the inbox, and why can nobody find it when it is urgent?

A made-up example. A heating and plumbing firm with 14 people. A regular customer calls and asks what price they were quoted in spring for servicing the boiler and whether the call out fee was included. The answer is somewhere in a March email the owner wrote in reply to a question about the quote.

Nobody in the office finds it within five minutes. Searching the inbox for "service" returns 40 hits. The owner is on site. The customer is waiting.

That is what company knowledge looks like in a small business. There is no wiki and no maintained database. What the company knows sits in sent emails, in quotes saved as PDF, in price lists and in the heads of two or three people. New staff have to ask. When the person who knows the answer is on holiday, that knowledge is unavailable.

The obvious idea is to give all of it to an AI and ask. The real question is which AI, and on what terms.

Searching by hand, pasting into ChatGPT or asking your own AI assistant: what is the difference?

In daily work there are three routes. They differ less in technology than in who controls your data.

	Search by hand	Paste into a personal chat window	Dedicated AI assistant on an EU server
Time per question	Often several minutes	Fast, if you know what to copy	A short question, an answer based on your emails
Contract with your company	Not needed	None, personal account	DPA for running the assistant
Training on your data	No	Possible, depending on plan and settings	Business access at the AI provider only, no training
Who decides what goes in	Nobody has to	Each employee alone	Fixed rules you set
Knows the context	Only whoever searches	Only what was pasted	Your tone, your services, your confirmed facts

The middle route is the most common and the riskiest. An employee copies the customer's email, with name, address and quoted price, into a personal account and asks for a reply. It is quick. But your company knows nothing about it, and no contract governs what happens to the data. We cover this in more detail in Is ChatGPT GDPR compliant for business?.

What do the German DSK, Spain's AEPD and the GDPR say about which company data may go into an AI?

Regulators do not ban AI. They describe fairly precisely what a clean setup looks like (this is not legal advice).

Germany's data protection conference (DSK) separates open from closed systems in its guidance on AI and data protection of 6 May 2024. In open cloud systems, inputs leave your protected area and may be reused. The DSK therefore says technically closed systems are preferable from a data protection point of view, as are applications that do not use inputs and outputs for training.

Spain's data protection authority (AEPD) goes further in its guidance on agentic AI (version 1.2, February 2026). It stresses clear access policies and compartmentalised agent memory; the exact implementation depends on the processing involved.

The GDPR sets the frame. Art. 5(1)(c) requires data minimisation: only as much data as the purpose needs. Under Art. 28(3)(a) a processor handles your data only on your documented instructions. Art. 32 requires appropriate technical and organisational security measures.

Translated to a business with 5 to 50 people: company knowledge may go into an AI when the system is closed, a contract is in place, your data is not used for training and access is limited to what is needed.

How does a dedicated AI assistant learn your company knowledge without your data training someone else's model?

It helps to separate two things that often get mixed up. Training a model means your data becomes a permanent part of a provider's model. An assistant that knows your business does not need that.

With the Verselo AI assistant it works like this:

1. Your tone, from your emails. During setup the assistant reads your sent emails and learns your greeting, your terms and your usual answers.
2. Confirmed facts only. It only keeps services, prices, deadlines and property details once you have confirmed them. Your emails stay in your inbox.
3. One server per company. Every client gets a dedicated, isolated server in the EU. No shared platform where many companies sit together.
4. Writing without training. To write the text, the assistant uses business access at the AI provider, where your data is not used for training. During setup we explain transparently which data is processed and why.

What providers say about this can be checked. OpenAI states on its business data page that by default it does not train on data from ChatGPT Business, Enterprise or the API. Anthropic says in its Privacy Center that it does not use inputs or outputs from its commercial products for training by default, unless the customer allows it or submits feedback.

Back to the plumbing firm. The owner messages the assistant from the phone: "What did we quote the Clarks in March for the boiler service?" The assistant finds the email, gives the price and whether the call out fee was included and, if asked, puts the reply to the customer in the inbox as a draft. For servers, contracts and data transfers outside the EU, read AI assistant on an EU server or US cloud.

Who on the team may ask what, and how do you start without an IT project?

Clear access rules are simpler in practice than they sound. Before you start you answer three questions: Which inboxes does the assistant read? Which kinds of email may it answer? Who controls it?

With the Verselo AI assistant, phone control is locked to your own account. At the start it sends not a single email on its own. Everything waits as a draft in your inbox, and prices and commitments never go out without asking. Every rule change is saved and can be rolled back in one click. More on servers and access is on our security page.

Starting looks like this:

1. A 30 minute call. We look at where your company knowledge sits and whether the assistant pays off for your inbox.
2. Setup. We set up your dedicated server and connect inbox and calendar. You sign the DPA with us.
3. Two weeks of drafts. You read along, approve and confirm the facts it should remember.
4. Automation on your terms. After that you decide, per kind of email, what may run on its own.

Firms bound by professional secrecy can see how this works with client emails on the AI assistant for tax advisors page. If your team mainly wants to query manuals, contracts and proposals in folders with separate access rights, the company knowledge page shows how the Verselo platform is planned to handle that after launch.

When is it not worth it? When your knowledge sits in a few well kept documents and few customer emails come in. Then a business plan with clear rules for the team is often enough.

Conclusion

Using AI with your own company data is not automatically a data protection problem when the system is closed, does not train on your data and only accesses what it needs; the specific processing still needs a legal review. How a dedicated assistant would work with your inbox and your company knowledge is something we can go through in a 30 minute call.